it's been a while

I made the move to the newly formed AI tools team at work about 6 months ago, which was about the last time you heard from me with any consistency.

It’s been BUSY. Turns out everyone wants to use the latest AI tools in the most optimal way possible, yesterday. Doing that across an enterprise is challenging, to say the least.

We have thankfully hired more people for the AI Tools team, so we’re settling into a more established-team-pace and workload. This means I have more time for personal projects again, which is kind of the whole reason I started this thing.

But first, let me take some victory laps on a couple things that have either happened, or have become more true that I predicted in this very newsletter. First, some high-level concepts that I think are most important, and then a “highlight reel” of sorts that Fable has clip-farmed for me [1].

We all have access to this ultimate solver tool that can basically do anything, and now it all hinges on “what” we do with it. I call this “Perspective” because it has to come from human experience. It is a vector or direction of where to point the LLM; what problem it should solve.

What is the highest-leverage trade you can make on token spend? What is the most valuable thing that you can create with your given token budget? You can't just sit on the sidelines and theory-craft this. You have to get in there and burn (tokens) to learn.

I think my best, easiest to understand, most reliable prediction is the Forest of Low-Hanging Fruit. Software nerds all have a similar set of shared experiences. We're all close to each other in vector space, or something.

At the beginning of the year, we started with writing our own agent orchestrators, todo lists, and fitness trackers. Claude and Codex have now Bitter Lesson’d one-shotted those for us, and now people have moved on to software factories. You know where this is going. I tend to stay very vanilla with my agent setup and just use whatever tools the labs give me because those are the best supported and are incentivized for the models to be trained on.

Software Developers who have been smart enough to pair with a domain expert to try to find some gap in the market are still seeing competition by others who may just be domain experts pairing with Claude. That leads me to my next most important concept that I've talked about before, Distribution.

Now that basically anybody can spin up a copycat clone app with very minimal effort, it very much depends on who knows the “What” you can do with AI and how likely they are to trust you with doing it for them.

“Trust” is a weird word that is doing a lot of heavy lifting here. It's influenced by a few sources, most recently this book I read recently on a recommendation from a friend, “Return to Real” by Ryan Levesque, and also more well-known marketers of Alex Hormozi and Gary Vaynerchuk.

People still buy from people they trust, but people are also trusting AIs more to make decisions on their behalf. I'm certainly doing this with all of the code that I write through AI. As AI becomes more accessible for non-tech-nerds through personal agents such as Muse, Dots, Duet, etc, purchasing decisions are next up to be delegated.

This is why I'm spinning this content production pipeline back up. I believe it is more important than ever to put as much of your human self out onto the internet as possible. You want to capture the human trust, but also the AI trust. The same thing that impresses a human will be similar to what impresses an AI.

Consistent Quality and Volume (“Qualume” from Hormozi) over time equals results (C/T=R from late MMA Coach Robert Follis).

This content pipeline was possible before, as I was automating more of my social media process with skills, but it's definitely even more doable now that Claude Opus 5.5 has crazy native video editing capabilities by itself and Codex has wild computer use functionality.

My goal with this project is to write my thoughts in newsletter format or record myself in a long form video, and then use the Agents to, as high fidelity as possible, transform, clip things up, and post them to all of my social media all at once, consistently, every day, indefinitely.

Boris Cherny and Claude made formal verification blow up overnight after he posted about formalizing the Claude Agent SDK. I've been reading the tea leaves on this for at least 6 months. Others have predicted this since almost a year ago, like Martin Kleppmann, the author of Designing Data-Intensive Applications, in his post “Prediction: AI will make formal verification go mainstream”.

I have only read one chapter of one introductory book on this and a number of surface-level blog posts on why it's good and who uses it already (spoiler alert, everyone doing fundamental infrastructure work that cannot be wrong). I need to do more human neuron work in this area to be actually informed on when and how to best apply Formal Methods.

Formal Methods are also getting more attention now because of the myriad of security vulnerabilities being discovered autonomously by abliterated attack agents. It would be great if we can prove that there are no security vulnerabilities in our code to begin with.

