Kevin Lacker

I'm happy to announce the launch of Acorn, a new theorem prover that includes an integrated AI. Theorem provers let you write mathematical proofs that are rigorously verified. But they are notoriously difficult to use. Acorn makes it easier, by using AI to fill in the details.

@angeris seems interesting

ok, weird, but I have a list of sci-fi books I've been meaning to read for a while, but it seems kind of boring to read them and then write some post about it or whatever so, any mutuals interested in a (hard-ish) sf mini-book club like thing

@ptrschmdtnlsn My advice is, don't run an internship in order to get one hardware feature working. Better to hire a contractor for this

Does anyone have advice on running internships? I'm sort of wondering what it would look like if my start-up took on one short-term intern to try to get one hardware feature working. I'm slightly concerned that the advice "internships are just for hiring, not output" holds here.

@spolu Humanity is still much funnier than the AIs

Writing, Coding, and Style Great writing may be the last refuge of human superiority on models. It's quite surprising to me that we still detect fairly well AI slop vs great writing but have all, as software engineers, accepted AI generated code as the new norm. We used to pride ourselves in our craft, sometimes even comparing code to an art or poetry. Let's be honest. It didn't take long for us to jump on the bandwagon and flood our codebases with... spolu.sh/notes/20260318…

@Miles_Brundage Products have an equilibrium of forces. Adding more features: 1. Makes the product better. (Good) 2. Makes the product buggier. (Bad) 3. Costs $ in engineering time. (Bad) When you weaken constraint 3, the equilibrium naturally shifts towards "more features" and "more bugs"

I'm a bit worried Anthropic has an org-wide case of AI psychosis that makes them think Claude is good enough that they can ship random product features without breaking things, but in fact they *do* keep breaking things, and they're not online enough to notice people complaining

@csvoss I wonder if it is because of you, personally. Is any real-life Voss closer to being the classic archetype of a science-fictional character? The LLMs don't just invent these associations out of nothing....

July 22, 2025. Day 1 of people discovering “Voss” is the stereotypical AI sci-fi last name. Finally, objective proof that my last name is excellent! Hahaha! Yes! Yeah! March 18, 2026. Day 239 of “Voss” being the stereotypical AI sci-fi last name. I am sent TikToks with slop plotlines revolving around the surname’s “rareness.” I slowly realize, with dawning horror, that the algorithms of the future will refuse to believe I’m not a bot. Is it karma? The feed will derank me. In the 2030s, my personhood credit score under Reformed Neo-Worldcoinism will tank. In the 2040s, the Pangram Labs Pan-Galactic Dyson Swarm will flag not just my own words, but any text mentioning me (or any of my ancestors) as nonhuman, dooming even those who knew me to be forgotten, until the 2050s, when anyone still in possession of the forbidden memories of the word “delve,” the word “tapestry,” or my name will find their uploaded em is mysteriously banned from logging in to the newly-upgraded (4D VR) X HyperSpaces (by HyperSpaceX). Before long, even this tweet will vanish into the oblivion of time— —I see you’re trying to read a story about a person named Voss. If you want, I can swap in a surname that scores as more genuinely human for improved narrative credibility. Should we do that next?

@rndhouse @e2 @e9 @L2 @F1 @S3 @p3 @e4 @E5 @l3 it does seem like you want a lot of these internal references for reasoning about the code. something like "when you access arr[i] here, i is actually less than n" usually involves a lot of tracing around looking at when things were set to what

More generally: Rather than sharing code as text, we should share representations closer to the program structure already captured in IDE indexes and related metadata. Then patches become: ``` replace @e2: (@e9 x * @l2 2) insert @f1.body[ah]: @s3 let @p3 z = @e4 (@e5 y + @l3 1); ``` And we can refer precisely to specific parts of the code at a higher structural level.

Eventually the agents are going to write most of the code. What sort of programming language do they prefer?

@lymanstoneky charging them extra for electricity seems reasonable, especially since datacenters are often limited by that. align the incentives

The simple way to do this is actually not to tax AI company profits per se, but to tax data centers on some volumetric measure, like per khw or something.

