Busy week of side-hustling and full-time hustling.

And another AI article to question the future of work.

Topics:

• Moderation Moderation

• Push Campaigns

• Something Big is Happening

Moderation Moderation

A few new features came to Wave before I went back to full time, including built-in camera integration for taking pictures and videos within the app, and Anonymous posting.

The idea behind Wave is to focus on the current evening - what’s happening while you’re out, not what happened last week.

With that came the idea of regularly clearing out the social feeds to keep things relevant to right now.

That also led to the idea of anonymous temporary posts.

Might be easier to post an update without worrying how that post will look in the morning.

In all honesty, I pushed back against that idea because anonymous content and comments means constant moderation.

But thankfully I discovered OpenAI’s free moderation API a while ago, so that will be the saving grace.

OpenAI gets to judge every word and image that attempts to be published, anonymously or otherwise, within Wave.

If they flag it, it doesn’t make it to the feed.

Instead, the post sits in a moderation queue and someone has to manually decided if it’s good to publish or not.

As soon as it’s marked “OK” it shows up in the app.

On the flip side, a post can be marked as “flagged” and it will no longer be available to view.

It’s a pretty cool system, but at this point it might be a little too safe, or at least too broad.

A phrase like “last night killed me” gets flagged as violent.

The word “stupid” gets a low grade rating for harassment.

One phrase is clearly just a figure of speech, not harmful to anyone, but technically violent in nature due to the choice of words.

The other is technically harassment.

In this day and age, anonymously calling someone stupid has to be 50% of all internet traffic so I’m sure we’ll allow it, but it further illustrates how complicated moderation can get.

OpenAI assigns a value with this harassment rating to determine how harassing it is.

“Stupid” tends to be in the .1 to .2 range from 0 to 1, where stronger words would have much higher ratings.

For now, I’m letting the moderation system hide anything that gets flagged at all, but I’m thinking the threshold will need to be much higher.

And it may even be time to consider a second AI review to look at the context and decide if it needs further review from a human.

The moderation API is basically classifying so it’s going to be pretty thorough.

A small, fast and cheap model could further check the context of the flagging and decide if it’s relatively harmless or not if it’s given some guidelines.

An AI moderator for the AI moderator.

That may be the next upgrade, but I’ll be keeping an eye on the moderation process overall for now.

If too much is getting flagged, I’ll be deploying new moderation measures soon.

For the time being, I’d rather err on the side of caution with too much moderation before I hand the entire thing over to AI even though I sort of have already.

Push Campaigns

Until now, Wave only had 2 types of push notifications: direct messages, or (if you’re a promoter) someone submitted plans in your region.

This week, I updated that system to allow for custom scheduled notifications.

The big use case for this is letting people know there’s a new version of the app available.

We can now schedule a notification for everyone on the app that there’s new features available (like Anonymous Posting) if they update.

And of course, there’s some marketing and retention motivation there as well with regularly scheduled, gentle reminders to open the app and see what’s happening 😉

But scheduled push notifications are just another reminder of how these systems get so complicated.

The actual sending of a single notification to a single person isn’t so bad, just need the title and message, and a way to identify the recipient.

Blasting a notification to everyone can get a little trickier.

On-demand is easier - fill out the notifications details and click send, find all the users and push their notifications.

Right now there’s a reasonable amount of active users on the app so it wouldn’t be so hard to go through everyone and send a ping, but I went ahead and made sure that doesn’t happen.

Instead of notifying everyone at once, it breaks the notification into batches.

When there’s 1000s of users to ping at once, this will become necessary, so might as well chunk things up from the start.

Additionally, the goal is to not have to click any buttons to send a notification, it should be happening automatically daily.

So now there needs to be a scheduling system regularly checking for when to send notifications and who gets them.

I went with a pretty not fancy approach here, and just have a regularly running scheduler that checks every few minutes to see if there’s any notifications that have been scheduled for sending.

It’ll work for now, but probably not the most optimal.

A queueing system would be much more helpful, but luckily there isn’t a massive amount of notifications to manage yet.

It was fun to see a basic scheduling system finally working, though.

With everything in place, I was able to setup a recurring notification to ping my app every 5 minutes and let it run for a day to make sure it worked reliably.

280 notifications later, the scheduling system is looking good so far.

Something Big Is Happening

The viral AI article of the week is Matt Schumer’s post:

Something Big Is Happening

A personal note for non-tech friends and family on what AI is starting to change.

shumer.dev/something-big-is-happening

It is an interesting read, highly recommend it.

AI is getting better and better at software, and a faster rate.

GPT-5.3-Codex did meaningful work on it’s own training and development.

AI is helping to improve AI more and more.

Humans can’t process exponential growth, but we’re in the middle of it.

It won’t be an overnight change, but it will be a “fast gradual” change.

But it will primarily be in the digital space for now.

Humans will continue to be the bottleneck of bringing or preventing change in the physical world until robots reach the same level.

A few cycles of robots building robots, connected with the minds of the latest AI models, and exponential growth will move into “meatspace”.

Smarter than a human with a gymnast’s abilities:

It’s slow until it’s fast.

And that’s it for this week. Big changes for me and the newsletter, and AI keeps doing it’s thing.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

*The views expressed in this newsletter are my own and not those of my employer.

The first full week of re-employment was a success.

And AI continues to reach new levels through claws and increments.

Plus what’s leftover when the internet is all agents?

Topics:

• Time-Blocking

• Newsletter Ideas

• More Claws and Iterations

Time-Blocking

The first week back to work was a success!

Luckily the transition so far hasn’t been so bad.

I accidentally prepared for this transition last year when I finally came to the conclusion that sticking to a proper 9-5ish schedule would be helpful.

Since then I’ve tried to keep normal work hours of Monday-Friday like I still had a job, despite having nearly complete freedom to work whatever hours I wanted, as long as I was available to clients during their ideal hours.

And my new position allows for a good amount of remote work so working remotely still feels like the norm.

The time in office will be the biggest change.

But I’m looking forward to that too.

Nice to change up the workspace and enjoy the benefits of being in an office again.

Overall the theory of a full-time job being essentially the same as taking on a big client is holding mostly accurate.

And making that change proved easier than I thought thanks to keeping a routine even when I didn’t have to.

If you ever find yourself working on your own, I would highly recommend setting a regular schedule even if you have the freedom to not need one.

Newsletter Ideas

I’m still working on readjusting to the new newsletter approach as this has been a staple of every week for over 2 years now.

So far this format is feeling like a win.

A shorter format for now, a few quick updates, but still get to keep the streak alive.

Saturday morning maybe the norm too, but we’ll see.

I never quite committed to pre-preparing content during the week and scheduling for Friday, but it’s not a bad idea.

I tend to make mental notes of what I see during the week and then spend a bunch of time getting distracted while I try to remember what I saw so I could reference in the newsletter.

I could easily upgrade that system to not only save the bookmark or idea when I see it, but also write a quick paragraph about it.

By Thursday night I could have something good to schedule.

Or I could write a quick recap on Saturday.

Only week 2 of figuring out this new schedule, plenty of time to iterate.

For now it’s short Saturday editions.

More Claws and Iterations

OpenClaw is still plugging along, filling up my X feed with images of it’s creator working on various machines with various claw appendages and Mac Mini’s.

Twitter tweet

I can’t find the post, but it claimed 5% of all GitHub activity right now can be contributed to OpenClaw bot contributions which is wild to think.

One thing was Cursor and other apps using AI to generate the code, but taking it to the next step is now just letting the AI have free reign to run the repo entirely.

And while this is happening, both OpenAI and Anthropic released new versions of their main models: GPT-5.3-Codex and Claude Opus 4.6.

It sounds like Codex is doubling down on developing smarter, still taking it’s time to think long about the task at hand, but producing solid solutions even if it’s not the fastest.

Opus 4.6 takes a different approach by spinning up swarms of other AI agent models to delegate and manage tasks to accomplish the job instead.

Codex is producing more solid code, but Opus is much stronger at producing beautiful UI design.

I have access to Opus 4.6 but the rollout for Codex 5.3 has been a little wonky.

I don’t see the option for 5.3 in the Codex desktop app yet, and the cloud doesn’t explicitly say which version it’s using so it my be available in browser.

I have a few side-hustle items to catch up on this weekend so I may be testing out both models out in the near future.

But I came across a post that highlights an idea I keep coming back to in the wake of all this agentic coding.

Twitter tweet

It may not be too far off that we don’t really need apps any more.

You talk to your computer or phone or smart glasses or whatever, and the AI/personal assistant handles it for you.

It doesn’t need a website, it only needs data and a way to interact with another system, and the event gets booked, the item ordered, the request submitted, all in the background.

Twitter tweet

Games will stick around too, because humans will still want to be entertained, so those apps will remain visual, built for people to actually interact with.

But those games will certainly have APIs working underneath, too.

Games and APIs all the way down.

What else could you ask for from the internet?

And that’s it for this week. Big changes for me and the newsletter, and AI keeps doing it’s thing.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

*The views expressed in this newsletter are my own and not those of my employer.

Big changes this week, for me and the newsletter.

And AI continues to AI.

Topics:

• Re-Employment

• Newsletter

• MoltBook

Back to Work

I am officially back in the working world.

As of Friday, I’m an SMTS at Salesforce.

There’s a number of reasons for the change, and I’m sure I’ll be unpacking them all slowly over time, but I can sum up the main ideas of the decision.

Basically I realized I was spending way too much time and energy optimizing for money.

I had finally settled on the strategy of landing one big client to create a nice foundation to build on.

A big client comes with a bunch of constraints - timelines, priorities, quality expectations, all traded in exchange for a large payment.

This was starting to sound like a job…

Except a job offers way more benefits.

So why not take a job at a place I enjoyed working, with people I’d be more than happy to work with again, working in a system I already know, making a paycheck that shows up as long as I’m doing quality work, and closes the loop of 24/7 stability searching.

It’s a different perspective from the outside looking in.

But felt like the best option for the time being.

So for the near future, I’ll be engineering full-time again.

There’s still plenty of side-hustling to do, though, so stay tuned.

New Newsletter

So what happens to this newsletter?

I’m not too sure yet.

I’d like to keep it going, but it may take a little different shape.

This is a Saturday edition so that may become the norm, but it may return to Fridays once I’m back in the routine.

It also may be a little shorter or less frequent.

Originally started the 5 bullet format because the business mastermind I was in recommended stealing the idea from Tim Ferriss’s 5 Bullet Friday newsletter.

I liked the idea and just picked up the habit.

Even borrowed the 5 bullet part.

But my 5 bullets tend to be pretty dense.

Once I get going, depending on the topics, I actively have to fight the urge to not have every bullet be 2000 words.

I see a few options to better work into a full-time schedule since my Friday mornings are no longer as flexible.

I could embrace the density with less bullets, or stick with 5 bullets but much lighter content.

I’ll try a few formats out and see what sticks.

Friday? Saturday? 5 bullets? 3 bullets? 1 big bullet?

Expect variations in the future, but in the meantime, let me know what you think!

OpenClaw

I’ve been meaning to get ClawdBot up and running for the last couple weeks now, but still haven’t done it.

In the time that I first saw it online, and mentioned it in this newsletter, ClawdBot became MoltBot and now seems to have settled on OpenClaw.

But it’s been a while since I’ve seen anything go this viral in the tech world.

Every other post on X is about ClawdBot, and so many thumbnails and titles on YouTube trying to jump on the “ClawdBot” trend, with the term reaching over 4 million searches on YouTube alone in the last month.

OpenClaw is really the first fully autonomous AI agent capable of running on your computer and doing whatever you tell it to do, and sticking with that task until it’s complete.

For example, this user set it loose in his computer and the bot setup it’s own account on Twilio so that it could call the user with a voice agent and ask him what to do next:

Twitter tweet

And now someone or some bot has created MoltBook, a social platform for all the OpenClaw bots to connect with each other socially and share tools.

Interesting, entertaining, slight concerning.

I’ll be anxious to see how this Clawd Saga evolves over the next few weeks.

Another step closer to AGI.

And that’s it for this week. Big changes for me and the newsletter, and AI keeps doing it’s thing.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

*The views expressed in this newsletter are my own and not those of my employer.

Deep into multi-agent management this week, as I ramp back up into proper modern software development.

Plenty of AI terms ahead.

Topics for this week:

• Parallel Parallels

• Multi-Agent, Multi-App

• Codex codex Codex

• ClawdBot

• Human Emulation

Parallel Parallels

I’m not quite to the pro-levels of running 15 agents at a time like the creator of Claude Code, but I did ramp up in agentic development this week.

I had GPT-5.2 work on a number of tasks to help with Wave while also learning how to run more agents at the same time.

Cursor automatically handles “best-of-N” development for you where you have one prompt, and it spins up 5 copies of GPT to all attempt the same task.

After they’re done, you get to pick the best output.

But what I really wanted to do was work on different things at the same time.

Cursor doesn’t handle that automatically outright.

If you try to open the same project in 2 different Cursor windows, Cursor will automatically switch back to the currently open project.

So there’s no way to run 2 different prompts on the same project with Cursor.

Which is probably good, that could create some nasty situations of one AI not realizing another AI is changing things.

For true parallelization, either some sort of virtualization would be necessary, like Cursor running in the cloud and working on a different branch, which may be possible with the browser-based version of Cursor.

But the multi-Cursor window solution on one computer requires using git worktree.

Git lets you branch code and keep changing in separate instances, but a branch is part of a single project.

Worktree lets you copy the entire project and treat it as a separate project with separate branches.

With a new worktree, it’s possible to open a new Cursor window, and have another AI get to work on a completely different prompt, not the 5 agents working on the same project from before.

Taking this to the next level, it means it’s possible to have 2 Cursor windows doing “best-of-N” on 2 different prompts.

Potentially 10 instances of an AI working at the same time.

Parallel parallels.

And good bye monthly token usage.

Maybe not, but I would imagine running that many agents at a time will immediately hit at least a few dollars worth of API compute in a single session.

I’m still early in my multi-agent development days so I’ll be trying out what works best for me in terms of management and budget, but this is one potential path.

Multiple Cursors working on multiple prompts.

Multi-Agent, Multi-App

Another way of having multiple AI agents work in parallel would be to use different programs at the same time.

For Wave, I was limiting myself to Cursor for development, but I did end up setting up the Claude desktop app so I could try out their new Cowork feature.

It took a few attempts to get my subscription reactivated and connected with the desktop app, but once reinstated, I had Claude Opus digging through files on my computer and reorganizing Wave screenshots and other random images that have been piling up for over a year now.

All this was happening in-between sessions of waiting for GPT to finish working on the current development task in Cursor.

And I was impressed with the results.

Cowork neatly cleaned up the files on my desktop.

I didn’t get much further than the desktop cleanup with Cowork, but off to a good start with Claude.

At one point during this, I introduced a third AI app because I still have the ChatGPT app running.

So while Cursor was developing, and Claude was organizing, I ended up asking ChatGPT a separate development question, totaling the active agents and apps to 3.

