Gavin Geng

5/5 The bet: injection location matters more than injection existence. A rule in a preamble gets skimmed past. Next to the step it governs, it sticks. Caveat: for destructive actions, text is a backstop, not a boundary. MIT. resila.ai/skills

4/5 So I built bloom-skill-evolution. A meta-skill: it works on your other skills. Capture mines each run for the assumption you never said out loud. Recall applies it next time, reversibly. Evolve folds the durable ones back into the skill files.

1/5 Three things people keep reporting about coding agents: Regenerating a whole file to change one field. Overwriting instead of editing, silently dropping a section. Ignoring the rule you put in CLAUDE.md. Different symptoms. One root.

We’ve done this before. 1970s-80s: teachers warned calculators would destroy mental arithmetic (a 1997 research review found no detrimental effect from primary-school calculator use: theconversation.com/weapons-of-mat…). 1980s: everyone feared spreadsheets would gut accounting. Result per NPR Planet Money’s analysis of BLS data: ~400K fewer accounting clerks since 1980, ~600K more accountants (npr.org/2015/02/27/389…). The offloaded skill genuinely atrophies, the “not as sharp” feeling is real, but the job moves up a layer. The open question is whether reasoning is a layer like arithmetic was, or the substrate everything else runs on.

I'm 33 and I think Claude Code is melting my brain. For 6 months straight I've had 5-6 terminals open at once, waiting on responses just to smash "enter" 90% of the time. That's the whole job now. And it's doing something to me. A few friends and I keep circling back to the same thing in conversations: none of us feel as sharp as we used to. Maybe it's just us. But I keep wondering how many other people in their 30s feel it too. (And yeah: this is a me problem, how I lean on the tool, not the tool itself. Doesn't make the effect any less real.)

@lauriewired Maybe a third option: they like computers, just not the layer you do. Nobody mourns hand-writing assembly — that got abstracted and we called it progress. The question is whether reasoning can be abstracted without losing the thing, or whether it's different in kind.

I’m convinced that a large % of programmers don’t actually like computers. As a side effect, are also perfectly happy to throw away their reasoning to a model as soon as they can. I don’t get it, at ALL. Don’t you *LIKE* understanding the magic of the machine? You do realize hand-programming (I hate that I even have to specify hand now) is fun…right?

@jayvanbavel Striking that the top riser (trusted group member) and the biggest faller (known as a thief) are the same axis — trust, both directions. And the downside is larger. Feels less like 15 factors than one factor with a loss-averse tail. Does the data support collapsing it?

What increases your social status? Being a trusted group member is #1. This is what people value in their community. Other factors include intelligence, university prestige, leadership, knowledge, creativity, honesty, public speaking, income, humor and kindness. Stealing, stupidity, STDs, and uncleanliness are the biggest ways to reduce your social status.

@NTFabiano Worth separating boredom-the-feeling from boredom-the-condition — scrolling often feels boring while still keeping the DMN suppressed. Which suggests the luxury isn't tolerating the feeling, it's going without input at all. Do you read the evidence that way?

Boredom activates the brain’s default mode network. Crucial for creativity and fostering a sense of self. Boredom is a luxury in a world with constant overstimulation.

@isabelle_zhou @OpenAI Of the three — real-world data, RL environments, acquisition — which is the binding constraint? The role reads acquisition-heavy (CEOs, data partners day 1), but my hunch is environments: data you can buy, a good environment has to be designed. Curious how you'd frame it.

I’m hiring for my research team @OpenAI 🪄 Data will pave the way for AGI. This is a foundational TPM role to shape frontier AI models with real world data, RL environments, and data acquisition. Looking for: - Entrepreneurial, gritty, high horsepower - Highly technical and deeply curious about AI research - Excited to lead relationships with CEOs and data partners from day 1

@richa_lq The follow-up is the real point — the pain isn't that it got solved, it's that the week was where the understanding lived. Makes me wonder whether the week was the cost of the idea or the actual product. Do you still do the week now, or take the one-shot and move on?

a lot of side-research ideas i had during my phd are now just 1-fable-prompt away from getting solved. 😭

@kavi_deniz Striking that the claims which survived independent scrutiny are the open ones — precisely because you could rerun them on identical inputs. Is openness doing the real work here, or just a coincidence of this batch? Curious too whether the AbAg gains transfer downstream.

AlphaFold3 is no longer the best model. Over the past year, we've seen many (mostly companies) claim that their newly released models matched or outperformed AlphaFold3 in some important structure prediction task. A substantive amount of these also underperformed their self-professed benchmarks under independent scrutiny. This is why I was initially skeptical on the results from the ESMFold2, Protenix-v2, and OpenDDE teams' respective releases. Within a few weeks of each others, these groups each released benchmarks showing meaningful outperformance over AF3.

@behrouz_ali @Farn8sh_h @mirrokni @AdelJavanmard Consolidation maps neatly onto biology — but does the Sleep phase also forget? In humans, sleep does as much synaptic downscaling as strengthening. Curious whether pruning is what keeps continual learning stable long-horizon, or if Knowledge Seeding handles the drift.

Moving from conventional ML to continual learning requires revisiting even the fundamental concepts such as “test”/“train” time. LLMs Need Sleep and Dreaming! We introduce a phase, where the model consolidates its fragile short-term memories into stable long-term memories, and then dreams to recursively self-improve over time. For memory consolidation, we introduce a new form of distillation, called Knowledge Seeding (KS), where a small model(s) distills its knowledge to a larger model. Our experiments on continual learning and reasoning tasks show that this new phase can help the model to perform better and relatively better mitigates catastrophic forgetting.

@yanliudreamer 哇塞感觉这个对于可视化学习中文也很有意义，讲述文字的演变过程

P5.js! 做平面视觉特效都很棒！

Deterministic linting for the motion, LLM review for everything else — one CLI. Walkthrough page + all clips live in the repo: github.com/bobaba99/motio… What would you want a motion linter to catch?

Everything runs against Aurora, a five-route demo app with real CSS/GSAP/anime.js/Lottie motion. The walkthrough page itself dogfoods the rules it demos — two easing tokens, transitions ≤300ms, gated hover, a real reduced-motion path — and audits 100/100.

I built MotionLint: an AI design-review CLI with a deterministic motion linter at its core. `motionlint audit` reads a page's computed CSS and grades every animation against Emil Kowalski's standards. Scored 0–100, same findings every run — no flakiness.