@n0riskn0r3ward yeah. just knowing “this repair should cost about x dollars” can save you a lot of money

feels very different when LLMs help me navigate getting my car fixed without getting ripped off than it does when i use it to write code. the gentlemen at the car repair shop were a little surprised when I told them I didn’t think they I needed to wait for the diagnostics from a Techstream/TIS and that an OBD II or similar aftermarket scanner should be able to diagnose the fuel injection issue on my older Camry just fine. Turns out Opus 4.6 and GPT-5.4 are correct about that.

@sutherlandphys my kids just really love both mining and crafting

Genuinely why does anyone use a Microsoft product

@proudmomom @minilek @DavidSacks @theallinpod just raise prices, industry is happy to pay cs grads enough to make the degree +EV

@minilek @DavidSacks @theallinpod CS programs should be supported by industry

Interesting to juxtapose the collapse of our CS major due to high instructional costs, with @DavidSacks three days ago on @theallinpod bemoaning a "massive supply shortage" of software engineers.

@minilek It would make a lot of sense to charge more for a CS degree

Lest there be any confusion, the collapse of the CS major is _not_ due to a lack of student demand. Each 'graduation year' covers a full academic year ('13 is AY 2012-13). The blue plot is the % of applicants to that graduating cohort (both freshmen and junior xfer) who applied to UC Berkeley intending to major in CS. The orange plot is the % of bachelor's degrees awarded that academic year that were in CS. '27 is a projection, but should be fairly accurate since the bulk of graduates either start off CS, or change major into CS by sophomore Spring (which for the Class of 2027, was last year).

@eshear @docmilanfar I feel like at some point, the helical nature is no longer inherently in the problem itself, it’s in the machinery that you have chosen to solve or understand the problem

@lacker @docmilanfar It’s helical in this instance bc there’s a winding count that matters, tho it’s quotiented out in this case so it’s harder to see.

When you see e, you know you’re seeing a self-scaling process. When you see π you know you’re seeing a periodic process. When you see i, you know you’re seeing helical process. When you see √ or ², you know you’re seeing a process half completed or a process doubled.

@eshear @docmilanfar Find the integral of 1 / (x^2 + 1)

@lacker @docmilanfar Give an example!

@eshear @docmilanfar you might use i while solving systems of equations that don’t have anything to do with helical processes

@docmilanfar Give an example please

@cayimby “inclusionary fees” is a pretty funny phrase

A San Luis Obispo homebuilder paid nearly $100,000 in inclusionary fees for an 8-home project, or would have had to sell a home for $450,000, which cost $1.325M to build. Fees like these block many projects before they start, weakening state efforts to spur construction.

@samth @AndyMasley it’s a good way to emphasize that you’re trying to be polite about it, when telling someone that you think they are wrong

@AndyMasley "directionally correct" is the rationalist version of "the fact that I believed it tells you a lot"

I was wrong and think this term is useful now based on the comments, sorry folks

@intspg @Math_files this seems obvious to me

@Math_files This seems disturbing - how can you look at something and say the proof is obvious when it is not?

One day, mathematician Shizuo Kakutani was teaching a class at Yale University. During the lecture, he wrote a lemma on the blackboard and told the students that its proof was obvious. A student hesitantly raised his hand and said that the proof was not obvious to him. He politely asked if Professor Kakutani could explain it. Kakutani paused for a moment and tried to think through the argument. After a while, he realized something surprising—he could not immediately prove the lemma himself. Smiling a little awkwardly, he apologized to the class and said he would return with the proof in the next lecture. After the class ended, Kakutani went straight to his office and began working on it. He spent a long time trying different ideas, but the proof would not come. The problem kept bothering him. Eventually, he decided to search the library for the original source of the lemma. After some effort, he finally located the paper where the lemma first appeared. The statement was written clearly. Then he looked at the proof. Instead of a detailed explanation, it simply said: “Exercise for the reader.” At that moment, Kakutani realized something amusing—the paper had been written years earlier… by himself. It was a gentle reminder that even great mathematicians sometimes forget the details of their own work.

link: substack.com/home/post/p-18…

The "Bismarck theory" of how AI will take over the world. Eventually the AI will be able to do your job for you, if you want. And over time, everyone will choose to do that.