So many CRITICAL and MANDATORY steps in the release instruction file. As it always is with AI, if it doesn’t work, just use more forceful language and capital letters. One more CRITICAL bullet point bro, that’ll fix everything.

Sadly, I am not too surprised by the developers of Lean turning towards AI. The AI people have been quite interested in Lean for a while now since they think it is a useful tool to have AIs do math (and math = smart, you know).

There are some comments speculating that some pro-AI people try to infiltrate anti-AI subreddits by applying for moderator positions and then shutting those subreddits down. I think this is the most reasonable explanation for why the mods of “cogsuckers” of all places are sealions for pro-AI arguments. (In the more recent posts in that subreddit, I recognized many usernames who were prominent mods in pro-AI subreddits.)

I don’t understand what they gain from shutting down subreddits of all things. Do they really think that using these scummy tactics will somehow result in more positive opinions towards AI? Or are they trying the fascist gambit hoping that they will have so much power that public opinion won’t matter anymore? They aren’t exactly billionaires buying out media networks.

Don’t forget the other comment saying that if you hate AI, you’re just “vice-signalling” and “telegraphing your incuruosity (sic) far and wide”. AI is just like computer graphics in the 1960s, apparently. We’re still in early days guys, we’ve only invested trillions of dollars into this and stolen the collective works of everyone on the internet, and we don’t have any better ideas than throwing more money compute at the problem! The scaling is still working guys, look at these benchmarks that we totally didn’t pay for. Look at these models doing mathematical reasoning. Actually don’t look at those, you can’t see them because they’re proprietary and live in Canada.

In other news, I drew a chart the other day, and I can confidently predict that my newborn baby is on track to weigh 10 trillion pounds by age 10.

EDIT: Rich Hickey has now disabled comments. Fair enough, arguing with promptfondlers is a waste of time and sanity.

I went deep into the Yud lore once. A single fluke SAT score served as the basis for Yud’s belief in his own world-changing importance. In middle school, he took an SAT with a score of 670 verbal and 740 math (maximum 800 each) and the Midwest Talent Search contacted him to tell him that his scores were very high for a middle schooler. Despite his great pains to talk about how he tried to be humble about it, he also says that he was in the “99.9998th percentile” and “not only bright but waayy out of the ordinary.”

I was in the math contest scene. I have good friends who did well on AP Calculus in middle school, and were skilled enough at contests that they would have easily gotten an 800 on the math SAT if they took it. Even so, there were middle schoolers who were far more skilled than them, and I have seen other people who were far less “talented” in middle school rise to great heights later in life. As it turns out, skills can be developed through practice.

Yud’s performance would not even be considered impressive in the math contest community, let alone justify calling him one of the most important people in the world. Perhaps at the time, he didn’t know better. But he decided to make this a core part of his self-identity. His life quickly spiraled out of control, starting with him refusing to attend high school.

It is how professors talk to each other in … debate halls? What the fuck? Yud really doesn’t have any clue how universities work.

I am a PhD student right now so I have a far better idea of how professors talk to each other. The way most professors (in math/CS at least) communicate in a spoken setting is through giving talks at conferences. The cool professors use chalkboards, but most people these days use slides. As it turns out, debates are really fucking stupid for scientific research for so many reasons.

1. Science assumes good faith out of everyone, and debates are needlessly adversarial. This is why everyone just presents and listens to talks.
2. Debates are actually really bad for the kind of deep analysis and thought needed to understand new research. If you want to seriously consider novel ideas, it’s not so easy when you’re expected to come up with a response in the next few minutes.
3. Debates generally favor people who use good rhetoric and can package their ideas more neatly, not the people who really have more interesting ideas.
4. If you want to justify a scientific claim, you do it with experiments and evidence (or a mathematical proof when applicable). What purpose does a debate serve?

I think Yud’s fixation on debates and “winning” reflects what he thinks of intellectualism. For him, it is merely a means to an end. The real goal is to be superior and beat up other people.