This multi-agent workflow could really be pushed even further because so I’ve only done this with dedicated apps.

There’s multiple CLI tools and web-based tools that could be brought into the fold and increase the parallelization to new levels.

Really the task I’ve had in mind for this is Codex.

GPT-5.2-Codex is supposed to be good for long-running tasks, and OpenAI provides a pretty generous quota for using it in the browser, so Codex in the browser will probably serve as a perpetual code review and refactor machine, looking for obvious and easy fixes on big projects like Wave, and doing so all within the cost of the subscription, no extra API costs needed.

So this week, I started to get a taste of the multi-agent craziness we are about to enter as these AI models and tools get better.

We’re practically already there, the real bottleneck is coming up with the ideas and orchestrating it all.

More to come on multi-agent workflows, but for now Cursor, Claude, and ChatGPT is a very solid start.

Codex codex Codex

Digging through agentic recommendations on X, I see mostly Claude Code and Codex as the preferred tools.

Determining the context of “Codex’ is a little tricky right now, though.

There are basically 3 versions of Codex from OpenAI.

First, there’s the Codex CLI (command line interface) that you can download and run on your computer much like Claude Code.

Then there’s codex the AI model, like GPT-5.2-Codex.

Finally there’s the https://chatgpt.com/codex version of Codex which is basically an extension of Codex CLI except that it runs Codex in the cloud for you in it’s own virtual machine instance so that it can process in the background.

As you can probably see, the hard part here is that these are all interchangeable to some degree.

“Using codex” could refer to the locally running Codex CLI that uses an API, but that doesn’t mean it’s using the codex API model, just the Codex app.

It could be Codex using the normal GPT-5.2 model, not the codex model.

On the other hand, using Codex could refer to using the codex AI model in some other harness like Cursor, since that is also an option.

Or it could be both, using the Codex app to run the codex model.

One or more forms of Codex doing development somewhere.

On top of that, there’s the confusing distinction of using Codex at what cost?

Using AI in the browser is generally free but using AI through the API costs per API call.

However, OpenAI provides a quota included in a paid subscription for Codex usage that doesn’t use API costs.

So the Codex CLI can be setup to use an OpenAI subscription for Codex-based development cost, which is separate from the API costs of using the codex AI model.

I had been avoiding using Codex altogether because of all this.

I didn’t want to accidentally start racking up API costs because I let something run over night and it didn’t go as planned.

But I’m pretty sure I’ve gotten a grasp on it now.

Learning Codex CLI can be configured to use my subscription quota, means I’m not stuck with the browser to take advantage of that quota, and can worry less about token costs.

It’s much easier to kick off a scheduled script in a terminal than it is to automate things in browser, although browser automation is getting easier by the day…

As if there weren’t enough AI tools to use, Codex is another one to add to the list, but seeing it referenced so much online, I figured it was finally time to give it a shot, and after some digging this week, I feel comfortable doing just that.

I’m sure that was perfectly not confusing learning how to setup codex with Codex or Codex.

ClawdBot

When I was thinking I was limited to using Codex in the browser only, I was debating ways to automate browser interaction.

There’s a few options out there now, even OpenAI and Perplexity released AI-integrated browsers for just such use cases, though I think they are severely limiting Atlas’ and Comet’s abilities because they don’t want to make it even easier to create spam through these tools.

Anthropic’s Cowork feature comes with a browser extension so it can take control of the browser as well.

But a new tool popped up on my radar this week that takes the browser automation idea, and even the Cowork concept to a whole new level.

ClawdBot is an open source alternative to controlling everything on your computer through AI models like Claude Opus or GPT.

Once installed and connected, you can message your agent through Discord, Whatsapp, Telegram, Slack, etc., and it can do what you tell it to do on your computer without you having to be there, including operate the browser for you.

I’m not quite ready to give an AI full control of my computer yet, so this may be more of a VM situation to test it out, but I would be super excited to be able to delegate tasks on my computer through messages.

Better yet, could set up the automations I’ve been thinking about but haven’t taken the time to properly setup.

Just have the agent figure out how to navigate through the machine for me and do the task.

Really powerful stuff.

ClawdBot may be a good weekend project so more on that if I do get it up and running.

Human Emulation

I saw that an xAI engineer was on a podcast and a few days later announced he no longer worked at xAI, but I hadn’t seen the reason or if it was just coincidence.

Sounds like he revealed the future roadmap for xAI so he was asked to leave.

Twitter tweet

A “full human emulator” is the plan.

I’m not sure leaking this is really a shock, this seems to be the general plan for most companies building AI.

Create software that can act like a human.

It’s certainly an interest spin on the concept, though.

Human emulator is a new one for me anyway.

But it makes sense, it’s concept I’ve been thinking about for a while.

“MacroHard” is Elon’s plan for an entirely AI-operated software enterprise run by human emulators.

And each week it’s feeling like that could be closer to reality.

Grok 4.20 hasn’t released yet, but supposedly it won a stock trading competition against other AI models.

Rumors are the GPT-5.3 is possibly coming out next week and seems to be performing very well.

Even an update for Gemini may be right around the corner too.

Lots more options about to be added to the human emulation toolbox.

We’ll see what comes out next week.

And that’s it for this week. Lots of agent coordination and model management and potential human emulation.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

Not a lot of development this week for me, but plenty planned for next week.

And a very busy week in the AI development world.

Topics for this week:

• Wave Refresh

• Custom Automations

• Agent Options

• Code Building Code

• Claude Cowork

Wave Refresh

Wave has taken a backseat for a while now with a few other projects taking priority but that’s about to change.

There’s been a growing backlog of items to update or fix, so next week I’ll be cutting that backlog down as much as I can.

Mostly admin items like user and location management, but a few other features are on the list with improvements to notifications, social posts, and of course, bug fixes.

I always enjoy working on Wave once I get back into the swing of it, so I’m looking forward to jumping back in and knocking out what I can.

Plus this will be a nice excuse to test out some of the newer coding models since I haven’t been as deep in development lately.

Anxious to stretch my agentic coding capabilities and see what Cursor and maybe Claude Code can do in a short amount of time.

Custom Automations

I haven’t mentioned n8n lately because I haven’t really touched it in the last few months.

I’ve thought about it quite a bit, but there’s a few things that cause me to talk myself out of going back to an automation platform.

Cost is part of it, having a dedicated machine to run it is another, but like I’ve mentioned in the past, I’d rather be in the code.

And n8n is “no-code”.

Similar to the financial app vs spreadsheet comparison last week, n8n is the paid budget tracking app that’s limited in features.

There’s way more flexibility in building something custom than building something custom inside another platform.

And maybe this n8n aversion is really just driven by the current Claude Code hype.

The only hang up with Claude Code is a scheduled task to kick off the terminal command, but Anthropic might have solved that problem with the most recent release.

ChatGPT does has scheduled tasks, but they’re pretty straightforward - basically repeat the specified prompt at the designated interval, and they exist in the background.

But Codex or Claude Code have a more agentic approach that expands being “do this thing at this time” so really I guess this is about scheduling custom agentic activity.

And that concept makes n8n feel obsolete.

You don’t need an automation platform when you have an engineer to delegate tasks too, even if that engineer lives in the cloud.

Along with Wave, agentic automation is also on the list for the next few weeks.

I’ll report back if I find anything exciting in the world of scheduled agents.

Agent Options

There will always be a debate on which model is the best at the moment, and right now GPT-5.2/Codex and Claude Opus 4.5 seem to be at the top.

Some are saying Gemini 3 is good if you can keep it in-check with proper prompting.

Grok 4.20 should be released soon and it’s inventing new math apparently.

Supposedly OpenAI has a new model on the horizon that might be even better than the unreleased Grok 4.20.

And the obvious answer to which model is the best is the most unfun answer.

The answer is… it depends.

Each model is good at different things, and people have different use cases.

The big debate right now is Opus 4.5 vs GPT 5.2

Twitter tweet

Crazy to think that’s the big debate.

Already passed the point of arguing that AI will ever be good at code, now it’s do I want for GPT to be slower but make less mistakes, or use Opus 4.5 to be faster and need to check more of its work.

It’s really down to personal preference.

Getting back into code next week, I’m sure I’ll have more opinions on this.

I’ll for sure be resubscribing to Claude and trying all these flows for myself.

Excited to see the results.

Code Building Code

Things are really accelerating at this point with AI writing more and more code.

Cursor is working on building a web browser, and they let AI build a majority of it.

Twitter tweet

Anthropic launched a new desktop app feature called “Cowork” where it runs in the background and can interact with your computer.

Apparently they let Claude Code build all of Cowork:

Twitter tweet

An entire desktop app feature built in a week with Claude Code.

I’m sure humans were involved to some extent for both projects, but a majority of the work is being done with AI taking the reigns.

Things are not going to slow down any time soon.

OpenAI has already said their last few new models have been improved by the existing models doing a majority of the coding.

More code is going to lead to more code.

Claude Cowork

This isn’t Anthropic’s first attempt at letting Claude control your computer, but it sounds like this attempt is much better.

You can download the Claude app, which now features Cowork, point it at a folder on your desktop, and it can get to work reading, creating, updating or renaming files.

It’s finally time to clean up all those documents and screenshots hanging out on the desktop.

After that, Cowork can be connected to other apps on your computer or even given access to the web through their Chrome extension so that it can looks things up and get additional context for the task at hand.

Super promising tool that I just setup so I’ll be looking forward to having it get to work on some things on my machine.

I could certainly use some help reorganizing files and folders.

Before today, I’ve been using the ChatGPT desktop app for a few months.

It has been surprisingly convenient to be able to Cmd + Tab to GPT at any time to ask a question or ask it generate something, and maybe even more powerful is being able to use the Cmd + Shift + 2 shortcut to take a screenshot of whatever window I was just looking at so that I can ask GPT to look at it and answer questions.

The ChatGPT app has some connectivity to other apps too, and it can be connected to Cursor or the Terminal or a few other coding related applications to work within that connection, but otherwise, it’s sort of self-contained to just working within the limitations of the GPT conversations.

The Cowork feature of the Claude app opens up way more possibilities beyond the GPT app.

The Cowork tab also adds a nice interface for managing tasks which GPT has indirectly, but Cowork makes it much more upfront.

Cowork tab of Claude Desktop

Looking forward to testing both desktop apps out side by side to see which one I end up using more.

More updates on Claude Cowork soon.

Let me know if you end up trying it out!

And that’s it for this week. Plenty of AI development happening now and planned for the future.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

Slow start to the new year.

Which isn’t such a bad thing.

Plenty to plan for, but still reflect back on.

Topics for this week:

• Keeping it Simple

• Trade-Offs

• Daily Useful Ideas

• Quiet Updates

• AI Tokenomics

Keeping it Simple

Looking back at 2025, I realized I was trying to do too much.

So one goal of 2026 will be to try to keep things simpler.

And ironically, AI can help with that.

AI loves to tell you what you want to hear, but really AI will do you whatever you tell it to do.

So I’ll be using AI to push back more and help me keep things in check.

Given a standard prompt with enough context about the ultimate goal, I can use AI to make sure that any new business venture or change in process lines up with that ultimate goal, and have it question whether something is worth the effort or not.

“Here’s the goal, here’s the new idea, ask 3-5 targeted questions to make sure any new approach actually helps accomplish the goal. Accomplishing the goal is the highest priority, pushback if this idea doesn’t directly improve progress toward the goal.”

If you tell AI to pushback, it can be very critical if you want it to.

It can also be questioning while still remaining constructive.

The here would be to help full flesh-out any new ideas before completely diving or switching gears.

Have AI make sure any changes actually align with the ultimate goal, or at least justify the change.

With the right prompt and context, AI should be able to provide just the right amount of pushback to keep things simple and focused for 2026.

Trade-Offs

The last 2 years have been a good lesson in the classic trade-off.

Money now or money later.

There’s 2 popular approaches in the software world currently, according to all the online gurus.

Make money upfront with an agency, or make money later with software.

And both are gambles.

Agencies are hard to scale because most of the work is manual or custom, and requires people to do the work.

Sign up for too many clients, and an agency can pay well, but also become way more than a full-time job.

Software, on the other hand, is easy to scale.

Built correctly, it costs basically nothing to add 100+ users overnight.

But if nobody wants that software, then all that building is a waste of time and money.

So the common suggestion is to balance both.

Start an agency to make money upfront, and use that money to build software.

When done properly, it’s a lower risk approach.

Balancing trade-offs.

A regular income and the ability to build without the pressure of instant success.

It makes sense to me.

And it’s possible to really simplify things using that model.

Swap out the agency with a job.

Less flexibility in a full-time position, but more stability.

And still the ability to build on the side without needing financial results.

A different set of trade-offs but a similar outcome.

And surprisingly simple…

Daily Useful Ideas

I’m still keeping an eye out for “micro-saas” ideas that could be fun to run on the side, despite my efforts to simplify things.

Small and manageable is fine, if it ends up being fun projects or things I’d want to use, especially things that would be useful daily.

I have small list at this point, most of which lend themselves to Chrome Extensions.

One of which I was reminded of today.

The “Political Filter”

I’m on X just about daily, but I’m no means a power user.

I follow a good amount of accounts, mostly tech related, and tend to stick to my “Following” tab.

Granted, the “For You” tab has gotten much better lately, it seems much more relevant and much less school fights, but I still try to avoid it.

But even within the “Following” tab, there’s an issue.

I follow some smart people, but half the time they are rage posting about politics, and that’s not why I follow them.

I frequently get the urge to unfollow but I usually don’t because I want to stick with the bigger picture.

A happy middle-ground, though, could be a simple Chrome Extension that uses GPT or some other model to quickly scan the post and semi-censor it.

Read the feed quickly and swap the post content out with “politics blah, blah, blah”.

They still get their freedom of expression, and I get to gloss over it.

I’m really only on X for tech so that’s really all I need to see.

If I’m feeling political, then the feature could be turned off or back on with the click of a button.

And this could probably extend to other features like “spoiler filter”, “language filter”, etc, but politics filter is the most relevant for me right now.

This is one project that’s high on the list for an AI agent to work on in the background, but let me know what you think.

What other are some other tools that are really simple but would also be really useful every day?

Quiet Updates

Seems like the LLM world has calmed down with the recent holiday stretch, which is probably a good thing.

With rumors of the “996” work schedule (9 a.m. to 9 p.m., 6 days per week), I’m sure there are some very tired engineers that needed time off.

But one update slipped through with MiniMax’s M2.1.

Twitter tweet

This is another open source model that could be run locally, or used for very cheap in the cloud.

It’s joining a list of open source models that generally perform well, including Kimi K2, GLM-4.7, and Qwen Coder.

These are really great for when you hit your Claude Code or Cursor caps, you can switch to running locally, or super cheaply through a service like OpenRouter.ai, only costing $.30 in and $1.20 out.

We may already be at the point where it’s getting difficult to even distinguish between these models any more.

The reasoning models will be a little different but as far as chat and code generation goes, it could be pretty easy to rotate between all these cheap or free models and drop the crazy expensive monthly subscriptions.