Devs are keen on FM now because in the “intent → spec → implement → verify” pipeline, they are realizing that the “Verify” step is the most important step and hardest to get right. You require a very clear definition of “correct”, which is ever-evolving and may just need to be updated iteratively as we go along. There are entire programming languages now being developed, like Bend, that have formal methods baked into them.

This summer, OpenAI and Anthropic showed that both of their unreleased, highest-end models can formalize and solve open problems in math. This unnerved me, and I think I can describe why now.

Access to the latest and greatest models is always going to be used first to further develop better and better models. That's this whole recursive self-improvement (RSI), superintelligence loop that Nick Bostrom and others have warned about for at least 10 years. That's the most logical place to allocate their limited, though vastly more than you or I could ever individually control, compute resources.

Unreleased closed models now just brute force open Math problems AS AN EVAL (as a way to “test” the new model) with up to 10,000 agents. What happens now? Do we keep walking up the dependency chain with Math → Physics → Chemistry → Biology → Psychology → Sociology → Economics → ???. What’s the end state wrapper?

This is the furthest data point for the Bitter Lesson.

For those unfamiliar, the Bitter Lesson (how I understand it in my head) is a lesson about how more compute yields more intelligence and that any clever thing you do around the model to optimize or "fill in the gaps" for a lack of model capability will get blown away by more compute in the next-generation of models. Human creativity is defeated by raw compute, thus it being bitter. The model tends to eat everything around it by folding it into its own capabilities.

This is where fretting about “moats” comes in. What is worth building (aka spending your Human time and attention on) when the next generation of models (that come out more and more frequently) can one-shot your current implementation?

The best answer I have for commercial application is what I read in a silly little md file in Gas City: the primitive test. It's specific to Gas City, but the thrust applies everywhere. TL;DR: You want to build and monetize dumb data pipes that the model and harness can just understand and use seamlessly. Your service becomes a thing the model doesn't have to reinvent the wheel for and can rely on to “just work”.

I always refer to social media posting sites like Buffer and OpusClip as services to emulate. I don't want to deal with (aka continuously spend tokens on to maintain) all of the intricacies of reliably posting to all social media platforms over however many years of experience they have maintaining these dumb pipes as a business. I’ll pay the $8/mo per account for them to handle it for me.

The primary thing that unnerved me about Math being solved is the realization that there are going to be very few people who are going to work on the most important problems with effectively unlimited access to the best technology and that the rest of us are going to have very limited access to it and little ability to “catch up” or “compete”.

Now that I state it like that, this has always been the case with Important Things like physics, chemistry, biology, etc (not everyone has access to a science lab with expensive equipment), but the technology is now an alien intelligence that applies across all domains. What stings is that we receive some dumbed-down, limited access to it vs. nothing at all, like a science lab.

This unnerving realization felt like being utterly helpless as a white belt rolling with a black belt in Jiu-Jitsu. You think because you’re a fit person, you lift a few hundred pounds of weights (younger me), you think you have some kind of chance of doing something against a rotund, seemingly out of shape black belt (present me) in a 6 minute round.

There is just such a large skill discrepancy that there is no way, ~0 (zero) percent chance in a grappling ruleset you have at tapping a black belt as someone coming off the street. While the black belt can choose to tap you as many times as they’d like during that same 6 minute round. The only way for the white belt to do better is to put in effortful practice day in and day out, for years, all the while, the black belt can keep getting better in that same time [2].

Math getting solved made me feel that what I was working on with my dumb primitive test data pipes was insignificant like a white belt compared to what people at the labs are working on. That all of the significant advancements in world-changing technology are now concentrated in a handful of companies.

Now that I state it like that, this has always been the case, given all of the hyperscalers owning all of the compute. But for some reason, this hit me more because it demonstrated that what we get as consumers of the labs’ technology is sort of the old leftover scraps after they've extracted their own value out of them.

What bothered me was that there is a significant delta between what labs get to use internally and what they sell to us to fund the future of their nice internal models.

How are you supposed to compete with worse tools and significantly less access to compute? How does a white belt compete with a black belt? You don’t.

You humble yourself and you learn from them. You work with them. You figure out what they aren't going to work on with the primitive test, and you focus your attention there.

If you truly do think you want to compete or have a drive to be at one of the labs, then you fast-follow. You figure out how to advance from “white belt” through the “belt ranks” to achieve a status that gives you access to work on the frontier problems with frontier compute.