Choice quote from Dave Karpf:

Policy moderation can never fail. It can only be failed.

Yeah, it’s not like reviewers can just write “This paper is utter trash. Score: 2” unless ML is somehow an even worse field than I previously thought.

They referenced someone who had a paper get rejected from conferences six times, which to me is an indication that their idea just isn’t that good. I don’t mean this as a personal attack; everyone has bad ideas. It’s just that at some point, you just have to cut your losses with a bad idea and instead use your time to develop better ideas.

So I am suspicious that when they say “constructive feedback”, they don’t mean “how do I make this idea good” but instead “what are the magic words that will get my paper accepted into a conference”. ML has become a cutthroat publish-or-perish field, after all. It certainly won’t help that LLMs are effectively trained to glaze the user at all times.

AI researchers are rapidly embracing AI reviews, with the new Stanford Agentic Reviewer. Surely nothing could possibly go wrong!

Here’s the “tech overview” for their website.

Our agentic reviewer provides rapid feedback to researchers on their work to help them to rapidly iterate and improve their research.

The inspiration for this project was a conversation that one of us had with a student (not from Stanford) that had their research paper rejected 6 times over 3 years. They got a round of feedback roughly every 6 months from the peer review process, and this commentary formed the basis for their next round of revisions. The 6 month iteration cycle was painfully slow, and the noisy reviews — which were more focused on judging a paper’s worth than providing constructive feedback — gave only a weak signal for where to go next.

How is it, when people try to argue about the magical benefits of AI on a task, it always comes down to arguing “well actually, humans suck at the task too! Look, humans make mistakes!” That seems to be the only way they can justify the fact that AI sucks. At least it spews garbage fast!

(Also, this is a little mean, but if someone’s paper got rejected 6 times in a row, perhaps it’s time to throw in the towel, accept that the project was never that good in the first place, and try better ideas. Not every idea works out, especially in research.)

When modified to output a 1-10 score by training to mimic ICLR 2025 reviews (which are public), we found that the Spearman correlation (higher is better) between one human reviewer and another is 0.41, whereas the correlation between AI and one human reviewer is 0.42. This suggests the agentic reviewer is approaching human-level performance.

Actually, now all my concerns are now completely gone. They found that one number is bigger than another number, so I take back all of my counterarguments. I now have full faith that this is going to work out.

Reviews are AI generated, and may contain errors.

We had built this for researchers seeking feedback on their work. If you are a reviewer for a conference, we discourage using this in any way that violates the policies of that conference.

Of course, we need the mandatory disclaimers that will definitely be enforced. No reviewer will ever be a lazy bum and use this AI for their actual conference reviews.

I’m a nerd and even I want to shove this guy in a locker.

Referencing the telephone game does not prove anything here. The telephone game is shows that humans are not good at copying something exactly without changes, which computers are better at. But the question here is if AI can achieve deeper understanding of a work, which is needed to produce a good summary. This is something humans are far better at. The AI screws up the summary here in ways that no reasonable person who has watched the TV series (or played the games) would ever screw up.

A game/sneer where you are a venture capitalist with billions invested in generative AI: https://woe-industries.itch.io/you-have-billions-invested-in-generative-ai

Although I never use LLMs for any serious purpose, I do sometimes give LLMs test questions in order to get firsthand experience on what their responses are like. This guide tracks quite well with what I see. The language is flowery and full of unnecessary metaphors, and the formatting has excessive bullet points, boldface, and emoji. (Seeing emoji in what is supposed to be a serious text really pisses me off for some reason.) When I read the text carefully, I can almost always find mistakes or severe omissions, even when the mistake could easily be remedied by searching the internet.

This is perfectly in line with the fact that LLMs do not have deep understanding, or the understanding is only in the mind of the user, such as with rubber duck debugging. I agree with the “Barnum-effect” comment (see this essay for what that refers to).