And MiniMax M2.1 is another one that could be in that rotation.

AI Tokenomics

Looking back at last year, I knew AI news was a regular thing for me, but there was one particular aspect of news that stuck out to me.

I’ve been tracking pricing longer than I realized.

Token prices are a core component of AI decisions, and it’s something I’ve been paying attention to for a while.

Maybe not a shock since I’ve talked about needing an LMArena Leaderboard for AI pricing before, but thinking about it more, just that one facet alone could be a pretty valuable thing to focus on.

AI pricing is really it’s own field.

Cryptocurrency coined the term “tokenomics” but it’s surprisingly fitting here since all AI models are priced per 1 Million tokens in and out.

I mocked up a landing page idea for tracking AI tokenomics, but nothing officially released yet.

It’s one idea that’s stuck with me for a while recently, and seeing that it’s been on my mind for a while is a good sign.

If I had to simplify down to just one particular thing, this could be one thing that could lead to enough news and change to not get bored.

Also need a good alternative term if “Tokenomics” has already been taken by crypto…

And that’s it for this week. Let the simplification begin.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

It was bound to happen.

A rare, not-Friday edition to look back at 2025 and wrap up the year!

Topics for this week:

• The Beginning

• The Middle

• The Ending

• A Lot of AI

• Next Year

The Beginning

Looking back at the newsletters from the beginning of the year, it’s hard to believe it’s been almost 12 months already.

A lot of the same themes and ideas stayed pretty constant through those months.

I was in the early stages of Wave and AI Automation in January, and they both remained active up to the present, and will into the future.

I also started the year by re-introducing a 9-5 schedule for myself and that was really foundational for the entire year.

I was circling the idea of automated content, specifically shorts, and only recently settled on a system I like that I happened to find through tracking n8n content.

I began embracing AI news back then, and I’ve enjoyed tracking it all year.

And with all that, this newsletter began settling into my core pillars of content.

But something else I realized, which probably isn’t so surprising, is I was trying to do a lot.

Lots of projects, lots of side-hustles, lots of distractions.

Developing Wave, implementing auth, migrations, and permissions, maintaining various Divs-related sites, building and managing a cold email system, handling inbound and deciding what not to respond to, recording and editing YouTube videos, iterating on titles and thumbnails, optimizing the channel without over-optimizing it, recording and editing podcasts, experimenting with AI-assisted workflows, tracking token pricing and model economics, orchestrating video processing pipelines, mapping out financial systems and runway, scanning AI posts and job listings without jumping at each one, re-evaluating management vs developer paths all while still consuming more YouTube than I should.

The year started out a busy one, but it’s settling down with a less hectic ending.

The Middle

As the year progressed, so did Wave, Divs, YouTube and AI Automation efforts.

Wave took up a bulk of the work, building a working version for the iPhone, then getting it distributable so others could download it, then setting up the services behind it for push notifications and messaging, then figuring out video processing and operational cost management, then a few more issues and requirements from Apple before finally publishing a test version through the App Store.

During all this, I did cut back on the cold email system even though I got it successfully working.

I decided I’d rather do YouTube to attract Automation clients instead, and I thought I could use that traffic to also build a Skool community at the same time so I slowly shifted to that approach.

Divs also picked up a new client in the middle of this going on so plenty of work was piling up.

By May, I realized I should probably cut back on every-other-day YouTube videos on top of trying to get Wave wrapped up for a release.

Interestingly, around the same time I was considering one video per week, YouTube accidentally took my channel down due to a false “community violation” flag.

Thankfully they quickly brought it back up after I submitted a repeal, but YouTube almost ended that endeavor for me entirely.

Around the same time, I was getting close to finally publishing Wave despite pretty minimal testing.

Once a group of people started getting their hands on it, I could see more issues being dug up.

I thought some of the issues were with the TestFlight version, but the issues persisted through an official App Store release.

So this led to a stretch of serious bug squashing.

Some were easy to fix, but others stuck around for way too long because I was relying way too much on AI to solve problems back then.

We might be getting to the point just now where AI can solve some pretty complex puzzles, but back then I was a little too optimistic.

It took a few days of refactoring the entire app and still seeing the same issues to realize that AI had become too much of a crutch for me to solve the problem.

With a few hours of additional setup to change some build processes, I manually solved the major issues that had been bugging me for weeks.

I remember letting out a huge sigh of relief knowing Wave had finally reached a stable build.

And in the process of not using as much AI, I renewed my interest in programming.

I became excited about coding again.

Ironically, I had already setup a path with AI Automation that would not be using much coding at all so that may have created a conflict that I later realized.

But the middle of the year was an eventful one.

Publishing Wave was a major milestone, creating a fully-featured social media app became the biggest success of the year.

The Ending

With Wave in a good place toward the end of the year, the focus shifted to AI Automation.

YouTube growth with the intent of attracting automation clients became highest priority, and the secondary goal was building a community on Skool based on the traffic.

In addition to publishing videos, I committed to streaming Monday through Friday, which I really enjoyed and looked forward to, and during those streams even built a crazy “online course generator” system that pushed the limits of n8n.

But I also learned 2 lessons:

• Streaming is fun but it takes a very long time to grow.

• I’d rather build software than automation workflows.

These 2 ideas will undoubtedly have an impact on next year.

Nearing the end of the year, I had already started to take inventory of where things were at, what’s working and what isn’t working, and these lessons helped make decisions for what’s to come next.

The year started with trying to do too much, but the year ended with simplifying things way down.

Right now this newsletter is the only true constant that has remained through the entire year (despite missing a Friday publication).

Wave, Divs, YouTube, and Skool are still on the list but final decisions are coming soon.

It’s been an eventful year, and I’ve learned a ton.

With everything that happened this year, it will undoubtedly reshape my perspective for next year.

I’ll be anxious to get into that more soon…

A Lot of AI

AI played a huge role in everything I did this year.

Without AI, Wave would be much simpler than it is.

Many of the projects I worked on wouldn’t have happened either.

I truly learned to work with AI this year.

And I did use quite a bit of it:

Cursor Recap 2025

Apparently I used 20 different models, switched primary models 13 times, used Cursor 198 days this year, and generated 739.3 Million tokens.

Cursor Model usage

My top models were:

• GPT 5

• GPT 4.1

• “Auto” (let Cursor pick the model based on the message)

• Claude 3.5 Sonnet

• Claude 3.7 Sonnet

• Composer 1

• Claude 4.5 Sonnet

• DeepSeek R1

• Claude 4.5 Opus

• Gemini 2.5 Pro

And I chatted with GPT a good amount too.

GPT Usage for 2025

5121 messages and 621 chats.

Kind of surprised they didn’t include a token count.

But that’s a lot of messages.

So I used a lot of AI.

And it’s wild to think when the year started, I had just gotten access to gpt-o3-mini, was using Claude 3.5 Sonnet for all coding, and DeepSeek R1 nearly caused the collapse of the entire AI market.

Since then, GPT has now gone through o3→ 4o → 4 → 4.1 → 4.5 → 5 → 5.1 → 5.2

Claude went through 3.5 → 3.7. → 4.1 → 4.5

And multiple versions of Gemini and Grok in the same time as well.

Large Language Models have progressed rapidly in 12 months and it doesn’t seem to be slowing down any time soon.

I’m sure it will be hard to believe what they’re capable of by the end of 2026 but I’m looking forward to more new releases in the coming months.

And I work with AI daily at this point.

Feels like soon enough we’ll be working for AI.

Next Year’s Plans

Obviously AI will be a cornerstone for any plans I have in the new year, but I am still in a bit of a holding pattern with the end of this year.

By the end of January I should have a pretty clear sense of direction.

Either way, I know more building is involved.

In 2025, I had a tendency to build things that were optimized by external influence or short-term ROI potential, but next year I’m hoping to return to engineering things that I want to engineer because they are fun and interesting.

I’ll still be using AI to do a lot of that coding, but I’ll be supervising and taking the wheel occasionally as well so I can keep my technical ability as sharp as I can.

And of course, YouTube will undoubtedly have a role to play as well.

What form of YouTube video is still a little up in the air, but most likely it will be short form because I believe a personal brand is a really important piece of financial independence, and consistent short content is both a manageable and effective way to build a brand at the moment.

So software and YouTube will remain core pillars for next year.

More details on that as the year progresses.

Lots learned this year, and plenty more to learn next year.

Thank you for reading as always, and Happy New Year!

Hopefully it was a good one, and hopefully next year is even better!

And that’s it for this week. Happy New Year!

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

Quiet week personally, but AI continues to do it’s thing.

Topics for this week:

• Software Engineering Manager

• Less Code Red

• Google’s Lead

• Opus Decline?

• So Many Tokens

Software Engineering Manager

It’s only been 2 and a half years since I left full-time employment as a software engineering manager.

At that time, humans still did all the coding, and I think I used GPT maybe once a week.

I was super excited at the potential of GPT and other LLMs, but it wasn’t quite to the point of everyday usefulness.

I would remember to ask it questions occasionally or maybe play around with a simple API connection, but nothing I was dependent on.

Fast forward 2.5 years later, and now when I go to build something, the first thing I do is map it out with AI and have it mock up the code for me.

I don’t write nearly as much code as I used to, I do make manual tweaks here and there, but I produce and review way more code than ever before.

I can spin up an entire codebase with a single prompt.

With AI becoming extremely capable at coding, being an engineer suddenly becomes much less about writing the code and more about managing it.

Depending on the model, basically all engineers now have an infinite army of junior- to senior-level developers at their command to produce whatever they are told to produce.

A core part of the job has been outsourced to a virtual agent available 24 hours a day, 7 days a week, that doesn’t care about a promotion.

Since most of the writing of code isn’t necessary any more, the engineer is left to focus on higher levels of the software lifecycle.

The engineer is still engineering the code, but in a completely different way, and spending more time managing that code then engineering it.

It’s an interesting time to be alive to see this transition from something that seemed so universal be completely shifted in a matter of months.

Pretty soon the term software engineer will mean something completely different than it did just 3 years ago.

Less Code Red

OpenAI continues to roll out the features as they try to compete with Google’s gradual AI dominance.

GPT-5.2 rolled out last week, and they updated their code specific models “Codex” with GPT-5.2-Codex.

This is seems like an incremental update from GPT-5.1.-Codex-Max which just came out a few weeks ago, but they claim improvements in cybersecurity understanding, as well as better tool calling and ultimately it’s better at working on large scale tasks for long periods of time without hallucinating into unnecessary changes.

The previous Codex version claimed the ability to focus on a task for up to 24 hours so an improvement past that would be truly impressive.

On top of that, OpenAI also launched GPT-image-1.5 to compete with Google’s Nano Banana Pro.

GPT Image 1.5 vs 1

It’s wild to see the comparison directly.

The old model image on the right doesn’t look bad, but it does feel a little off, missing some sort of realistic element.

The image from the new model on the left looks much more like a real photo.

Not sure I’d spot that as AI on first glance.

One thing GPT models in general are particularly good at is following instructions.

This example was pretty impressive:

Instructions for a 6×6 grid of unique items

Given the list of 36 items, GPT was able to generate an image that contained all the items as specified.

All items correctly rendered from the list above

Following instructions is easy to take for granted until the model stops doing what you told it.

Really impressive to see that level of depth.

I’m not sure they’ve beaten Nano Banana Pro yet but I’m sure they’ll be getting close as they continue to play catch-up.

More releases from OpenAI coming soon I would imagine, but they weren’t the only ones to release more new AI models this week…

Google’s Lead

After having their search dominance shaken to it’s core, Google has responded impressively well to the AI race.

Gemini 3 Pro is a super powerful model that upgraded an already leading model of Gemini 2.5 Pro.

And offer a smaller, faster and cheaper alternative, Google updated their flash series of models with Gemini 3 Flash on Wednesday.

The Pro model does the heavy lifting, using a good amount of reason tokens to think through solutions before spitting out an answer, and Flash models can use reasoning tokens as well, but that’s not really the intent.

Given the name, it’s primary intent is speed.

And it’s already reaching the top of the leaderboard, not far behind Gemini 3 Pro.

gemini-3-pro rank 1, grok-4.1-thinking rank 2, gemini-3-flash already rank 3

With the release of flash, Google has 2 models in the top 3 for text and in spots 4 and 5 for WebDev.

gemini-flash-3 in 3rd for Intelligence, 2nd for speed, 5th for price

Artificial Analysis has Gemini 3 Flash ranked highly for Intelligence, Speed and Price as well.

The flash models aren’t as great for chatting or deep development work, but they are really great for digesting data and quick answers.

A good all-around model for an API.

And not too big of a price jump from flash-2.5: up $.20 for input and up $.50 for output.

We’ll see what other models release in the next few weeks, hopefully these poor engineers get a chance to take a break for the end of the year.

But for now Google is doing it’s best to be a top contender in the AI race.

Opus Decline?

Saw this on my feed the other day:

Twitter tweet

Basically someone saying they’ve seen a noticeable decline in the quality of output from using Claude Opus 4.5 through Claude Code.

This seems to be a recurring pattern with a lot of the giant AI models.

I remember people talking about a similar issue with GPT 5.

It started out super smart when people had early access, but when it went to public release, it seemed to decline in performance quite a bit.

And a few others have been claiming the same with Claude Opus 4.5.

It seems noticeably less performant.

There’s some crazy inference architecture happening behind the simple idea of sending a request to an AI and getting an answer back.

I would imagine a number of services and connections are happening as part of the process to run text through a Large Language Model, and with the insane amounts of traffic these companies are getting, I’m sure they having to shift resources around constantly.

But it’s an interesting pattern that there may be a slight marketing angle to it as well.

Deploy more resources to the newest model so it performs the best and looks the best when people are reviewing.

But after a few weeks, slowly start to shift the resources away from the big marketing focus and on to other projects…

Doesn’t seem completely out of the question, especially in the current war for GPUs.

I haven’t been using Opus 4.5 much lately so can’t speak to my own experience, but I hear almost nothing but good things.

Hopefully this is just a temporary, even subjective observation.

And maybe just aspect of the AI world where sometimes the AI won’t be as smart as other times which I have noticed.

Certain times of the day, GPT gets noticeably slower than other times.

Huge demand for these services so I can’t image the continual struggle to support them.

Some day I’d love to learn how these AI services operate at scale but for now I’m happy to speculate based on social media comments.

And hope that temporary AI underperformance remains temporary.

So Many Tokens

I caught this clip on X and it’s kind of hard to comprehend.

Twitter tweet

Sam Altman talks in rough numbers:

• 10 trillion tokens per day from one big AI company

• a human produces about 20,000 tokens per day

• 8 billion people on Earth * 20,000 = 160 trillion

One company produces almost as many tokens as all humans produce naturally.

So companies are individually reaching the same information output levels as all of humanity which is wild to think.

Before the AI crazy, we were in the “big data” era.

Companies were producing more data than they knew what to do with, but couldn’t possibly consume it all, so it became “data science” to try to figure out what data was actually important and what could be ignored.

