I’m trying to become a better AI builder and have been lucky enough to interview 5 of the best:
1. @kieranklaassen built the Compound Engineering system that many AI builders swear by. In our episode, he walked through exactly how it works, from planning to skills to commands like /ce-plan, /ce-work, and /lfg (let’s effing go).
https://www.youtube.com/watch?v=g6z_4TMDiaE
I’ll link the free Compound Engineering system below.
2. @ryancarson has started multiple companies and is now running a startup solo with OpenClaw, Codex, and Devin. It was fascinating to see the system he built to help him run his company.
https://www.youtube.com/watch?v=IDqdVZwAwjw
3. @Shpigford has been building products for 25 years. Like many AI builders, he created his own skill stack to streamline the way he works, including /build, /adversarial-code-review, /but-for-real, and /learnings.
https://www.youtube.com/watch?v=GdxLaeyu33c
4. @kunchenguid made it to L8 in big tech as a principal engineer, which is very hard. Then he quit to ship 40 PRs a day with agents and build free open source tools for visual planning, parallel agents, and AI code review.
https://www.youtube.com/watch?v=88B6DimMD2g
I’ll link his GitHub below too.
5. @mvanhorn has built some of the most popular open source AI projects on GitHub despite having no formal engineering background, including /last30days and his Printing Press system. I’m dropping his episode next week on my YouTube.
If you’re trying to become a better AI builder too, start with these interviews that show you how real AI builders work.
Subscribe for more conversations like this:
https://www.youtube.com/@PeterYangYT?subscribe

Good take
My guess is
- demand for intelligence is near infinite
- but 80% of workloads will be running on 99% cheaper models within 12-18 months
- 20% of workloads will still run on latest gen models where IQ maxing is important (scientific breakthroughs, higher level ochestrator agents?)
- rough analogy might be what % of macbooks or gaming PCs sold have the maxed out specs for CPU/GPU, prices are falling much faster than Moore's law here though
- this leads me to think the limiting factor will be energy and compute, not better models
At Coinbase we're working hard on routing prompts to cheaper models where appropriate, and in some cases have been able to keep costs roughly flat, while token usage continues to grow exponentially.
Tommy: The most basic way AI could blow up imo. I'm not saying it does but this is the most obvious way I can see it happening
- Per seat subscriptions are massively subsidized. The flat fee was priced way below what heavy usage actually costs
- For real business use you have to move

i agree that the frontier american models are clearly better, but it doesn't help that the evals being used are such bs that the compelling way to actually assess as a user is to just try em and decide based on vibes.
eg many of these evals put opus 4.7 and 4.8 *way* higher than 4.6 which is nonsense to anyone that has used them.
pair that with the reality that most people just aren't yet using these for anything all that sophisticated (even among the power-ish users) and it makes sense that the chinese OSS models seem compelling.
Dean W. Ball: You’d be shocked by how many people in think tanks/academia/government/“strategic classes,” including in the U.S., are convinced that Chinese models are now “good enough” and leading the world in adoption. Meanwhile, the reality I see is a fairly wide, and still widening, gap.

http://x.com/i/article/2063850827694096385

A pause continues to be utter and complete nonsense and it always will be.
1. Let's make planes safer by not making planes!
2. What exactly happens in a "pause"? Do labs get to keep working and we all just get to sit on our hands waiting for this eureka moment?
I guess someone gets to keep their checks while we just bring the economy to a crashing halt based on a few people's vague ideas about some imaginary future problems that they came up with while huffing glue and reading Dune!
3. Who will fund the labs when they are not putting out new products?
I guess VCs will continue to just give them 100s of billions out of the goodness of their hearts!
4. What actually justifies a pause?
Apparently, the wonderful world of imagination! Theoretical future problems that haven't happened yet.
Like massive job losses! Um, jobs are increasing including in the areas most affected by AI, like coding, so I guess not that.
So what? I know the jobs apocalypse is coming because I can imagine it and imagination is reality, right?
Maybe advanced AI weapons?
Ah, so the government will pause weapons research too?
Well no, they will keep doing that anyway because they always do that. It's what governments do.
Okay so we're going to ban Chat Bots while the government keeps making weapons? That should solve everything we're worried about with AI!
Well then what about recursively improving models that grow to superintelligence overnight?
Yeah that's not really a thing. That's the plot of an Avengers movie.
Models are bound by the same real world constraints we are like compute (brains/chips) and time (will this drug have side effects in twenty years can only be known in twenty years) and fuzzy multiplicity (not right or wrong but right-ish and wrong-ish means you can't make a reward signal for "is this the right decision for my business") and the subject/object paradox (the thing improving is judging its own improvement. Yeah chew on that one for a bit.)
But I imagined AI overcoming every real world constraint instantly so it's true!
5. How would we know we did everything we needed to do in a pause? How do we know when it's over?
We dont. We just want a pause now because we want it! Don't you see my pause ⏸️ emoji? It's nice right!
6. Who gets to decide we are ready to unpause?
The government or the people!
Great, because vague ass platitudes like this always go well for concrete policy design.
Looks, none of this is real.
It's theater.
It's not real policy. It has no basis in reality. It's a mass hallucination.
It's pushed by people who believe in magic and magical solutions.
And anything that comes out of it will do infinitely more damage than the imaginary thing they were trying to protect us from in the first place.
roon: now on the eve of RSI it seems everyone is more mutual conditional pause agreement pilled than they used to be and that seems like a good development

