Kevin Connolly

Researcher at heart (physics PhD), fascinated by data, tech, piano, languages, literature - possibly everything. Most waking hours @ashbyhq.

[Kevin Connolly @wintercarver]

Have 100% been thinking about @stewartbrand's pace layer's too, have been discussing "layers" and their rates of change quite a lot with my team lately.

Am personally more interested in everything in the "days - months" range and for day-to-day system design in the context of applying AI across a large team, think that's where a lot of uncertainty/value is (for my area of concern).

[Kevin Connolly @wintercarver]

@astrodanish May be more appropriately called regret or ambition, depending on your degree of agency and light cone coordinates.

[Kevin Connolly @wintercarver]

For the more visually inclined.

[Kevin Connolly @wintercarver]

Started reading All you Need is Kill the other day - it's the Japanese sci-fi novel that the movie Edge of Tomorrow is based on. Apparently, the title, which stands out to me as one of the best book titles I've ever come across, is considered to be a dark variation of the Beatles "All You Need is Love".

[Kevin Connolly @wintercarver]

Though, only one can be right...

[Kevin Connolly @wintercarver]

Paradoxical factoid: the original meaning of the word 'factoid' meant a false or fabricated statement presented as the truth.

[Kevin Connolly @wintercarver]

Hey @sama why don't you hook this boss up with some of that sweet, sweet 750 tok/s technology? (neutral third party opinion, fan of progress).

[David Turturean @DavidTurturean]

I solved my 3rd Erdős problem, #870, with ChatGPT-5.5-Pro, then verified it by formalizing the whole proof in Lean 4, sorry-free and axiom-free! With about 180,000 lines of Lean code, as far as I can find, no one person has formalized a single problem at this scale. 🧵 1/n

[Kevin Connolly @wintercarver]

@docmilanfar Yeah, just like Bach: "I was obliged to be industrious; whoever is equally industrious will succeed equally well." - just a matter of effort, haha...

[Kevin Connolly @wintercarver]

Lots of fun tidbits in this discussion on Machiavelli from @dwarkesh_sp with @Ada_Palmer.

https://t.co/3R1fyRk3rm

[Kevin Connolly @wintercarver]

Reflecting on my own journey through physics, one salient point is the slow transition over many years from a focus on fundamental truths, nature of the universe type pursuits, that are fundamentally based in validating or invalidating hypotheses with a focus on theory, to a growing (and still increasing) appreciation for pure engineering. Even as a lowly experimental physicist, I'd say my center of gravity, intellectually, was in the concept of theory and foundations of nature. I'd work on giant machines, detector hardware, calibration, analysis - the limits of engineering were right in front of me, indeed, part of my day-to-day work for years, but my frame of mind was in "nature of the universe" theory, not "excellence in engineering."

If I were to do it again, and much credit of course goes to Musk in popularizing the recognition that science and engineering are inextricably linked and that much of science is driven by engineering. Certainly a lot of nuance in there, but without getting too worked up in the silly version of that debate, I'll just defer to Feynman's famous quote about theory vs reality, and reality is governed by measurement, and measurement is governed by tools, and tools/machines are governed by engineering. Better engineering means bettter science (but not all engineering is about science, of course).

Much of the history of physics is a story of tool breakthroughs and innovations, and I wish the importance of engineering mastery and innovation was emphasized more heavily in that telling.

[Kevin Connolly @wintercarver]

@pitdesi Indeed... unlikely that any critique raised within first 24-48 hours is on the "oh yeah, we failed to really consider that fundamental flaw" list. Similarly have no idea if they will succeed, but feel confident failure will not come from any reason being noted today.

[Kevin Connolly @wintercarver]

@robbietilton @DavidSHolz Evangelion-generation geek here: Rei in large vat is End of Evangelion, not TV series. But love the comparison, color theme is on point.

[Kevin Connolly @wintercarver]

@DavidSHolz Can't help but see this as a mini Super-Kamiokande - @DavidSHolz any inspiration therein? Obv many differences and super cool. Congrats!

[Kevin Connolly @wintercarver]

一寸先は闇。

[Marc Andreessen 🇺🇸 @pmarca]

Five dimensional chess doesn’t exist. Everyone is furiously improvising all the time. The future is utterly uncertain.

[Kevin Connolly @wintercarver]

Curious, if you’ll entertain the question, what prior works on thermodynamics were an influence or inspiration in your approach to the topic? Stat mech and thermodynamics were probably one of my favorite topics in undergrad/grad studies (personally, enjoyed Landau & Lifschitz the most, tho various texts are enjoyable in their own ways).

[Kevin Connolly @wintercarver]

A modern and somewhat niche lamentation, but as an appreciator of software development and product craft, it���s too bad we have ended up here with release notes. Yes yes, it takes time and effort, but this represents a certain disregard for the user. Kudos to @GeminiApp to at least noting new features periodically… can’t you at least have your agents throw us product nerds a bone, @ChatGPTapp @claudeai @grok ?

[Kevin Connolly @wintercarver]

@shadcn Not quite clear to me from your sketch, but curious: is that one giant ongoing "note" or multiple distinct notes? If the latter, where do you envision each individual note going after it is made? Do they become implicitly pinned message equivalents?

[Kevin Connolly @wintercarver]

@thsottiaux I’m decidedly not a tokenmaxxer by skill or disposition, but it did feel like a technological marvel of sorts to check in on a task running on my laptop at home with ease from my phone to see ~1.5M tokens over a ~90min run completed. Flow, the tokens did.

[Kevin Connolly @wintercarver]

Kind of odd to reflect on the odds of the Riemann Hypothesis being proven by AI or in deep collaboration with AI going from generally ~0% just a few years ago to generally not-0%. No irrational exuberance here, just pondering it - though I do find it an exciting prospect.

A proof, regardless of origins, would be a welcome win, though I s'pose it may induce an intellectual existential wave for some. With the recent result from OpenAI disproving the planar unit distance problem, it does make me wonder whether getting a list of open problems written down and committed to as "significant indicators" now would be of any use. There is a certain propensity to rationalize away these results post facto, and the breadcrumbs of success are rapidly approaching full loaf-sized realities.

[Kevin Connolly @wintercarver]

"fit vision of the project" - imo probably going to be the longest standing challenge in end-to-end product building, but who knows, AI-related predictions are tough. This is apropos of "taste" being the trendiest tech word of 2026: there is just an unbelievably deep and fractal-like space of opinion for any service/product/feature/sub-feature, deep or trivial alike, that simply shipping it off for production is risky with respect to what you're going to get out at the end.

Interestingly, I'd wager for new, smaller projects the risk is highest as the constraints are lowest (i.e. not much has been built). The larger a framework of product surface exists, the more technically challenging any development decision becomes, but the *taste* factor becomes progressively, implicitly defined. I suspect this is why a lot of vibe coding side projects plunge into the Miyazaki Mire of despair, amongst other factors...

Anyway, I like this this idea of VISION.md. Good exercise for establishing ones priorities, principles and standards. Probably a good example of easy concept, hard problem.

[Peter Steinberger 🦞 @steipete]

I built an autotriage skill for codex that has a set of guidelines + reads VISION.md from my repos, so issues/prs that have a clear way of
- fit vision of the project
- being inferrable in code with high confidence
- clear fix
- can be live tested

Are now worked on autonomously. Codex can use a VM + computer vision (via https://t.co/2T5aNF5jTT , new parallels backend) to verify fixes, so it can work without interrupting me. I manually review suggestions. Since it was tedious to type in issues, I added an issue browser into https://t.co/NfEoHIQPil that parses common clipboard formats by codex so I can click through them conveniently.