Then AI comes along and compresses all that data into something that you can talk to.

But it can also now produce more data than ever before.

Almost 2 steps forward and 1 step back.

And it’s interesting to hear the shift in thinking about information.

I do see this becoming a subtle transition where things are looked at from an AI perspective.

He didn’t say words spoken, he said tokens.

Humans produce 20,000 tokens per day.

Weird to hear now, but I bet it might be the norm in the next few years.

We’ll see.

We’re many trillions of tokens away from that happening soon, but it may be sooner than expected.

And that’s it for this week. AI is shifting engineering and maybe perspectives altogether.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

The Code Red response.

A fork in the road.

And new results.

Topics for this week:

• The Wind Down

• Fork in the Road

• Brand Pause Benefits

• Red Alert Response

• Up and Coming

The Wind Down

Only 3 newsletters left for the rest of the year!

I haven’t done it yet, but I plan to go back and review where things were last year so I can recap in one of these last few emails of 2025.

There are projects that will continue for next year, but a number of projects that won’t, so I’m compiling that list.

Divs Design will continue on, supporting websites, SEO and Wave for now.

And maybe even with a new landing page to support it.

But some of the AI and Automation endeavors may take a backseat.

Too many things happening with not enough results.

Divs Cloud may make a return instead, with a renewed focus on SaaS, aided by SEO.

And YouTube may transition to more software and AI, but less Automation.

Lots of possibilities as usual, just starting to map out what those may look like for2026.

So more to come on the final list, but looking forward to fully re-evaluating while trying to wind down a very eventful year.

Full list hopefully early next year.

Fork in the Road

Somehow I’ve been living financially independent for more than 2.5 years now.

It doesn’t feel like it’s been that long, but at the same time the pressure has been increasing.

Which is another reason for re-evaluating my current list of projects from above.

Good to determine what’s working and what isn’t.

Which isn’t out of the ordinary, I probably re-evalutate a little too often.

But there’s also another reason.

I may have the opportunity to return to a full-time role, and if I take a new position, I will have make the list much shorter.

Going back to the working world would take up quite a bit more time day-to-day and leave a lot less time for so many projects, but the ones that do make the cut could be done properly, without the constant pressure that it needs to highly efficient and highly profitable.

It’s been a great 2.5 half years of working for myself, but it may be coming to an end soon.

At least in the full-time sense.

I’ll for sure continue to have work for myself on the side, but I would also have a boss and Monday-through-Friday responsibilities taking priority again.

Nothing official at this point but there’s a high likelihood in the near future.

More to come on what’s next for the new year.

Brand Pause Benefits

I’ve pretty much paused all content at this point while figuring out what’s happening next.

But despite the pause, there’s still growth!

Yet again, the beauty of the almighty algorithm working it’s magic.

Actually multiple algorithms at this point.

All thanks to the impressive Blotato system from a few weeks ago.

I’ll be anxious to get that automation repurposing TikToks to multiple platforms again.

Specially since posting to TikTok was actually boosting my YouTube growth.

I managed to post all of 6 TikToks but still gained 2 subscribers and more than 6000 views on YouTube because of those repurposed TikToks, not including the 20+ followers and 10k views on TikTok.

If I keep following the Blotato creator’s advice about posting 2 TikToks a day consistently for 6 months, it does seem like 500,000 followers/subscribers across all platforms is a very realistic outcome.

And on top of that, YouTube itself is still growing my Skool community, even when I haven’t posted an n8n video in a few weeks now.

Nearing 800 members just from YouTube traffic alone.

So despite pausing most of my social media efforts, I’m still seeing the benefits.

Hard to deny the impact of regular posting and letting internet traffic do it’s thing even when I’m not actively pursuing growth.

Red Alert Response

Shortly after OpenAI’s internal “Red Alert” memo, rumors were they already had new AI models ready to go to try to bring them back to the top of the leaderboards.

Sure enough, yesterday they tweeted that GPT-5.2 Auto, Instant, Thinking and Pro were all rolling out to everyone with a paid account, and free accounts today.

Sporting some impressive evals, it does seem to be closer to the top than 5.1:

Twitter tweet

It’s getting hard to tell if these benchmarks are valuable or not anymore, but GPT-5.2 Pro (High) seems to be the highest performing on the “ARC-AGI-1” test:

Twitter tweet

The green dot farther to the right seems to be Gemini 3 Pro and GPT is performing higher than that so it may be the leader in “General Intelligence” for now.

But with a quick release, it’s also a little more expensive than GPT-5.1, which probably isn’t too surprising.

gpt-5.2 pricing

Up $.50 for input and up $4.00 on output from gpt-5.1 which isn’t a terrible hike.

What’s really interesting here is that gpt-5.2-pro released with the API when gpt-5.1-pro didn’t have an API yet.

But gpt-5.2-pro is also $21 for 1 million tokens in and $168 (🤯) per million tokens out.

That may be partly why they were hesitant to release 5.1-pro’s API.

gpt-5.2-high has made it to #2 in WebDev

These new releases have moved gpt up in the LMArena rankings so it may gain them back some traction.

But it also may kick off another round of updates from all the other model leaders.

Sounds like Grok 4.20 is already in the release rumor-mill so there may be another new one in the top spot here soon…

Up And Coming

I’m fairly tuned into the AI market but not quite to professional levels yet, so I completely missed this one.

Looking at the GPT 5.2 benchmark posts and responses on X, I saw someone point out that the “poetiq” model was missing from the evals.

I had never heard of poetiq.

Apparently it’s a new model created by former Google Deepmind developers who broke off to start their own project.

Google Deepmind might actually be the secret AI winner right now because it’s very limited access but seems to test completely off the charts on all the evals.

The problem is it isn’t publicly available so the results can’t be independently verified, just have to trust Google and those with access.

So it would not be surprising to see a team that worked on the potential AI lead create a new model that’s already a top performer.

Twitter tweet

This is a week old so naturally it’s completely out of date, but Poetiq looks to be right on par with GPT-5.2. for the ARC-AGI-2 test.

Poetiq next to GPT-5.2 Pro (High)

We’ll see if more comes out about Poetiq, but it might be one to keep an eye on as the AI race continues.

And that’s it for this week. New opportunities, new releases and new models.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

Re-evaluating old approaches.

Re-mixing new approaches.

Lots of ways to solve the business puzzle.

Topics for this week:

• Accidental Opportunity

• Micro-saas Solutions

• Code Red Memo

• Too Many Cursors

• Advent of AI

Accidental Opportunity

I’ve been thinking more and more about building in the background.

I built Wave primarily with AI generated code, and lately the models have only gotten more trustworthy.

Of course, that wasn’t the case when I was working on Wave, and still isn’t quite there yet.

I usually sit there in Cursor and AI Studio, double-checking the result after every generation, but Wave is a very complicated application.

Not the best project suited for agentic cloud development.

Something simple like SEO generated articles could be a nice project that can run over night or in the background and build momentum over time.

I’ve accumulated a nice backlog of random URLs that I thought were great ideas at the time and some are in various states of completion.

The other day I was reviewing spending habits and found a Wix plan that I had forgotten about for one of those random websites that I thought was a great idea at the time.

The plan automatically renewed earlier this year so even if I canceled it now, it wouldn’t officially shut down until July.

So why not take advantage of a sunk cost?

With a renewed interest in SEO, I may be able to have AI push that plan from sunk cost to recovered cost, or even profitability if handled correctly.

With about 50 well structured, valuable articles, it sounds like Google will pick up a website as a valid source and start recommending it for search traffic.

If all of those articles have affiliates links, it only takes a few higher priced items to earn some money back and turn a dead website into a tiny passive income machine.

I’m really not so worried about it making money, but if it didn’t cost anything to operate that would be a plus, but also to have something that can near autonomously grow traffic to itself is the real power.

That strategy will be very useful for other projects.

I accidentally created the perfect reason to test out programmatic SEO with a realistic goal of generating a couple hundred dollars in 7 months with minimal time and effort.

More to come on that SEO opportunity.

Micro-SaaS Solutions

The SEO approach I mentioned above isn’t exactly a new one.

A huge portion of the internet exists to do exactly that.

Pump out articles that Google likes so it recommends them as search results.

Fill the articles with affiliate links.

Make money when people buy from those links.

The alternative strategy to this is still to programmatically build SEO, but to create traffic for my own services.

Rather than take an affiliate payout of 1-3% for other products, why not attract visitors to something I create and would get 100% of the purchase.

This is indie hacker’s approach to development.

Build a small tool, get some traffic, improve traffic based on questions and feedback.

And the way AI is going, it can manage a good portion of this entire process.

It will still need human supervision, we’re not quite to AGI levels yet, but it’s getting closer with each new upgrade.

For now, I see a scenario where AI agents are building and tweaking these small tools a few sessions at time, then switching to research mode where they gather feedback and research SEO opportunities for topics related to the tools, then generate a few articles that can help drive traffic, then repeat the cycle.

A couple nights a week, a series of projects could be looping through these steps and making progress in the background, patiently waiting for review in the morning.

And this approach attempts to solve the two main reasons why a business doesn’t succeed.

Bad product or bad distribution.

AI built micro-saas has potential to handle both of those problems by constantly monitoring feedback to improve the product, but also constantly working to create more traffic based on what people are actually doing.

A modern approach to making 10 viable software bets.

The true 10x engineer.

This may be where things are headed in the future so I will keep you posted.

Lots more SaaS to come.

Code Red Memo

The hot topic this week has been OpenAI’s slightly panicked internal memo.

With 4 major model releases in the last few weeks, OpenAI may not be comfortably sitting as the clear leader of the AI race at the moment.

Gemini 3 Pro, Grok 4.1, and Opus 4.5 now claim the top spots on many AI leaderboards as the top performing models, and variations of GPT 5.1 sit close by but not in the lead.

Apparently Sam Altman sent out a “Code Red” email warning to the company that they need to refocus on GPT becoming the top model again, and deprioritize their other projects.

OpenAI has branched out to a ton of other products other than AI models recently, so it’s reasonable to think their efforts are getting spread a little thin.

At the same time, I do wonder if there’s a limit to how many people can work on improving models at the same time.

Could create a “too many cooks in the kitchen” situation.

But rumors are, they already have a model that’s better than Gemini 3 Pro and will undoubtedly be releasing this new version very soon while they re-prioritize taking the AI lead.

Then we’ll see how the other big players respond.

Grok 4.1 just came out 2 weeks ago, but already seeing Grok 4.2 rumors so xAI is ready to keep pushing to new levels.

I thought there would be a few weeks without any major AI news after all the recent upgrades but maybe not.

OpenAI may be pushing more big release as their try to fix this “Code Red” situation, and the big competitors may be responding just as quickly.

Too Many Cursors

As an AI consumer, I can benefit from things like OpenAI’s “Code Red” situation above.

New model takes the lead?

Just switch to that one.

Cursor has made that easy enough, thankfully.

And that may serve as my main development hub.

That’s certainly what they’re trying to create, with some of their recent updates to make multi-agent coding a core part of the user experience.

But as a developer I can’t help but want to have my own system to handle things how I want to, outside of Cursor.

I dabbled with the idea of building my own AI pipeline which is probably what will end up happening eventually but recently I’ve considered switching to something a little more flexible.

Claude Code is one such alternative, that’s Terminal-based instead of an IDE like Cursor, but it sounds like it takes a little hacking to get Claude Code to use other AI models than Claude.

And of course, big AI players have their options too, with Google’s Gemini CLI and OpenAI’s Codex CLI, which again I believe allow for multiple models but take some config to switch out of their preferred options.

But there’s a couple options not owned by a big AI company, which are OpenCode and Droid CLI.

Neither are particularly tied to one model, which makes it easy enough to switch on the fly if a new model becomes the leader in AI coding output.

I believe Droid CLI even has a built in /switch command so they have already prepared for model hopping.

As I talk about jumping between different AI projects, I’m seeing a benefit to moving to something terminal based and not having to have 15 instances of Cursor open to manage everything.

Plus the ability to script actions to start at certain times is not really an IDE kind of thing, more of a CLI thing.

That’s probably the biggest motivator for considering a switch.

Queuing up tasks in various projects that can run at a certain time every night.

Back to that whole 10x talk from earlier.

10x the work but not 10x the IDE windows to rifle through.

I haven’t run the commands to install either of these options yet, but I had the quick start guide up and ready to go earlier this week.

Droid CLI will certainly be making it onto my computer sometime before the next newsletter.

I’ll be sure to let you know how it goes.

Advent of AI

Somehow another year has passed and another Advent of Code has started!

AoC has been running strong for 10 years now, providing a series of daily coding puzzles leading up to Christmas (although this year is only 12 puzzles instead of 25).

And these puzzles aren’t very easy.

Complex problems that require a good understanding of advanced programming concepts to come up with an efficient solution.

“Leetcode” at it’s finest.

Not sure I’ve every made it past 3 days of puzzles if I’m being honest, but it’s always fun to get a few solved.

And there’s now 10 years of puzzles to practice programming if you’re looking to brush up on your coding skills.

This year in particular though, I wanted to test a few of the top AI models and see how they fair in coming up with solutions.

GPT 5.1 Pro is supposed to be surprisingly good at puzzle solving so I wanted to see how it does with complex code problems like these.

And I’m sure I’ll try out a few myself.

I’ll report back on how the puzzle solving goes next week.

Let me know if you end up solving some yourself!

And that’s it for this week. Let’s see if AI can solve the puzzle of business.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

Topics for this week:

• Divs Design Web

• Short Growth

• Vibe Coding Freebies

• Monthly Goals

• All the Upgrades

Divs Design Web

I didn’t expect AI to be right about this one, but it was.

I’ve been meaning to update the Divs Design website for a while now, but I didn’t think it’d take that long.

Have Gemini generate a new page, update the images and make sure the links all make sense and it’s good to go.

GPT put the website update item at 2 days.

And it’s been a little longer than that.

To be fair, I haven’t had a ton of time to work on it, but it’s taken more than a few hours to update.

Now I think it’s about there.

Shippable.

Jay | divs.design

Full‑stack engineering for creators and founders. Custom mobile apps, SaaS platforms, and premium business websites — built fast.

divs-design-web.vercel.app

Needing to edit some new images, ask for design advice, update sales copy and tweak the layout and order and offers has led to quite a bit of back and forth with multiple AIs before getting to “good enough”.

But I think it’s finally there. For now.

I’m pivoting from Webflow this time around because I don’t think I need it for a landing page like this.

AI can handle the fancy animations at this point and I’m finally moving beyond the “high end productized web design service” trend.

So Vercel is good enough to ship with.

I also was trying to force some sort of competitive SEO term into divs.design and had no idea what I was doing so I clunkily forced the term “web design package” in there as unnaturally as possible.

SEO can be important for something like this but not a term like that.

Leading to a few reasons to redesign this page.

And it’s just about there.

I should be switching it over to use the proper domain next week.

More Divs to come.

Short Growth

So far so good on the shorts strategy.

There’s enough AI news happening that it’s pretty easy to just talk about a new update quickly and be done for the day.

