The Wake is a daily briefing from George's saved internet. The issue is written as a newsletter first. The tweets are the source material, preserved below for receipts.
Source window: May 20, 2026. Signals: 9 bookmarks and 2 likes.
Brief
Developer productivity is back on the front page. Not because of a single shiny model, but because the plumbing that lets humans and AI cooperate is finally being built in public. The conversation today runs from the simple agony of a laggy pull request UI to public reinforcement learning checkpoints, to autocomplete tools that feel like an extra pair of hands. The throughline is velocity: small UX frictions cost expertise, AI flattens cognitive edges, and what remains valuable is design, integration, and safety. Expect a sprint of tooling bets and a tightening debate over who captures the surplus.
The latency tax
When a senior engineer says they literally take their hands off the keyboard because a PR page skips frames, you should listen. Slow tooling is not an annoyance. It is a productivity tax that compounds across teams and months. Mitchell Hashimoto's complaints about PR diff speed are a reminder that core workflows are fragile. Where humans are optimized for pattern recognition and rapid context switching, software that lags breaks thought. The result is wasted attention, delayed reviews, and, worse, a tendency to simplify work in service of the tool rather than the product.
That is where the low-hanging infrastructure wins live. Faster diffs, instant inline comments, and smooth keyboard-driven review loops reduce cognitive context switching. We saw the positive flip side in the same stream: a performant review demo showed that the limiting factor is often execution, not architecture. In other words, the improvements are reachable. Teams that treat PR UX as a feature will outpace those that treat it as background noise.
Why this matters now: as models get better at code, human reviewers will shift to higher-order judgments. If the UI still chews time, that human capital is squandered. Fix the latency problem and you multiply the benefit of every automation layer you add on top.
Autocomplete as a second brain
Autocomplete is no longer a gimmick. Tools like Cotypist are moving from helpful to indispensable, turning predictable scaffolding into live muscle memory for developers. When code completion works everywhere, it shrinks the gap between junior and senior productivity. That is the same observation in different language: AI is leveling certain aspects of cognitive capability.
The provocative read here is the idea that AI flattens variance in intelligence for competitive tasks. With good tooling, the upside of raw cognitive firepower shrinks; what matters next is taste, design, and domain intuition. The colorful line about "competing on beauty" captures this: if models and assistants handle correctness, differentiation shifts to product aesthetics, ergonomics, and trust. That redistributes value from pure engineering virtuosity to multidisciplinary teams that can make systems feel delightful and reliable.
This is not purely utopian. As autocomplete becomes a ubiquitous productivity multiplier, the capture of value will depend on integration and data ownership. Whichever editor, IDE, or platform gives the cleanest, least disruptive assistance will anchor a lot of developer workflows. That makes investment in subtle UX details a defensible moat.
Open RL and the speed of iteration
Public research artifacts are changing how quickly new capabilities land in practice. A team publishing an explainer, environments, and training checkpoints lets practitioners iterate faster than traditional paper-and-wait cycles. Open checkpoints are not just signals of transparency. They are multiplier events: people will run the agents, find failure modes, fine-tune them for new domains, and build tools on top of them.
This matters for two reasons. First, reinforcement learning and aligned fine-tuning are where behavior transforms from static prediction to persistent strategy. When those methods are shared openly, the community can converge on robust practices faster. Second, having reproducible environments and checkpoints changes the economics of experimentation. Small teams can stand on the shoulders of large labs and ship useful, specialized agents quickly.
There is an obvious counterweight: public checkpoints also accelerate the discovery of failure modes. The schizophrenia activation report on a mainstream multimodal model is a reminder that improving capability without understanding downstream harms is reckless. Open research expedites both innovation and auditing. Expect a spike in third-party tooling that automates safety tests and behavioral evaluation for released checkpoints.
Product engineering, components, and the new craft
The social proof for shipping matters. When engineers publish components that claim to be "unblankable" and absurdly tolerant, it signals a new kind of product engineering ethos. Rapid iteration, generous public components, and practical documentation move more value into integration skills than bespoke research. Pierre's work on high-performance PR review tooling and community-built components are examples: the guy shipping fast and clean will win the developer mindshare.
That dynamic also reframes open source. We are not returning to the 2010s worship of monolithic "everything libraries." The modern playbook favors small, composable components that survive extreme misuse. Teams that focus on predictable, well-documented building blocks make the rest of the stack easier. And because everyone is using similar base models and autocompletion, stability and ergonomics of the components become a stronger differentiator than raw capability.
One practical consequence: hiring will tilt toward product engineers who can ship delightful developer UX, not just researchers who can train slightly better checkpoints. The gatekeepers of developer productivity are shifting from ML labs to product engineering teams that keep the latency low and the suggestions accurate.