The most obvious indication of AI I can see is the countless paragraphs that start with a boldfaced “header” with a colon. I consider this to be terrible writing practice, even for technical/explanatory writing. When a writer does this, it feels as if they don’t even respect their own writing. Maybe their paragraphs are so incomprehensible that they need to spoonfeed the reader. Or, perhaps they have so little to say that the bullet points already get it across, and their writing is little more than extraneous fluff. Yeah, much larger things like sections or chapters should have titles, but putting a header on every single paragraph is, frankly, insulting the reader’s intelligence.

I see AI output use this format very frequently though. Honestly, this goes to show how AI appeals to people who only care about shortcuts and bullshitting instead of thinking things through. Putting a bold header on every single paragraph really does appeal to that type.

Chicken penne is now rivaling spaghetti Bolognese as a core layer of pasta infrastructure.

If rationalists could benefit from just one piece of advice, it would be: actions speak louder than words. Right now, I don’t think they understand that, given their penchant for 10k word blog posts.

One non-AI example of this is the most expensive fireworks show in history, I mean, the SpaceX Starship program. So far, they have had 11 or 12 test flights (I don’t care to count the exact number by this point), and not a single one of them has delivered anything into orbit. Fans generally tend to cling on to a few parlor tricks like the “chopstick” stuff. They seem to have forgotten that their goal was to land people on the moon. This goal had already been accomplished over 50 years ago with the 11th flight of the Apollo program.

I saw this coming from their very first Starship test flight. They destroyed the launchpad as soon as the rocket lifted off, with massive chunks of concrete flying hundreds of feet into the air. The rocket itself lost control and exploded 4 minutes later. But by far the most damning part was when the camera cut to the SpaceX employees wildly cheering. Later on there were countless spin articles about how this test flight was successful because they collected so much data.

I chose to believe the evidence in front of my eyes over the talking points about how SpaceX was decades ahead of everyone else, SpaceX is a leader in cheap reusable spacecraft, iterative development is great, etc. Now, I choose to look at the actions of the AI companies, and I can easily see that they do not have any ethics. Meanwhile, the rationalists are hypnotized by the Anthropic critihype blog posts about how their AI is dangerous.

You cannot honestly call it “trust” if you still have to go through the output with a magnifying glass and make sure it didn’t tell anyone to put glue on their pizza.

When any other technology fails to achieve its stated purpose, we call it flawed and unreliable. But AI is so magical! It receives credit for everything it happens to get right, and it’s my fault when it gets something wrong.

Unfortunately, I don’t think anyone is ever going to go through all 19,797 submissions and 75,800 reviews (to one conference, in one year) and manually review them all. Then again, using the ultra-advanced cutting-edge innovative statistical technique of randomly sampling a few papers/reviews, one can still get useful conclusions.

After the bubble collapses, I believe there is going to be a rule of thumb for whatever tiny niche use cases LLMs might have: “Never let an LLM have any decision-making power.” At most, LLMs will serve as a heuristic function for an algorithm that actually works.

Unlike the railroads of the First Gilded Age, I don’t think GenAI will have many long term viable use cases. The problem is that it has two characteristics that do not go well together: unreliability and expense. Generally, it’s not worth spending lots of money on a task where you don’t need reliability.

The sheer expense of GenAI has been subsidized by the massive amounts of money thrown at it by tech CEOs and venture capital. People do not realize how much hundreds of billions of dollars is. On a more concrete scale, people only see the fun little chat box when they open ChatGPT, and they do not see the millions of dollars worth of hardware needed to even run a single instance of ChatGPT. The unreliability of GenAI is much harder to hide completely, but it has been masked by some of the most aggressive marketing in history towards an audience that has already drunk the tech hype Kool-Aid. Who else would look at a tool that deletes their entire hard drive and still ever consider using it again?

The unreliability is not really solvable (after hundreds of billions of dollars of trying), but the expense can be reduced at the cost of making the model even less reliable. I expect the true “use cases” to be mainly spam, and perhaps students cheating on homework.

Now I’m even more skeptical of the programmers (and managers) who endorse LLMs.

The basilisk now eats its own tail.