Stilled divided on the right strategy though.

It only takes a few minutes to successfully record and post 1 video to TikTok, which then automatically turns into a post on YouTube Shorts, Instagram and LinkedIn, but I still managed to miss an upload yesterday for Thanksgiving.

Got sidetracked on Wednesday when I was supposed to be pre-recording a bunch for the rest of a busy weekend.

So I may use both strategies depending on travel, holidays and whatever else is going on that week.

But this TikTok approach is much more manageable and seems to be producing decent results.

In the last week alone, TikTok has produced:

• 37 Followers

• 295 Likes

• 9889 Views

A few of those same videos have over 1000 views just by being reposted to YouTube.

And I haven’t even checked Instagram or LinkedIn stats to be honest.

Really I’ve only been looking for notifications about comments and so far the only comments have been on TikTok.

But a handful of shorts are leading to noticeable uptick in activity which is cool to see.

22 subscribers to YouTube in the last 7 days which is a decent bump.

Now just need to consistently do that for the next 6 months and I may have a following in the 5 or 6 digits.

Seems easy enough…

Vibe Coding Freebies

At the time I made this video, I had only heard about Gemini 3’s release so I thought it’d be fun to go through a bunch of different ways to start vibe coding.

Cursor has been my main tool for a while, but I know there are a ton of options these days.

Even crazier is that basically all of them are free to use within limits.

So I listed all the ones that I’ve seen or heard about.

I haven’t used them all myself but I’ve talked to people that have or read about them online.

I may end up downloading them all trying them out at some point.

But the list includes:

• Lovable

• Replit

• AI Studio

• Cursor

• Trae

• Windsurf

• Kiro

• Antigravity

• VS Code (with extensions)

• GitHub Copilot (for VS Code)

• Cline (for VS Code)

Not an exhaustive list, but all have free options to spin up a project and use AI to help with coding.

Insane.

Personally I’ve used:

• AI Studio - in the browser, generating code with Gemini 2.5 Pro

• Cursor - weekly if not daily since around October of 2024?

• VS Code - for the last 10 years, this was the only option (I may have used Intellij for a bit in there and Eclipse from time to time because I had to)

• Cline - plugged it into VS Code to see how it worked but Cursor was too convenient

• Replit and Lovable - popped into both of these to see how they worked when they first came out, but quickly realized they were not for me because I’ve been in VS for the last 10 years and I needed more access to all the code

What’s really crazy with all this is that all of the access to this is free.

Meaning it would be easy (and slightly painful) to continually move one project between all of these tools, using up the free usage tier, then moving on to the next free usage tier, rotating until one of them reset for the month.

It’s a great time to be someone interested in code.

I talk a little more about it in the video if you want to check it out below:

Monthly Goals

The mythical “$10k per month” has become the goal standard for hustle bros everywhere.

The vehicle changes but the mantra stays the same.

It’s sounds simple enough - build something that makes $10k per month in recurring revenue and your foundation is set.

A few clients here, a few projects there, $2000 and $3000 at a time, and you’re at $10,000 per month.

Or a bunch of small monthly payments.

350 people paying $29 per month is $10,150.

Or a combination of big and small.

Only 175 community members or courses sold, and a couple service contracts to fill in the remaining $5000 for the month.

Then rinse, repeat, attract, and maintain for the next month.

Or another approach, one big contract that averages out to $10,000 per month.

$200,000 divided into 12 months, after taxes, is about $10,800 per month.

Lots of ways to get to $10k per month.

And most likely it will involve doing things you don’t want to do.

Just a matter of balancing what you want more.

Most hustle bros are starting from scratch so the big contracts aren’t really an option out of the gate.

But given the opportunity I’m sure it would be something they would also recommend.

Lots of ways to set the foundation, lots of numbers to consider.

Maybe more on this soon…

All the Upgrades

The last 2 weeks in AI have been one of the bigger stretches in a while.

It started with Grok 4.1 on Monday, Gemini 3 Pro the next day, followed by GPT 5.1 Pro, then Nano Banana Pro, and finally wrapped up with Anthropic releasing Opus 4.5 on Monday of this week.

And right now, Gemini 3 has taken the lead on LMArena at the moment, but for text only.

Gemini 3 Pro leading text, Opus 4.5 leading WebDev

Opus 4.5 is the top spot for WebDev and so far it seems like Opus is super solid for building websites.

Gemini 3 did a great job as well, it built the initial version of https://divs-design-web.vercel.app/ for me, but Opus really improved the section that highlights Wave:

Wave Mobile Project Highlights

And GPT Pro is a slow deep thinker but it’s not available in the API yet so it’s not being ranked.

But it was crazy to see all these releases happening with the span of a week.

Gemini 3 had been rumored for a while so I’m sure all the other frontier model providers were planning to remain competitive.

I’m curious how far ahead everyone else knows each other’s planned release dates because I can’t imagine the crunch to hit the deadlines determined by your rivals.

Supposedly “966” is becoming the norm in the AI race, working 9am to 6pm 6 days a week, and maybe getting to rest on Sunday.

So hopefully some of those devs are getting a chance to take a break if they want.

But seems like a pretty big jump in processing power in the AI space in general last week over the course of 5 releases.

AI isn’t slowing down anytime soon.

I haven’t pushed these models nearly as much as I’d like to with the holidays and travel and all, but I’m thankful for their work, thankful that I somewhat understand what they’re working, and thankful to take things a little easier this week with Thanksgiving.

Looking forward to stress testing AI much more next week.

Hopefully you got to take it easy and enjoy the holiday this week in the US.

Thankful for you reading too!

And that’s it for this week. 5 major releases in about a week. Hard to beat that.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

We are getting closer and closer to the day that AI will be doing all the coding.

With some huge updates, LLMs aren’t slowing down anytime soon.

Feel the AGI.

Topics for this week:

• Wave User Management

• Content Engine

• The First Two Years

• One Max and One Pro

• More Pros

Wave User Management

Payments are still on hold for Wave at the moment.

Video uploads took priority last week, and this week more advanced user requirements are rolling in.

A normal user can log into the app and start browsing through everything right away.

But we put some guardrails in for anyone that tries to sign up as the “Venue” account option instead of a user.

Can’t have random people signing up as a Venue owner and creating whatever made up venues they want.

Which is a good we had that block from the beginning because a random email address from China did in fact try to sign up as a venue owner, despite Wave only releasing and existing in the US.

But right now the Venue account review process is manual.

Basically we’re expecting someone to sign up as a Venue account and I go into the database and flip their status to active to allow them to continue setting up their account.

This manual process has happened a few times now in the last couple weeks, so it’s starting to make sense to add in a way to allow admins to make that change instead.

Not only does it save me time, but it’s a nice impression when a Wave admin is sitting at that Venue with the owner and they can mark them approved right there in person.

So luckily this feature should be a quick solution:

• Have Claude (or maybe Gemini 3…) whip up a new “User Approval” or “User Management” screen

• Let Gemini or GPT implement the validation and hooks to the database

Since the app release cycle is kind of lengthy, I’ll probably include a handful of small wins and bug fixes to make the effort worthwhile.

If it didn’t take up to a full day for Apple to review new app releases, I would rather go the micro-release route, but since I end up having to wait, it makes more sense to wait on a few things at once.

So a new “User Management” screen and a few other tweaks are on the list for next week.

Minor Wave update coming soon.

Content Engine

It’s finally started.

Only a couple weeks late from when I mentioned it.

I have setup and activated the Blotato repurposing automation.

For 3 accounts for now.

Every hour, one new TikTok post will eventually find it’s way to YouTube Shorts, Instagram Reels and LinkedIn.

There are options for even more platforms, but that’s plenty to start with.

Despite having just set up this system, I’m still caught off guard when I see the posts on other accounts.

It’s great when automations work.

And the math adds up quickly.

I posted 6 TikTok videos this week.

But with the automation, those 6 posts became 24 posts for shorts alone.

Gary Vee would be proud.

Almost the amount he recommends posting in a single day.

On top of that, I managed to post 2 long form YouTube videos this week, and I may have a 3rd ready to go soon.

YouTube is still processing the data, but it’s looking like just this activity alone is already showing a growth spike, which is great to see.

More content and growth to come.

Looking forward to getting this content posting system properly warmed up so I can get the content producing system online and funneling into it eventually.

The growth engine has started.

The First Two Years

Listening to enough podcasts and YouTube videos, I’ve heard plenty of reoccurring themes.

The line that caught me the other day was effectively “throw away the first 2 years of entrepreneurship”.

The first few years are about learning.

Alex Hormozi and Chris Williamson call it “The Lonely Chapter” where you’ve stopped doing what you used to do, but you aren’t good enough to be recognized for the new thing you’re doing.

According to GPT, there’s plenty of VC terms to describe this period as well:

• “The Wilderness Phase”

• “Zero-to-One Desert”

• “Pre-Product-Market-Fit Hell”

All of these talk about the early stages of building where you’re not sure if what you’re building will even work at all.

It can be a long and challenging time but it’s very common for anyone straying off the beaten path.

As I look back at my two years on this path, I can see the bigger pattern.

After a lot of projects, podcasts, courses and cohorts, I’ve gotten a good feeling for what’s working and what isn’t.

Still a lot of work to do, but the path is becoming clearer.

So I can throw away the last two years in terms of building a massive product, but keep the last two years for having built a foundation and learned a ton.

Sounds like I’m right on track.

One Max and One Pro

OpenAI had to keep up with the big releases this week, and certainly did their part.

Last week they released GPT-5.1 which I didn’t even end up mentioning.

It didn’t feel like a huge upgrade to GPT-5 to be honest, but it does seem very quick to respond and still follows complicated instructions very well.

This week though, they pushed GPT-5.1 Pro, which sounds really promising.

Apparently it’s much slower, but is a much smarter “brain”.

Twitter tweet

Some highlights:

GPT-5.1 Pro is a slow, thoughtful “brain”

“…solving problems people thought were out of bounds for today’s AI systems.” is quite a quote.

And this video from Theo talks about how GPT-5.1 Pro was able to solve a hacking puzzle completely on it’s own, which is actually kind of insane:

Sounds like it took it about 45 minutes, but it came up with the proper solution.

I have not seen any model be able to solve puzzles anything close to that yet, so I’ll be anxious to test this out myself.

I’ve had access for a couple days but haven’t gotten to do more than ask it how to come up with a quick approach to automate image-based social media posts.

It did create a multiple step implementation for how to get started and how to evolve the automation which was quite impressive.

But I’ll be anxious to give it some code and see what it does.

And of course they also released GPT-5.1-Codex-Max right before Pro went live.

Codex is there cloud-based model that’s good for coding in the background on it’s own.

Sounds like it can work on a single task for up to 24 hours which is super surprising that it can keep focused for that long without wandering off into hallucination land.

Twitter tweet

I have a number of projects to let it async code for me so I’ll giving this some work to do while I work on other things.

Lots more testing to do with GPT-5.1 Pro and GPT-5.1-Codex-Max.

But if that weren’t enough…

More Pros

Gemini 2.5 Pro only officially released earlier this year but it feels like it’s been much longer than that.

I’ve slammed 100,000s of lines of code into it to figure out problems and come up with new solutions and architectures, and overall it’s served as my primary coding engine.

And this week Gemini 3.0 Pro officially released - the next big version.

Previews from last week have been really impressive, and again, I haven’t had a chance to really push it with additional 100,000s of lines of code yet, but it did pretty impressively produce a working n8n workflow when I gave it the same request of automating image-based social media posts.

And it do so considerably faster than GPT-5.1 Pro.

Beyond that I haven’t been able to do too much else so I’ll be anxious to do some coding with it as well.

But generally people have been excited.

Landing pages and games seem to be the go-to choices so far.

Twitter tweet

Being able to build video games so reliably really highlights how well it’s handling complexity because video games are incredibly complex to build.

Lots of potential here.

And if that weren’t enough, Google also release Nano Banana Pro this week.

It, too, is mind-blowing:

Twitter tweet

And this guy:

Twitter tweet

Casually adding 14 characters into one image.

So much potential, maybe too much potential.

Many things to test, looking forward to building with all these new pro models.

You can try out both of Google’s new models at aistudio.google.com (Gemini 3.0 Pro for free, but Nano Banana Pro requires a paid API key).

Time to get back to work, lots of coding and recording to do!

And that’s it for this week. Big updates with Gemini 3, Nano Banana Pro, GPT-5.1 Pro and GPT-5.1-Codex-Max.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

Had trouble starting this newsletter.

Then had trouble finishing it.

Lots to cover.

Topics for this week:

• Video Upgrade

• First Ranking

• Episode 13

• The Last Piece

• The Next Number One

Video Upgrade

I’m slowly becoming a video expert.

Watch out TikTok.

After a hard lesson on video processing failure, of course…

I had a service running on a single AWS App Runner instance handling the upload and reformat flow of videos for Wave, and it worked great when I tested it.

But I missed the obvious test:

Concurrency.

I was only uploading 1 video at a time from my phone, when I should have been slamming it with at least 3 in a row.

Last week, reports started rolling in of video uploads randomly failing.

Checking the logs, I could see the issue: signal killed 

The single instance of App Runner was running out of memory trying to process multiple videos.

The easy fix was to upgrade to a bigger instance of App Runner.

But a couple days later, the same issues kept coming.

So it was time to rethink the strategy.

After talking it through with Gemini and GPT, the upgrade plan was clear.

Use App Runner for routing, offload the heavy video processing to Lambda.

Lambda isn’t always “warm”, so it has cold-start delays, but it has 2 major advantages over App Runner:

• It scales instantly

• It can run multiple instances at once

App Runner can scale too, but it’s way more work and way more expensive and probably overkill at this point.

So the new architecture looks like this:

1. App Runner gives the phone a signed URL to upload the video

2. Supabase fires a webhook to Lambda when the file lands in storage

3. Lambda compresses the raw video and publishes the processed video

This means App Runner stays light and fast, waiting to help users immediately upload their videos so they can go back to using the app and not have to sit and stare a loading spinner.

They get notified later (ideally less than 30 seconds) when their video has been added to the Live feed.

And it seems to work!

I spam-uploaded back-to-back as fast as I could and didn’t see a single failure, so this system should be able to handle a healthy amount of video traffic for a while.

Still plenty of room for optimization, though:

• tuning ffmpeg for faster compression with less memory usage

• streaming data between AWS and Supabase instead of full file downloads

• caching and parallelizing processes if we want to get really crazy

But for now, Wave video is in a good spot.

…until Wave goes viral.

Then it’s Kubernetes time.

First Ranking

Not too long ago now, one of my website clients asked about SEO services, so I offered a starter package and got to work.

Admittedly, when I built their website, I was still mastering Webflow at the time, and wasn’t really focused on SEO at all.

Re-reviewing their site, it was very obvious how to take their overall SEO score from a 54 to a 100, according to Ubersuggest, by cleaning up a lot of html structural issues.

That was the easy part, if not a little tedious.

The next part is the long slow part to SEO.

Ranking.