I think this is more or less the ego blow of what the Mathematicians were feeling, but here's where I think human attention and Perspective are still needed.

In a previous newsletter, I quoted Axl Rose, and here I quote him again.

The world is big, with a lot of people in it. The number of problems in the world is probably only going to go up. Problems tend to get divided into “fields”.

AI labs aren't there to one-shot your field of study. Their incentives are only to make better models that they can sell to you so that you can do work in your domain of expertise. If something you made gets one-shotted by a newer model, now it's on you to think bigger, more ambitiously, further down the line.

This is uncomfortable. We don't get to rest on our laurels. We don't get to do something once and then make our career off of it and retire riding off into the sunset. If we (I) have 30 more years to work, and our primary value now as humans is defining “What” the agents work on, we need to get more creative, fast. This may feel like wanting to run faster and spin up dozens of agents and work like hell around the clock, but I think it's more athlete-shaped than that.

Coding agents allow us to push ourselves to mental extremes harder and faster than ever. With anything extreme with severe consequences when things go wrong, there has to be a discipline around it. I have never worked more or harder in my professional life than the past 18 months, and I’ve felt the weight of burning out and getting sick multiple times.

Another thing I just so happen to have experience with that pushes you mental and physical to extremes is, drumroll please, you guessed it, Jiu-Jitsu. When you first start, it's almost like an addiction. You never want to miss a class for anything because in every class, you feel like you're improving so much and it’s so fun [3].

You end up training so much, so often, that you never take a rest day, and you end up getting injured or sick because your body is run down. The only way to progress in Jiu-Jitsu is to stick around consistently, long enough, injury-free. You need to develop a schedule for yourself where you can consistently train and then recover. Sure, you can build up to more training volume over time, but there is a certain limit that your body can naturally take.

AI Agents feel like this type of mega-reinforcing-to-the-point-of-personal-detriment type of technology similar to Jiu-Jitsu class. Training is not a strictly negative thing, but the dose makes the poison. Steve Yegge called this out with his AI Vampire article in February, but I think more and more people are probably viscerally feeling this now.

I'm struggling with this myself, but maybe something simple is to take a full day on the weekend off from all short-form attention splintering technology and rest. I've started to really watch and pay attention to long-form TV shows and movies again. It's still challenging to try to not pull out my phone and have a second screen while watching the big screen.

Our attention is getting pulled in so many directions constantly throughout the day. We need to somehow reintegrate our perception of time and experience content (or real experience) that is real life-paced, and not edited or engineered for clickbait audience app retention purposes.

I'm going to sound like a broken record, but Yegge recently talked about in a podcast that we need to now exercise our minds in the same way that we have to go intentionally exercise at the gym. We can go through life not lifting a finger. Now we have to intentionally go lift some weights to stimulate those positive physiological processes in our body, and now we need to do the same thing for our human neurons.

In next week's article, I'm going to deep dive into what it means to be a software engineer in 2027 and how we are going to have to operate to thrive in this new era.

Ok that’s enough of my human neurons talking. Let’s see what Fable says I got right this year:

The full scorecard, with sources for every claim and the ones I got wrong, lives at Enterprise Vibe Code Predictions.

[1] If you ever watch a TV show and they, in the middle of the season, do a montage clip recap of previous shows (otherwise known as a “clip show”), that means they ran short on time and just didn't want to produce new content. This is my version of it, and it is indeed lower effort and is just trying to get me back into the habit of doing this.

[2] Of course, we're humans and are limited by our biology and individual lifespans and health spans. The black belt slows down and can eventually be caught, so who the white belt is really competing against is the cohort of people who started around the same time as them. Skills are learned linearly in time (x-axis), but I believe the absolute level is on an exponential curve (y-axis).

[3] It’s an example of the Law of Small Numbers. Just by sheer class number attendance early on, adding one class feels like a gigantic step forward in your understanding of Jiu-Jitsu, plus it’s fun as hell. At two classes, you feel like you doubled your knowledge of Jiu-Jitsu, at three, tripled. Jiu-Jitsu is so large that you probably don't stop experiencing a new technique or position every class for your entire first year.