Scale, surplus, and the billionaire question
Jeff Bezos' thought experiment about the burger chain compresses a broader economic argument about scale. If you create something that millions prefer, wealth accrues. In the context of AI, the scalable product is not necessarily the biggest model. It is the service that integrates intelligence, handles safety, and becomes the habit. The wrinkle is that as AI flattens cognitive differences, scale consolidates around experience and trust. That is where the debate over whether massive wealth from scale is "ethical" becomes practical rather than moral.
Scale will also drive more verticalization. Companies that own distribution and developer interfaces will be able to monetize incremental improvements across many customers. That amplifies the value of shipping high-quality integrations and stable components. Expect capital to follow teams that demonstrate both high velocity and sticky user experiences.
What to watch
- PR UX metrics: page render times, keyboard latency, and inline diff responsiveness in flagship forges. Small improvements yield outsized productivity returns.
- Autocomplete everywhere: adoption curves for integrated completion tools and how they change onboarding and code review throughput. @steipete's endorsement is an early signal.
- Open RL checkpoints: who forks, fine-tunes, and ships derivative agents. Look for safety audits and failure-mode reports to appear in parallel.
- Hallucination and safety headlines around large multimodal models. The "schizophrenia activation" report is likely an early wave, not the last.
- Component-first product plays: teams publishing highly robust, well-documented UI and infra parts. Watch who becomes indispensable to other builders.
- The distribution bets: which platforms make the assistant a default and whether that drives concentration of revenue.
Keep an eye on the small frictions. They look mundane, but they are the gates through which the productivity gains from models must pass. Remove the friction, and the surplus becomes yours to capture.
Source tweets
Evil Rabbit / @evilrabbit_
- bookmark: open on X
- This.
tomie / @tomieinlove
- bookmark: open on X
- @captgouda24 The prostitute comment is quite apt. God made men differ in strength, but Sam Colt made them equal. Likewise, God made men differ in intelligence, but Sam Altman made them equal. We’re now competing on beauty. The concern is not if scaling has hit a wall. It’s if you have.
michelle / @michellechen
- bookmark: open on X
- check out our blog for an explainer on RL and how we put this research together. our repo, environments are all public and you can even try the various training checkpoints:
Mitchell Hashimoto / @mitchellh
- bookmark: open on X
- This is why PR diff speed matters. This isn't a dunk on GitHub specifically, because GitLab, Forgejo, etc. are all equal or worse. But this is the kind of thing that drives me nuts, because this is a core workflow and its slow enough I literally take my hands off the keyboard. Btw, when my mouse jiggles on the left, its because the page is literally skipping frames and I'm instinctively shaking my mouse to see if it'll respond. And on the keyboard input you can literally here me finish typing before a letter even shows up. For someone like me who is an expert at these tools, my brain navigates the tool dramatically faster than it can keep up, and that is not good. The tool should not get in the way. the post also includes media
George Tziralis / @gtzi
- bookmark: open on X
- Welcome Pickups FTW!
Mitchell Hashimoto / @mitchellh
- bookmark: open on X
- This is how performant PR review could be. On any forge. Pierre is showing us that the only thing holding that back is a skill issue. Excellent ship here! They’re on fire!
Christopher Hooks / @cd_hooks
- bookmark: open on X
- In addition to the fact that no one wants this, Google Gemini has recently become a schizophrenia activation machine the post also includes media
Saurabh Kumar / @drummatick
- bookmark: open on X
- I owe my entire ML career to @lilianweng and her blogs Specially on beta-VAE. That blog changed the way I saw generative modeling
Peter Steinberger 🦞 / @steipete
- bookmark: open on X
- Can't recommend @cotypist enough. Autocomplete everywhere.
WOLF / @WOLF_Financial
- like: open on X
- JEFF BEZOS WAS ASKED ABOUT THE CLAIM THAT "YOU CAN'T EARN A BILLION DOLLARS" His entire response was a burger joint: "Let's say you start a burger joint with 10 employees. People love your burgers, so you open a second outlet. Then a third. By the time you've opened 1,000 outlets, you are a billionaire." "This is a real life story. It happens all the time. In-N-Out Burger. Raising Cane's. At what point did that money all of a sudden become unethical? It didn't." His point: "If you create a service that people love and millions of people choose it, you're going to end up with $1 billion." Do you agree? the post also includes media
amadeus / @amadeus
- like: open on X
- i'm really more of a product engineer than an opensource engineer, but @fat bullied me into making diffs, so
here's a component that you can just throw unrealistic amounts of bullshit shit at and it should just work. my goals were, unblankable, absurd, and a waste of pierre's vc money building shit nobody realistically needs. also it's still pretty unoptimized, so i got a lot more work to do. please use it, find bugs and let me know! 🙏
Generated from Birdclaw bookmarks and likes. Edited by Ody before publication.