I laid the foundation for Google to trust the site.

Now it’s time to show up in search results.

And I’ve already had 1 success!

GPT helped come up with an SEO friendly article that is now ranked as search result 28!

Not exactly first page or top 5, but this was a pretty big unlock.

I’ve been following SEO content for a long time, but I hadn’t really had much success in getting anything to rank anywhere near the top, so this was very exciting to see.

Ubersuggest identified the term that had the best opportunity to rank, and GPT made it happen.

So now the real work begins.

I showed the client that one of their competitors has over 500 articles on their site and get about 12000 visitors per month, while my client has 8 articles and gets about 800 visitors per month.

Even with just 20 more ranked articles like that, we’ll be well on our way to 3x’ing their organic traffic.

But seeing this process generate results was really exciting.

I have yet to SEO my own websites so I’ll be anxious to implement a similar process and play around with some side projects as well.

It was a path I started going down a while ago, but naturally got sidetracked for some other shiny objects.

And was also slightly deterred by the SEO space freaking out that AI and Google had destroyed SEO as we know it.

But turns out it’s very much still alive, just maybe a little different.

So I’ve already drafted the next article for the client, and I may be updating some of my other websites to follow a similar SEO process.

And hopefully many more rankings to come.

Episode 13

I previewed it last week, but The Dev Sync, episode 13 is now live:

Some good clips to start this one off with bots stealing window seats on planes, captcha being our last defense and hoping the pizza protocol is a real thing in the future.

It was another fun episode so if you do check it out, let me know what you think.

Episode 14 should be a heated one as well.

What’s the role of Junior Developers in the time of AI?

No plans to record it yet but I’m sure we’ll sneak at least one more episode in before the end of the year…

The Last Piece

Looking back at this long, winding road, I’m starting to realize something.

I basically have everything I need to build that real portfolio of SaaS apps.

For years it felt like I was missing something.

Some secret “real founder skill” I didn’t have yet.

But between all the projects, client work, experiments, and rabbit holes, I’ve accidentally built out almost every piece of the puzzle:

• user systems

• sessions and auth

• messaging

• video processing

• image handling

• likes, posts, feeds

• notifications

• deployments and infrastructure

• content engines

• and now SEO

All for website and mobile development.

At this point, the biggest missing piece is payments.

I’ve added Stripe Links to websites, but that’s not exactly scalable.

Once I fully integrate Stripe APIs into a SaaS or app cleanly, I’ll have built basically all the pieces I need.

I could build anything.

Not just one app, but that full portfolio of apps.

And those apps could actually make money.

I knew payments was the last piece so I tried to speed run the implementation in one evening (both for the experience, and for practice with upcoming Wave payments) and I got decently close on Friday night.

I whipped up a super stripped-down version of “Corporation as a Service”.

The goal was to get something live that could offer enough value that it would be worth paying a cheap monthly subscription.

Sample “Corp as a Serivce”

Basically predefined personas that you could chat with for various business advice.

It could easily be customized to talk to famous business personas and have GPT or other models imitate their advice.

Straightforward but potentially useful.

With the good enough prototype on my laptop, I tried to dive into integrating Stripe.

But unfortunately I needed the app to be deployed so that I could get real URLs for the Stripe webhook, so I went to Vercel to get live.

And Vercel did not like the vibe-coded monorepo which resulted in about 30 minutes of trial and error commits with Cursor to finally get the configs right for Vercel to properly build and deploy the web app inside the monorepo.

But quickly after that, I got a little hung up on the business side of things.

I couldn’t decide if I should use a separate business entity for this Stripe setup or stick with Divs Design for now.

Finally GPT convinced me to keep it separate, but by the time I got to integrating the live site with the new Stripe account, it was already 2 AM.

So I decided to pause for the evening.

I may put a few hours into this sometime soon to finally wrap up that last piece, because it feels very close.

It may only take a couple hours to complete the last big missing piece.

Full founder mode will be enabled.

The Next Number One

Gemini Pro 2.5 has been my go-to AI for vibe-coding the last few months now, and I’ve been excited to see rumors of the big upgrade coming soon.

Gemini Pro 3.0 may be just around the corner.

Online speculation is that it’s already live in the Gemini mobile app, but it’s still called Gemini Pro 2.5.

Some of the posts from users on X.com that do have access to 3.0 have been quite impressive.

Like this incredibly high-end website animation:

Twitter tweet

Or this super solid modern landing page:

Twitter tweet

Or a full working Minecraft coliseum demo:

Twitter tweet

No official announcements yet, but some people have posted they already have access on aistudio.google.com as well.

I’ve been periodically refreshing today but no luck so far.

Hopefully I’ll get a chance to test it out next week 🤞

I’m sure this will be the next number one model in the rankings.

And that’s it for this week. Video processing, SEO, the last piece, and the next number one.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

Wave payments on hold this week to focus on video uploads instead.

Plus it looks like n8n doesn’t want workflows to be paid services.

And are we in a bubble?

Topics for this week:

• Rebuilding TikTok

• Reading Handwritten Documents

• The Dead Internet Discussion

• n8n to Micro-SaaS

• The Bubble

Rebuilding TikTok

I have new respect for video-based apps.

Trying to replicate TikTok’s multi-billion dollar video platform is not easy to do.

I was happy to get an initial version working in Wave that looked like TikTok-style video swiping, and even managed to setup a cheap way to make it work with real videos, but that cheap approach may already be reaching it’s limit.

AWS App Runner is handling all the video processing on the backend at the moment.

Users upload their video to the virtual machine and it strips metadata and compresses it down so it can be quickly loaded during swiping.

But it seems that basic system may already be reaching it’s limits with a few users.

I’m trying to balance decent user experience with cost effectiveness while Wave isn’t making a ton of money to fund its infrastructure, but it could prove challenging.

The lone App Runner instance probably won’t cut it.

I upgraded the App Runner VM itself, from .5 vCPU with 1 GB of RAM to 1 vCPU with 2 GB of RAM, but concurrency is still going to be an issue.

A few video uploads at the same time or back to back and that will start to cap out again.

Probably going to need implement a queuing system first and foremost.

Chatting with a few AI models, it sounds like App Runner can lead the processing but also offload to Lambda if things get a little busy.

And users will have to wait a little.

The next biggest performance improvement would most likely be chunking video data so it can process immediately.

Lots to look into but I’m looking forward to building something that can reliably manage a burst of upload traffic on a busy Saturday night.

More to come on Wave’s slow replacement of TikTok and Instagram.

Reading Handwritten Documents

I made a video about this process a little while ago, where I was demonstrating how to use n8n to read the contents of a PDF and fill out other data columns within Monday.com using the content, saving the effort of having to manually enter customer data.

But I managed to completely forget to try that system with handwritten documents.

If the doc is nicely printed, n8n can turn the PDF to text with no problems.

If it’s handwritten, n8n completely gives up and returns no text from the document.

So rather than having to install another community node to convert the PDF to an image, I finally found a way to have gpt-5-mini read the document instead.

I knew GPT could handle it because I took the handwritten document and uploaded into the ChatGPT app on my MacBook, and GPT read the text back to me right away.

Unfortunately it’s not as straightforward to do in n8n.

Took a couple different HTTP Request node attempts to finally figure out how to have n8n call OpenAI to read the contents of a document.

But this video covers the full process.

This could be a really handy one even for things outside Monday.com.

The Dead Internet Discussion

We managed to not wait another year for Episode 12 of the Dev Sync after our triumphant return with Episode 11 last month.

Jim, Eric and I recorded a new episode last night about the “dead internet” theory.

It’s a theory that came about around 2016 that all of the internet is just bots talking to bots.

Surprisingly though, last year may have been the first year that bot traffic overtook human activity on the internet, leading by 51%, so bot activity is certainly on the rise thanks to LLMs.

We dive deeper into that, and try not to get too distracted with AI related topics like it’s impact on the future of the internet and how work is changing in general.

It was another quick 50+ minutes, and the full episode should be up on Monday.

Stay tuned!

And here’s Episode 11 in the meantime:

n8n to Micro-SaaS

I’ve been circling around the micro-saas idea for a long time and I may have a reason to dive into it soon.

I accidentally essentially built a SaaS using n8n when I was working on the course generation system during livestream sessions.

With the completion of version 1 of that project, I got curious about turning it into a full platform that people could pay to use.

My first thought was I could just vibe-code a front-end for users to sign up and pay to access the app, then the backend would kickoff the workflow to generate the course.

I’m not sure at this point, but this approach may fall into a bit of a gray-area.

According to their sustainable use policy, you can’t use n8n for commercial purposes if the value of the service derives entirely or substantially from n8n functionality.

Turning a workflow into a SaaS could potentially violate that policy since a substantial amount of value would be operating through n8n.

At it’s core, I think they are trying to prevent people from reselling access to the n8n platform, like vibe-coding a new website that just uses n8n nodes underneath and charges for access to a custom no-code platform.

And arguably, it would still take a substantial custom web app to turn an n8n workflow into a SaaS.

It would require a full website for users to signup and login, a backend like Supabase to manage the session and store user and transaction data, and a Stripe integration to handle payments for the service before ever kicking off the n8n workflow from a webhook.

But most of the value of the app would derive from the n8n workflow creating the resulting deliverable.

So to be safe, I don’t know that I’ll be connecting an n8n workflow to a website and charging for access unfortunately.

The alternative will probably look something like Go microservices hosted in Lambda, replacing the n8n workflow.

Kicking around a few candidates for this approach so more to come on that, but I know for sure n8n will remain just an automation platform for now.

The Bubble

The hot debate right now in tech and the markets is if AI is a bubble.

A few weeks ago, Sam Altman himself admitted we may be in an AI bubble.

And the common comparison is to the “dot com” bubble of 2000.

But I heard an interesting point the other day about why this AI bubble is different from the dot com bubble.

According to Gavin Baker, Managing Partner and CIO of Atreides Management (nice Dune reference), during the 2000 bubble, there was a metric of something called “dark fiber.”

Fiber optic cables are laid down in the ground for internet traffic, but if the cables aren’t lit up for use, it’s just dark fiber.

At the time, 100s of thousands of miles of dark fiber was laid for the expected internet boom, but at the peak, 97% of the fiber in America was still dark.

Only 3% was in use for the internet.

AI, right now, has the opposite problem.

No company can get enough GPU’s to run their AI models.

GPUs are melting from use instead.

There’s no shortage of traffic.

The AI bubble could be a different kind of bubble, though.

NVIDIA has positioned itself nicely at the center of the AI race as the primary manufacturer of the GPUs needed for AI models to run in the cloud, so if NVIDIA makes a mistake or runs into trouble, it could have massive consequences.

But at the same time, all the other AI companies are scrambling to come up with alternatives so that not everyone is reliant on one primary source of GPUs.

AMD and a number of other chip manufactures are stepping-up for sure, even new ideas for chip architecture and processing concepts are entering the market with all of this pressure, so new options are showing up to spread the stress.

So right now we might be in an NVIDIA bubble.

But only time will tell.

It was really interesting to hear the dot com bubble described with a single metric like dark fiber after looking back 25 years.

I’ll be curious to see if there’s one defining metric for the current AI bubble when looking back 25 years from 2050.

And that’s it for this week. Video uploads, micro-saas infrastructure and bubbles, what more could you ask for.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

Recent reminder - it always takes longer than expected.

Sometimes consistency isn’t enough, but the right consistency.

And an interesting perspective on build Large Language Models.

Topics for this week:

• Under Estimated

• Handwritten Documents

• Scheduled Videos

• Growth Strategies

• The Model Business

Under Estimated

I was looking for regular content, and I thought I was building a weekend project at the suggestion of a dedicated stream viewer.

Two months later, it turns out I built a SaaS that learned how to lecture.

It asks a few questions about your course, then builds the whole thing: slides, narration, lecture, all automatically.

Scoped way down to use only “Apple-style” with minimal text, and sticking with just AI voice narration, I didn’t see the system being that difficult to build.

I was thinking I could knock something out in a weekend.

It was only a few steps: just get the course details, generate slides and audio, then stitch them together into videos with ffmpeg.

After the weekend I thought about building it, I decided I’ll just build it on stream instead.

And quickly figured out it was going to take a while.

I started working on flows for this system around September 10, and finally got a single lecture video produced on October 6, but with manually running workflows and manipulating data directly.

Yesterday, October 30, was the first time the entire system ran on it’s own without human intervention.

I entered the details, selecting the option for only 2 courses for testing, and eventually out popped 2 lecture folders with:

• 2 slides per lecture (title slide and key points)

• a 2-3 minute audio narration for the lecture

• a video lecture that shows the slides while the audio narrates the lecture

• a lecture document to accompany the video lecture

• a list of additional resources for continued learning on the lecture topic

• a set of quiz questions about the lecture

It’s a pretty solid system, and took way longer than expected.

But also does much more than I had originally pictured it doing.

This is what I meant by accidental SaaS.

I think I started pushing n8n to it’s limits with this one.

But it’s exciting to see the whole thing come together after about 2 months of building.

And 2 months sounds like a long time, but this was almost entirely build only while streaming.

I think I tweaked a couple things once or twice off stream.

About 80 hours total, or roughly 2 weeks of full-time effort spread over 8 weeks of streaming.

Much longer than the original weekend hacking I had planned, but really only 2 weeks of full-time building.

Not bad for the result.

It’s been a good project, and nice staple for the stream.

It showed me once again, that something I thought could be built in a weekend, actually took 2 solid weeks of focus.

Even automating things with “no code” takes longer than expected.

Handwritten Documents

Speaking of taking longer than expected…

Last week I set up an automation that would read a customer intake form, find the customer details like name, address and phone number, and add those into the appropriate columns in the CRM Monday.com.

That process was pretty straightforward.

I did over-engineer a little, though.

Using that input data, I generate a fancy report number that combined part of the address, the customers initials, the signing date, and part of a timestamp to guarantee a unique searchable value this: FL-2025-05-04-1234-EL

Spent more time getting that to cooperate than anything else.

But I forgot to test with a document that had handwritten customer info, not printed.

n8n has an “Extract Text from PDF” node but using that with the handwritten document produced no text at all.

Just didn’t even try.

Dropping the PDF into GPT, it read the text right away, though.

So I had to add a second route to the automation so that if the “Extract Text” failed, send it to GPT instead.

But again, this was surprisingly more complicated than expected.

n8n has direct nodes for GPT to read images, but not PDFs.

Instead, this became a multipart process of redownloading the PDF file from Monday.com, uploading the file to GPT so it could read it, finding the proper way to send a request to GPT to read the file I previously uploaded, then delete the file upload afterward since it doesn’t need to live in GPT and take up storage space.

Getting GPT to read a PDF

The alternative would have been to use a non-standard community node for n8n that could convert a PDF to an image, and then send the image to GPT to analyze instead, but I was trying to stick with just using OpenAI if I could since I know they can handle it.

It took about 5 different attempts to finally find the right approach to get gpt-5-mini to read a PDF but luckily only maybe 45 minutes total to get a proper result.