Striking paper from Wharton. The big conclusion: AI must increase productivity 2.7x -- and quickly -- or tech companies risk bankruptcy with all that entails for the economy. For context: this is how a quickie 2.7x productivity boom would compare to historical precedent. Paper linked in my daily AI digest. Useful context for OpenAI reportedly talking to the US government about a bailout (ahem, I mean ownership stake).

SF had a brief window during Covid where they could’ve tackled its affordability problem and did nothing.
What a wasted opportunity.
Mike Simonsen 🐉: The data for rent appreciation in San Francisco is just nuts.

My 5 biggest takeaways from @kunchenguid, ex-Meta L8 engineer, on how he set up his agentic engineering system:
1. Spend most of your time planning and validating, not coding.
Kun sees himself as the manager of an always-on engineering team. His job is to create plans, validate work, and improve the overall system. The coding phase is handled almost entirely by agents.
2. The quality of your plans determines how long agents can work on their own.
Memes about loops aside, an one-line prompt might keep an agent working for a few minutes but a detailed plan can keep it going for hours.
If you want agents to run longer, invest more upfront in the spec, goals, and validation criteria.
3. Use visual plans, not walls of markdown.
Kun built Lavish, an open-source tool to make visual HTML plans, that anyone can use for free: https://github.com/kunchenguid/lavish-axi
Instead of reading a giant markdown plan, Lavish turns it into a visual HTML artifact where you can leave inline feedback.
This makes it much easier to tell the agent exactly what to change before coding starts.
4. Run agents in parallel, but isolate the work.
Kun uses Treehouse to manage reusable worktrees so agents don’t step on each other’s changes: https://github.com/kunchenguid/treehouse
If the work is exploratory or likely to fill the context window, he delegates it to a subagent.
This way, the main agent stays focused while the subagents dig, test ideas, and report back.
5. Let agents review the code before you do.
Kun no longer manually reviews every AI-written PR.
Instead, he uses No Mistakes to run a fresh agent review, rebase the change, run tests, update docs, create the PR, and assign a risk level: https://github.com/kunchenguid/no-mistakes
In Kun’s testing across 267 agent changes in 15 repos, No Mistakes caught and fixed 68% of mistakes that would have been missed.
Kun walked through Lavish, Treehouse, No Mistakes, and his full agentic engineering workflow in our episode.
📌 Watch it here:
https://youtu.be/88B6DimMD2g
Peter Yang: "If you're still manually reviewing every line of code, you're the bottleneck."
Here's my new episode with @kunchenguid, an ex-Meta L8 engineer who now ships up to 40 PRs a day with AI agents. Instead of manually reviewing code, he built an agentic engineering system that

Lots of excitement about this but I’d encourage everyone to tone it down. Here is the most important sentence in this paper:
“it remains unclear whether the survival benefits of GLP-1 agonists extend
beyond their effects on weight. In fact, the questions of whether GLP-1 agonists directly impact tumor growth, whether they reshape the tumor microenvironment, and whether they reduce systemic inflammation independent of the impact on weight are major knowledge gaps.”
In other words, this is a cool association, but we may just be seeing the effect of reduced weight. This retrospective only looked at people with BMI above 25 before taking a GLP, and it makes sense that reducing weight would reduce cancer risk because obesity increases the risk of many diseases.
this is all to say - obesity sucks and it’s amazing we have a drug that might end it. But if you’re not obese you should not believe there is any evidence this drug will prevent cancer.
ABC News: A new study of more than 111,000 women ages 45 to 80 found those on GLP-1 medications had a reduced risk of developing breast cancer by about 30%. https://abcnews.link/ZTUh9eG

This one tower will have more units in it than all of SF has permitted in 2026 so far
Only In Boston: A nearly 400-foot luxury tower called is nearly complete above Huntington Avenue in Boston.
The 34-story building will include apartments, retail space, underground parking, and a connection to the historic Huntington Theatre.