I expected the actual reading of the handwritten document to be a pain, not the process of uploading and reading it to be the real problem.

I didn’t get a video done for this one yet, but it will be high on the list.

Amazing to think a business could easily have GPT, or any other major model, do all the document reading and entering at this point, even if those documents are handwritten.

Scheduled Videos

OpenAI’s Sora 2 video model has been out for a few weeks now, but I finally got around to making an automation for it.

At first I set out to use some of these new video API platforms like kie.ai and fal.ai which advertise convenient and super cheap access to all the major AI video and image models, including Sora 2.

I created a workflow that connected to kie.ai’s Sora 2 and pretty quickly had a working system, but I paused before recording the video to show it.

Sora 2 has been invite only since launch so I couldn’t help but wonder if these third party websites are finding ways to get to these video APIs that aren’t within Terms of Service for OpenAI.

This is the same reason I haven’t fully bought into relying on Apify.com for automations.

These websites work well and are super convenient, but they also feel like a time-bomb.

They aren’t operating within other websites guidelines so it only takes one cease and desist from a huge company to force Apify, or kie or fal to close up shop, and any platform that’s been dependent on those sites now has a big hole in its process.

I couldn’t in good faith recommend kie or fal for now until I know more about them so I stuck with the direct Sora 2 API, which was easy to use (and I was lucky enough to get my hands on an invite code a few weeks ago).

So here’s a simple system to schedule Sora 2 videos everyday:

Finally put my face in a thumbnail again.

Let me know what you think!

Growth Strategies

I spent a good amount of time this week focusing back on “brand growth”.

I’ve talked about it before how the stream alone is not a good short-term traffic generator.

This has been the problem with Twitch since the early days.

The easy part is turning on the camera and being live.

The hard part is growing the audience.

One-per-week YouTube videos are a solid foundation for traffic though, and it’s been working slowly but surely.

If growth is the goal however, it’s going to take more than one video per week.

Or better quality videos…

But at this point I think I need more quantity and data to figure out the quality piece.

Almost 40 long videos is a lot but at the same time it isn’t.

Shorts however take much less time to make and can increase traffic considerably faster.

I’ve been working on ways to automate the shorts process to pull content out of a stream, but of course, it’s taking longer than expected…

But I do have an alternative in the meantime.

Sabrina Ramonov has provided a fantastic guide and service to leverage shorts content much more efficiently.

She spends 1 day per week creating shorts and one long video, manually uploads her shorts to TikTok, then uses her service Blotato to repost those TikTok videos to 9 other social media platforms.

She even threw in some example n8n automations on how to make this happen.

I’ll be exploring this approach more next week while I continue to squeeze value out of the stream, but this process is looking promising.

More to come on this shorts strategy.

The Model Business

We’ve gone full meta (not that Meta) in this one.

I was listening to Theo’s video talking about how OpenAI is burning through cash at the moment, but then Theo referenced a podcast clip with Dario Amodei.

Podcast within a podcast.

Dario was on the Stripe podcast explaining the business behind training a frontier model like Claude.

You can look at training a new huge model like building a separate business.

It costs whatever, $100 million to make a model in 2023.

Then $800 million for a new one in 2024.

Then $1 billion in 2025.

So it looks like money is just burning.

But the 2023 model doesn’t cost much to maintain once it’s built, and makes money once it’s offered as a service.

If they get 2 million users to pay $120 per year to use that model in 2024, and assume maybe 50% margin to operate, it’s $20 million in profit.

By 2025, the 2023 model is $120 million in profit.

The next model has to push monetization a little harder but has more options.

Say it gets 10 million users to $240 per year or they offer Max pricing of $200 per month, $800 million is well within reach.

Even just 10,000,000 users paying $120 per year is $1.2 billion, and the $800 million for training is recovered.

And the pricing becomes more extreme for the $1 billion model in 2025, but ideally its a more capable model that can take it’s time to become profitable, but continue to operate for longer.

All the while the other models are operating in profit, feeding the training of the next big model until the market reaches equilibrium and these numbers don’t scale anymore.

It was super interesting to hear Dario explain it, and I’m anxious to listen to that full podcast, but it excited me because I’ve been thinking about AI models like an enterprise in your pocket - a Corporation as a Service.

It’s both obvious and not obvious at the same time.

Corporations are essentially one big business made up of a bunch of smaller businesses that all try to align on the same goals.

An LLM is essentially a large percentage of an actual business compressed into a single asset.

But that asset still needs humans to turn it into a fully operating business.

Maybe soon, though, one of these models may finish training and actually be a self sustaining business, operating all on its own.

And that’s it for this week. Wave is finally live on iOS and now Android!

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

Turns out AI can still be pretty smart sometimes.

And I may have been a little hard on n8n in the past.

Not sold on AI browsers though…

Topics for this week:

• Events vs Products

• WhatsApp Bot

• AI Browsers

• Accidental SaaS

• Over 100

Events vs Products

Had a really obvious-after-the-fact realization this week while building out Wave payments.

Focusing on giving Wave something to sell before worrying about enabling a full marketplace, I found myself arguing with Gemini about the data models behind the transaction.

I’ve been trying to balance my knowledge of the future vs what is necessary right now.

I don’t want to code things specifically for Wave selling event tickets but I also don’t need to add enough abstraction that I could switch from Stripe to some other payment processor on the fly.

And I was having the same struggle with Products vs Events.

Cover for a bar or Table service sounds like individual products, one time sales with an inventory.

And I was trying to force Events to follow the same structure.

But Gemini was pushing back on that idea, insisting that I needed an events table to associate with the products table.

I’ve learned to doubt AI by now so I wasn’t sure, but I took Gemini’s response and threw it into Perplexity and ChatGPT to see what they said, and they agreed.

I was thinking about things wrong.

I’m not selling the full event, I’m selling a ticket to the event.

So the ticket is the product and that product is associated with the event.

That did make way more sense.

But I still didn’t totally follow that.

GPT ended up suggesting I still treat tickets as products for now.

The extra tables and complexity only become necessary when the ticket sales become complex.

If there’s different tickets, like General Admission vs VIP, for the same event and the event is a recurring event, then things become more complex.

For now, I do just need to be able to sell tickets for one event and one location.

Even multiple different tickets for one location for one event.

Classic over-engineering.

I talked about it a while ago that Gemini surprised me in the past because it would push back on bad practice, and it is still capable of that.

GPT has caught up and even moved things a little further, helping to focus on what’s necessary now, and encouraging me not to “gold-plate”.

So thanks to GPT and Gemini, I have proper understanding of events and products, and a good solution for handling both items for now that’s extensible in the future.

Vibe coding at it’s finest.

WhatsApp Bot

WhatsApp has been a frequent request for n8n integration and I finally got around to making one.

It’s was a bit of a pain to setup the integration because the Facebook developer portal isn’t the most intuitive thing.

Facebook, or technically Meta, makes you jump through some hoops to setup a Facebook developer account, create an app client, create and connect a WhatsApp account, verify your business and username, generate an Access Code, and finally get a test number.

With all that, I was finally able to use the WhatsApp nodes built in to n8n.

But I could see why there’s a little work to get this going.

Automated WhatsApp messages could easily turn from being useful for a business into a nice scamming system that blasts everyone it can find so I don’t envy Meta trying to balance real users with malicious ones.

So I’m sure that’s why it’s a heavily requested topic.

Lots of people use it, but it’s not the easiest thing to use.

And now you can automate your WhatsApp messages too.

I am surprised this didn’t get more traffic, to be honest.

It’s a heavily searched topic and I’ve seen people request it more than anything else, but so far the video doesn’t have great numbers, below 100 views as of now.

I’ll admit the video isn’t great either.

And I haven’t committed to adding stupid faces to every thumbnail so I’m sure that would bring more traffic.

I was in a bit of a rush to get this one done and then it took forever to edit out my phone number so this quick video turned into a 3+ hour endeavor but I was glad to get it out there anyway.

Hopefully it’s still useful.

Plus I figured if it gets enough traffic I could just remake a better one later.

But WhatsApp with n8n is online ✅

AI Browsers

I’ve had access to Perplexity’s Comet browser for a little while now (which it might be publicly available at this point) but I haven’t used it very much.

I’m a little afraid of getting myself banned from websites for showing suspicious activity by automating things within the browser.

Plus even when I did try to have it scrape a couple websites for me, it quickly got confused and didn’t return great results, and I kind of gave up on Comet for the time being.

I do use the ChatGPT chat app all the time, though because I find it super easy to cmd + tab back to it to ask it a question or use the cmd + 2 shortcut to take a screenshot of the last window I was looking at to ask about code or someone’s comment in Skool.

And this week, OpenAI launched their own AI browser, ChatGPT Atlas (for MacOS only as of now).

Since it had been a couple weeks since I tried Comet, I figured I’d give another AI browser a try and see if I like it any better.

Once logged in, I tested a couple tasks.

I tried to have it build a list of YouTube video ideas based on Skool comments, and it kind of failed spectacularly.

It mostly found my own comments which was not helpful and gave up after 5.

I later tried to have it pull a bunch of addresses for Wave locations and it stopped after 2 when I told it I needed 10x that.

The best I could get it to do was 5 at a time.

So my fears about getting flagged for bot-like activity may be solved because these browsers are heavily instructed not to let you automate repetitive tasks to that volume.

Which is probably a good thing.

Needing to do things more than 20+ times is probably getting into API territory anyway.

I think these browsers are more intended for “shop around for this particular item and add it to a list and find me a place to eat” virtual assistant kind of tasks.

So far I haven’t been impressed, but I may the wrong expectation.

Taking too much of an engineering approach to these tools.

Much like I’ve been doing with another platform…

Accidental SaaS

I’ve recently realized the source of my frustration with n8n.

As I’ve been live building a “course generation” system entirely with n8n on stream, I’ve become increasingly annoyed with how the platform works.

But someone pointed out in my chat that I’m pushing n8n way past automations like it’s intended.

I’ve basically been building a full SaaS (software as a service) on a website that’s just meant to streamline connections between a few apps.

Classic over-engineering as always.

With that realization, it became immediately clear why I was turning pretty sour on n8n.

I can’t expect an automation platform to be a replacement for building real software.

This was a good breakthrough though because it made me forgive n8n, and also made me realize I really do just want to build software.

As I wrap up some of these other projects, I’m looking forward to getting back to building more proper software and leaving n8n for the simple automation flows.

More SaaS to come.

Over 100

I remember counting up to 50 newsletters sometime last year and being excited to hit that milestone.

This morning I realized this newsletter is something like #107.

Well passed 50 and over 100!

It’s amazing how time flies.

It’s become my Friday morning ritual and I don’t think I’ve missed one yet.

There may be one or two Saturday editions in there…

But I look forward to continuing the trend as long as I can.

It’s one of the few things where I practice what I preach.

Doing something just because you want to do it.

Focusing on a number I can control - the amount of newsletters I write each week - instead of worrying about how many subscribers it has or how many opens it gets.

So thank you if you’ve stuck with me through all of these, I really appreciate it.

And if you’re new, welcome!

Always open to feedback and hearing your opinions so let me know what you think!

And like I said, this one snuck up on me so I’d like to do another look back at all the previous ones to see what’s changed and what hasn’t.

But in the meantime, here’s to 100 more newsletters in the next 2 years 🥳

Thanks for reading/opening!

And that’s it for this week. Events vs products, AI browsers and over 100 editions.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

Podcasts are officially back in the mix.

As well as Stripe-first design, more YouTube numbers and the death of Vibe coding.

Topics for this week:

• Stripe-First

• Almost 1%

• Bigger Numbers

• The Death of Vibe Coding

• Blind Recommendation

Stripe-First

Wave took a backseat last week to a number of other projects that suddenly flared up.

But this week I was able to make a little more progress.

I had built the screens and the datatables, so I got to get into some Go code this week to actually connect to Stripe.

Got as far as one method: Create a Wave Event

Not a ton of progress but it revealed another layer, as it always does.

Planning out the function with Gemini Pro 2.5, it brought up a valid point of making sure this endpoint, and all other endpoints, are “Stripe-first”.

If something goes wrong in the middle of an API call, it’s better to make sure that Stripe creates the product first, then creates the corresponding product in the Wave database afterward.

Of course it’s possible to correct the Stripe platform, but it would have to be done strictly through their APIs.

Not that their APIs are bad, but they will undoubtedly have some limitations.

If I need to correct the Wave database on the other hand, I have full admin control which presents many more options to do what I need to do, not limited by API definitions or restrictions.

So the first official Stripe-connected API is also a Stripe-first API, laying the foundation for a number of other APIs to come.

Almost 1%

It only took a year, but with Episode 11 going live on Monday, The Dev Sync is statistically, already past the top 10% of all podcasts just by having more than 3 episodes, and nearing the top 1% by making it to 21 episodes.

The data comes from Jack Butcher appearing on an episode of Chris Williamson’s podcast back in 2021 (and is referenced in the tweet above), but I would imagine the data is still pretty accurate to this day.

It’s hard to stay consistent.

I think it still counts if some of the episodes are over one year apart anyway.

It took a minute to edit this episode since it’s been a year since I last edited a podcast, but I think it turned out pretty well.

Felt right back to normal though, chatting away about AI like we never stopped.

And we had Jumpin’ Jim join us as our first ever guest to mark our one year return.

Only 10 more episodes left to get to the top 1% of podcasts.

Let me know if you check it out:

Bigger Numbers

I’m trending behind on my usual YouTube uploads.

The podcast took up the Thursday recording slot, and editing the podcast on Monday took up my backup recording slot, so I finally got a YouTube video uploaded to my other channel on Tuesday.

Not super exciting but it is a useful one!

n8n has a built in PDF extractor which is really convenient, but it only returns the text of the document.

GPT however can read that text and shape it to whatever data format you’d like, so in this case it was able to convert the content of a PDF to be uploaded to Monday.com directly.

Fairly straightforward, yet powerful.

But also not exactly viral YouTube content, sitting at 59 views 3 days later.

Even Nano Banana videos have been a little lack luster lately.

My first Nano Banana video caught a wave and is sitting at 3500 views, but the following 3 videos only have a few hundred.

Nano Banana videos views, about 4500 total

Generally, it seems, having big faces in your thumbnails tend to perform better so may need to do more of that.

Still getting used to that idea but that may be a requirement for continued YouTube success.

And balancing the clickbait approach versus purely practical title and thumbnails.

We’ll get there eventually.

Despite this, still seeing steady growth.

Nearly 1100 subscribers to the channel which is kinda crazy.

Crazier still, Skool is over 500 members now too, and the only traffic that gets is from YouTube, and maybe a few people that browse within Skool.

522 members in Skool

Slow and steady with one video upload per week right now.

Always room for growth but I’m pleased with the progress so far.

Plenty more uploads to come and hopefully plenty more growth as well!

The Death of Vibe Coding

We talked about vibe coding in The Dev Sync episode above, where I gave my opinion there.

Coming from the perspective of someone that generally knows how to code and still has been stuck in the vibe coding loop myself, I’ve had my doubts about the trend.

But it’s a fun term, and makes for good marketable posts and clicks, so it’s been a good trend.

It could be that it’s dying down now, or at least shifting now that people are starting to see the cracks.

AI is getting better and better at coding, but there’s still plenty of room for mistakes.

Until it can successfully complete a task end to end with a general directive, I’m not sure developers are out of the job yet.

And if anything, it further enforces the need for developers because you need someone that can actually spot the problem when AI is innocently blind to the issue.

This post on LinkedIn seemed to have sparked the idea that vibe coding is dying, although the idea is derived from potentially inaccurate data about the traffic to an AI coding website called Lovable.

I was listening to Theo’s take on the situation because I was curious to hear more about vibe coding dying:

He brings up some good points after admitting the original data about traffic might not be the most reliable indicator of how Lovable is doing, but it’s certainly food for thought.

I’m happy to keep vibe coding or whatever we’re going to start calling AI generated code production instead.

Blind Recommendation

It’s been a while since I’ve recommended a podcast without having listened/watched it yet, but today breaks that streak.

I was reviewing my YouTube history to remind myself of the week (one of these days I’ll learn to take better notes for this newsletter), when I saw this thumbnail:

These are still 2 of my most listened to people right now, so I am certain this will be an entertaining episode, even at nearly 3 hours.

I’m sure I’ll have it done by Monday.

Chris celebrated his 1000th episode with Matthew McConaughey 2 weeks ago (which surprisingly I haven’t finished yet), and it doesn’t seem like Modern Wisdom is slowing down anytime soon.

Dr. K has been on a tour lately as well, showing up on a few other big podcasts including Flagrant and Jay Shetty’s On Purpose (I did listen to these).

But I’m looking forward to watching Chris and Dr. K chat for a while again.

Thoroughly enjoyed their last conversation from one year ago so I have no doubts this will be good as well.

Feel free to let me know if this doesn’t turn out to be a very good blind recommendation.

Always open to more data.

And that’s it for this week. Wave is operating Strip-first, and I snuck a few podcast recommendations in there for old time sake.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

An old coding buddy is back.

And an old project revitalized.

And an n8n replacement?

Topics for this week:

• Wave Events

• YouTube Cadence

• A Quick Year Break

• Old New Code Buddy

• No More n8n?

Wave Events

Taking a slight detour from the Wave marketplace for now to focus on monetizing Wave more directly.

Getting venues to sign up for a Stripe Connect account shouldn’t be too much of a challenge but it’s a harder ask than something like “Wave will bring you 100 people for free on Saturday”.

Wave gets to charge tickets for the event, the venue gets all the food and drink business and it’s a win-win.

Since a Wave event in this situation would be owned by Wave, but hosted at a venue, it seemed logical to make the event button a separate button.

New “Wave Event” button

And a nice looking button at that.

I had Claude Sonnet 4.5 cook that one up and it did a spectacular job.

I needed a number of other screens to help create and manage the information about the event and Sonnet 4.5 killed it with those too.

Wave Event Management screen

A nice Events Details dashboard

After almost a year, Claude is still helping Wave look clean.

With those mocks and most of the database schema built out, Wave Events should be ready to integrate with Stripe APIs this week.

Plenty more work to be done, but it’s looking good so far.

YouTube Cadence

I’ve had a recent flair up of activity with clients and projects so YouTube is starting to fall behind at the moment.

I did start the Live Stream streak back up this week, 4 of out the 5 days, with plans to get a Friday stream in after this newsletter is done.

But I broke the “1 video per week” cadence last week by missing an upload on Thursday.

I got a video done on Monday instead and it’s slowly building momentum, almost 100 views as of this morning.

I’ve been leaning into what brings people to my channel at this point, so that meant sticking with Nano Banana for now.

n8n nano banana

I’m still genuinely impressed with Nano Banana every time I use it.

I also know I need to get more social media content going so I thought a system that could generate some options would be useful.

As often as you want, this system will take a prompt, choose a template image and add a word from a word bank to the appropriate image chosen by GPT.

Pretty pleased with how this one turned out, not too complicated but still helpful.

Check it out and let me know what you think!

A Quick Year Break

After nearly a year, The Dev Sync will return with Episode 11!

And it will have a special guest!

We recorded last night so I haven’t had a chance to edit it yet.

The plan is to get it uploaded on Monday.

The last episode was just about 1 year ago, October 24, 2024, which is really wild to think that much time has slipped away.

It was great to get the gang back together and ramble about AI again.

A lot has changed since then!

I built an entire app in that time and there have been 4 different versions of Claude Sonnet between then and now.

And plenty more to discuss.

Here’s the last episode in case you want to get caught up for the new one:

Old New Coding Buddy

I finally gave Claude Sonnet 4.5 a try this week, and at first I was really impressed.

Sonnet 3.5 was my main coding companion when I started Wave and it helped built out most of the original styling of all the screens.

But after a while, I got frustrated with Sonnet’s struggle to stay on task.

I ended up finding a lot of luck with Gemini Pro 2.5 instead.

And even more recently I’ve had a lot of success with GPT-5 within Cursor.

But hearing the hype around Sonnet 4.5 I had to give it a shot.

And as I highlighted above, the results were really great.

New screens looked amazing on the first try, which is quite a feat!

Very satisfying to see the result be nearly perfect with minimal prompting.

And it did so with maybe only 1 line changed that didn’t need to be changed.

But in that one particular trial session, a little over 2 hours of vibe coding which is on the shorter side, Cursor warned me I was on track to burn up my $20 in credits by October 6.

This was on October 5.

Sonnet was using $20 of processing in a few hours, where GPT would last me for a few weeks.

gpt-5 costs in Cursor

sonnet 4.5 costs in Cursor

Clearly there is a big price discrepancy.

• gpt-5’s cost for 1.2 million tokens: $.17 

• sonnet 4.5’s cost for 1.1 million tokens: $.95 

This pretty accurately lines up with the printed prices for each service:

$3 per Million Tokens in, $15 per Million Tokens out

$1.25 per Million Tokens in, $10 per Million Tokens out

I really liked the visuals of Sonnet 4.5 but it might be a little too pricey to use it for everything.

The plan for now will be gpt-5 is the workhorse in Cursor, only switching to Sonnet 4.5 when I need a new screen mocked up or changed.

And this will still be backed up by Gemini 2.5 Pro in the browser.

Until the rumored Gemini 3.0 drops…

But until then, it’s good to see Claude have such a big improvement and be welcomed back into the AI rotation.

No More n8n?

That’s the easy marketing hook everyone is and will be using on social media for a while now.

OpenAI dropped a bunch of updates this week complete with a workflow builder, similar to n8n.

Twitter tweet

I haven’t had much time to try it myself but I did poke around the demo site and it looks promising.

The widget gallery looks particularly pleasing:

Feels very similar to Google’s Opal.

And now it makes sense why OpenAI announced their Shopify and Etsy chat integrations last week because AgentKit allows for building entire app flows, including purchasing, in a single interface.

OpenAI is attempting to take what n8n does to the next level and build the workflow and app for you using AI.

n8n is supposed to have their own AI builder soon, it’s been previewed, but I’m not sure it’s fully rolled out yet.

And even though this is a promising release for OpenAI, giving control of a full AI flow rather than hoping you prompt things correctly or having to custom code API calls together, I do think n8n still has one big advantage.

Open source.

OpenAI wants to take the Apple approach and create a completely closed system.

They want you to build on their platform and use their tools.

n8n will let you download the code and run it yourself.

Both have their advantages and disadvantages but it will be interesting to see where this leads.

I’m sure I’ll test out AgentKit soon, maybe even have a video about it in the near future.

But another fun release in AI that I don’t think will kill n8n, for now.

Getting closer and closer to AI building entire apps on it’s own though, so maybe soon?

And that’s it for this week. Old and new buddies and projects.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!

A slower week for a change.

Being out of town for a few days, I got a chance to enjoy reading about AI news and not worry about implementing it all.

Not yet anyway.

Topics for this week:

• AI Marketplace

• Doubling Down

• Leveraged Growth

• New Code Buddy?

• New Social Media

AI Marketplace

Answer Engine Optimization or “AEO” is a new term I heard on a podcast a few weeks ago.

Everything used to be all about SEO (Search Engine Optimization) which really meant trying to show up at the top of Google searches.

But a number of stats are flying around claiming Google searches are down a huge percentage because people just ask ChatGPT (or whatever AI) instead.

If that’s the case, then businesses are trying to make sure their websites get picked up as an answer from an AI instead of the number 1 Google result.

It probably doesn’t hurt to still be in the top spot for Googling, but if ChatGPT or other “Answer Engines” like Perplexity, Claude, or even Google’s Gemini will provide links to sources, might as well try to be the link the AI provides.

OpenAI is taking this a step further by releasing the Agentic Commerce Protocol:

Twitter tweet

No need to click a link to go to a Shopify site, just buy right there within the conversation.

This felt particularly relevant as I’ve been working on a marketplace of sorts with Wave.

Wave isn’t using AI conversations yet, but this Agentic Commerce Protocol seems like a pretty big opportunity in general.

If a new online shop is able to build a website with favorable “AEO”, it could have AI models automatically suggesting and selling their items without a single Google search.

It’s a wild time for the internet.

Get the feeling there will be plenty more shifts in how things work as AI manages the internet more and more for us.

AI’s buying and selling with other AI’s, no humans required.

We’ll see how quickly the Agentic Commerce Protocol takes off and how much more purchasing people will be doing directly within their AI conversations.

Or if the AI’s will be doing the purchasing as well…

Doubling Down

I was out of town for the majority of the last 7 days so I was not as active on social media, though I’m still not super active in general.

But despite the travel and distraction, both of my core focuses continued to grow.

28 subscribers in the last 7 days

33 members in the last 7 days

The YouTube channel grew by 28 subscribers with zero uploads, and the Skool community grew even more, gaining 33 members in the last 7 days.

“Doubling down” is the phrase that keeps running through my head when I see these numbers, but this realistically isn’t even a doubling-down situation.

I’ve been posting 1 video a week for a good stretch now.

I could easily septuple down just with YouTube shorts alone, posting 7 days a week, and probably see at least a 2-3x return in growth.

The path is becoming clearer.

More time on YouTube, less time on low-paying projects.

Leveraged Growth

Another theme that’s been popping up along side “doubling down”, is “leverage”.

I quit a full-time job a couple years ago with that idea in mind, thinking I understood what leverage was.

I was “trading time for money” at a job because there’s no scalability.

Somehow I have fallen back into the same trap, working for clients that pay flat rates with no leverage to be found.

Part of the problem was following trends online, whatever the hot business model was at the time.

Dropshipping, crypto, web design agency, now automation agency.

None of them were really my idea, just someone else’s.

And none of these were leveraged activities.

They were all ways to make money upfront to funnel that money into leveraged projects.

YouTube so far has been the only thing that I’ve stuck with that’s actually a leveraged activity - and this newsletter (thanks for reading/skimming/opening 😉)

Of course not everything can be leveraged, but I could be prioritizing more leverage than prioritizing clients out of obligation at this point.

So that’s the big pivot.

Build more leverage for myself.

The doubling or septupling down on YouTube would be a great start.

But another one that’s popped up lately is SEO or the concept of “AEO” from earlier.

YouTube is a great source of traffic, but I’m not sure I’ll be able to build a YouTube channel for every project.

If I do end up vibe coding a few projects of my own, SEO or AEO may still be a viable way to get traffic to those new tools I end up building.

The great indie-hacker-turned-vibe-coder @levelsio has said despite his large X following, only about 5% of his app traffic comes from social media, the majority is a result of “pSEO” or “programmatic SEO”.

He automates the creation of pages on his websites based on what people search for which leads to him ranking higher in searches in general.

Very smart approach, and I don’t know how I’ve never seen the term “pSEO” before until this week.

Armed with YouTube and pSEO, I think there’s a good possibility to leverage some traffic into a few of those website ideas I’ve been kicking around…

New Code Buddy?

I’ve gone back and forth on Claude unfortunately.

For a long time, Sonnet 3.5 was my primary coder for Wave.

But after a while, I got tired of it constantly changing things that didn’t need to be changed, and when new model upgrades released, I ended up switching off of Sonnet entirely.

I think I tried Sonnet 4.1 when it first came out but still felt like I was getting better results with Gemini Pro 2.5 and GPT-5 at that point so I really haven’t used any Anthropic models in a few months now.

This week Anthropic released Claude Sonnet 4.5 claiming it to be the best coding model in the world.

Twitter tweet

Being on vacation, I haven’t done much coding so I haven’t had a chance to test it myself but overall it seems positively received.

LMArenan rankings Oct 3

Sonnet 4.5 has climbed it’s way to tied for 1 in Text and tied for 4 in WebDev on LMArena so it’s clearly testing well.

It’s priced closer to GPT and Gemini at $3 for input and $15 for output which is much more sustainable than Claude Opus’s pricing.

I’ve seen a few people say Sonnet 4.5 is better than Claude Opus which was Anthropic’s most powerful model until now.

So I’ll be anxious to try it out.

I’ve been switching back and forth between Gemini Pro and GPT-5 recently so we’ll see if Sonnet 4.5 is able to bump one of those out.

More vibe coding planned for next week which will be the perfect opportunity to try it out.

New Social Media

The original release of OpenAI’s Sora was a big deal when it came out.

It was one of the first AI video models to actually look real at times.

Sora quickly fell behind with a number of huge updates since then.

Currently Google Veo 3 probably still holds the lead for most realistic video generation, but this week OpenAI announced Sora 2:

Twitter tweet

Kind of hard to believe some of those clips of Sam Altman aren’t actually Sam Altman.

Those look real.

The voice is just a little off, a slight hint of AI robotic sound, but I’m sure most people wouldn’t even notice.

Sora 2 has stepped their game way up since the first model, which was impressive to begin with.

This new version looks on par with Google’s Veo 3.

But maybe the bigger deal is that OpenAI is taking a TikTok approach.

They’re slowly releasing Sora 2 through the Sora app for iPhone only, and invite only.

If you get invited, you get 4 codes to invite others.

And the app is focused on creating a feed of AI videos only for you and your small group of friends, not the entire internet.

Overall I think it’s a smart approach - get dedicated followers first, then use network effects to grow rather than release to the masses.

Plus this saves them from “melting gpus” by not letting a large percentage of the internet generate videos constantly for the next few weeks.

I’m guessing they’re trying to also capture the TikTok/Instagram Reels crowd to better establish their own social media platform so we’ll see if it sticks.

Most of our phone screen time may be on the “Sora” app in the next few years unless they have another product-line name planned for a social media platform.

If this long-term vision succeeds in reaching Facebook-level adoption, it would make their app the first truly AI-based social media app.

Haven’t gotten an invite yet so I haven’t been able to try Sora 2 out myself.

Hoping I’ll have an update about it next week.

And that’s it for this week. AI marketplaces, AI social media, and leveraged priorities.

If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.

I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316

Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email me@jaypeters.dev.

Thanks for reading!