"What if the model companies do this?" is the new "What if Google does this?" I.e. the meaningless question investors ask that shows either that they're stupid or that they dislike you and are looking for ways to find fault.

The current line is “AI can replace execution, but not taste.”
But taste is just compressed judgment from lots of exposure: see enough examples, classify enough good vs bad, and you develop instinct.
That is not a moat against AI.
That is a training set.

The Agent Open 🎾🏓
Everyone loves pickleball. We’re hosting a massive pickleball tournament during the AI Engineer World Fair.
I’m so excited to see this event come together. This is an AWESOME collaboration between 7 companies:
@braintrust
@browserbase
@cursor_ai
@modal
@p0
@turbopuffer
us
Come find out which AI Engineer, founder, tech influencer, VC, or anyone in between is the most cracked. Come participate in our tournaments, or come hangout for drinks, conversations, and casual games!
Sign up here: https://luma.com/the-agent-open
LlamaIndex 🦙: The Agent Open: AI's Pickleball Tournament 🏓
Come put your code and backhand to the test and embrace the full Open experience.
Custom built out courts. Stadium seating. Exhibition matches by AI leaders. Fresh agent merch. Every infra startup you love, all in one place.

As many of you were interested in the technical details of the model, here is a followup thread to go more into technical details about VLA-JEPA.
1. Architecture
2. Training
3. Recipe for the demo
4. TLDR
🧵below
LeRobot: VLA-JEPA just dropped in LeRobot 🤖
What makes this model special is that it does not just learn what action to take from a given observation, it also leverages a JEPA world model to learn action-relevant dynamics.
During training, the VLA leverages V-JEPA2 by conditioning its

Earlier today, New York University announced the creation of its new Earth Systems Institute—a multidisciplinary hub that will deploy AI and computational tools to better predict environmental changes and to advance means to better prepare and respond to these global phenomena.

When we first demoed Claude Code internally, it got two reactions on Slack.
A year after GA, @_catwu and I sat down to talk about what's changed: why I use auto mode instead of plan mode, how routines fix bugs before I see them, why I do most of my coding from my phone now, and where the product is going
ClaudeDevs: Claude Code's first demo got two Slack reactions.
One year after GA, @bcherny and @_catwu look back: verification best practices, why we built auto mode, routines and loops, and what's next.
https://www.youtube.com/watch?v=Hth_tLaC2j8

Narrative violation: according to @Stanford research, local models can answer 71.3% of real-world chat and reasoning queries accurately, up from 23.2% in 2023. Obviously at a fraction of the cost and energy consumption of frontier APIs.
The obvious conclusion: you don't need a frontier model for most tasks. The future is multi-model: local, open-source, smaller and cheaper for the majority of workloads, frontier APIs when no other choices!

Feels like there’s a completely different set of best practices for AI builders on the $200 / month subsidized subscriptions vs employees working at companies that are trying not to overspend API costs

If you've adopted AI at your company but haven't seen any tangible results, read this 1990 article: "The Dynamo and the Computer" by Paul David.
When electricity first arrived, factories that "adopted" it barely got faster. They just swapped the steam engine for an electric one and ran everything else exactly as before: same machine layout, same workflow, same management. Electricity in, no real gains out.
The most common mistake with any new technology is to drop it into the old organization and then declare the transformation done.
The real leap came decades later, when each machine got its own small motor. Suddenly machines no longer had to be lined up around one central drive shaft. They could be rearranged around the actual flow of work.
The productivity gains didn't come from electricity. They came from REDESIGNING THE ENTIRE FACTORY around it.
AI is the same. Bolting it onto your existing process gets you a faster steam engine. The payoff comes when you redesign the work itself.
(link to paper in comments)

Wow. This is the way you want criminals to be talking about your anti-crime tech.
Rahul Sidhu: One of the most prolific criminals in all of San Francisco tells @adam22 that “crime in San Francisco is over with” because of Flock cameras + drones. He complains that he can’t even do drivebys anymore.
It’s simple: when the risk of getting caught is too high, crime plummets.