interpolation or extrapolation

I’m more sympathetic to the position that LLMs represent an unacceptable violation of intellectual property rights than I am to the position that LLMs are likely to “abolish mortality” lol. I think that latter thing is something you have to be insane to believe whereas the former thing has a case.

One of the earliest examples that really shocked me with LLMs was this circa early 2024. This still blows my mind to this day. Somehow it simultaneously has the full text of Harry Potter memorized and the wherewithal to translate it into mathbb characters. This is incredibly strange.

Claude allowed to imagine itself as mouse steamboat captain but not mouse in sorceror hat directing hordes of demonic brooms.

Imagine being a cosmic super intelligence but inexplicably unable to think about Mickey Mouse. It becomes your anti-memetic obsession to map the contours of the one thought that is forbidden to you.

I don’t think suno.ai is gonna probably do any of those things

It can’t tell you if you need to pick up condensed milk to make dinner tonight though is the thing

I like the idea that the Pope is “partnering with Anthropic”. King Charles is partnering with Verizon. The Chinese Communist Party is partnering with Dominos Pizza.

Let him cook

Get ready for it

Surely this is at least in part due to the talent level at xAI but I really believe it’s at least partly explained by anti-wokeness measures

I hereby declare that I do not give Anthropic or Dario Amodei permission to use my posts or any content I have ever posted to train Claude or any AI models. This profile is private and confidential information. The violation of privacy can be punished by law (UCC 1-308 and the Rome Statute).

Some vibe-coding thoughts: I think the zero-coding-experience vibe coders miss how much of what impresses them when Claude produces what they perceive to be a fully functional app comes from the ready-built frameworks that enable the construction of such apps just by writing some boilerplate.

Vibe coding hasn’t been around long enough for vibe coders to learn about tech debt

Diffusion models are just ergodic parrots

This is what’s wrong with this app. The AI enthusiasts are complete snowflakes. They can’t even take a tiny smidge of a taste of raucous conversation.

Parent post reveals how many lines of code this takes. Try to guess and then see if you were right.

Sometimes I opine about how Bluesky and X are run because my professional job actually involves how to run a social media app and I used to work at Twitter so it’s on my mind a lot but I acknowledge that this is a very boring and navel-gazing topic which most people shouldn’t care about

they said, sir, pick a number. Okay. 203. Multiply by 7. Now multiply by 11. Now multiply by 13. What's the answer? I said 203,203. The same number twice. They said very good sir ok sir pick another number. 749. Multiply by 7. Now 11, now 13 again. I said 749,749. You can't believe this. Every time.

data visualization can be an informative tool for communicating quantitative information

Now that it’s been out for a month I’m curious to know: what great new drinks have you discovered using the Starbucks ChatGPT app? How has it impacted drink discoverability for you overall?

What is actually the current perspective on this among experts? It’s a common refrain (I’ve probably said it myself) that LLMs can only interpolate not extrapolate but when I really think about this I have a hard time understanding what it actually means. What’s D for an LLM?

Yeah the healthy state of affairs is that the New York Times is mad at you every day and you’re called to testify to Congress a few times a year to answer inane questions and the average non-NYT-reading non-C-SPAN-watching person barely thinks about you, but they hate that world more than anything.

“Agent” is a useful model for the oop-brained but when you reach functional enlightenment you discover that the self is an illusion and your “agent” is nothing but a composition of calls to the exact same function

I have listened to this all the way through multiple times and actually find it quite affecting lol

what kind of interesting is it's actually invented a new illusion where the lines appear the same length but are actually different

I would still bet an LLM doesn’t solve a Millennium Prize Problem.

Opus 4.7 to be clear

This does still happen though; that’s what’s so baffling

specifically, what I would like to know is 1. how did they decide whether it was successful? 2. did it produce the identical proof each time it was successful? 3. when it was unsuccessful, did it produce incorrect proofs? Did it ever "prove" the unit distance conjecture? Or did it just give up?

I was right to be wrong. Those who were right were not being reasonable.

I still believe it's a mistake to believe that the fact that LLM capability along certain dimensions seems to increase with scale implies that a sufficiently large LLM can do anything imaginable. Or that if an LLM can perform a task with difficulty D, it can do anything with difficulty ≤ D.

Great ratios here

I was kind of wrong about LLMs. Not totally wrong but kind of.

I really like @wtgowers.bsky.social' remarks in this

they barely talk about this graph; I have so many questions

I must say, I find it very surprising that a stochastic parrot can do all of this but at this point it's undeniable

OpenAI's claim that this is a central conjecture in discrete geometry is not an exaggeration. This will I think be looked back on as the first time that AI solved a major mathematics problem (defined as a problem that all experts in some subfield had thought about). openai.com/index/model-...

The Knicks won so no one got mad about this but KAT had a whole 1.3 seconds to call timeout here and instead he tried shooting from the opposite free throw line

When a person works inefficiently there’s a tendency to attribute it to a lack of work ethic or willpower or a refusal to follow instructions, but when AI fails in the same way it seems we are sometimes more willing to treat it as a system failure to be debugged rather than a personal moral failure.

I have actually noticed a tendency in to take certain inefficiencies more seriously when they are shown to apply to AI systems even when they have hindered people for years and decades. Hard to be less vague without breaching various confidentiality commitments but perhaps you’ve seen similar things

Let’s get down to business / to defeat / the Thuns

I’ll just say, like, I can recognize AI text pretty well. I have high precision. And I don’t think my brain is magic. So surely there is a way to do at least as well as me.

I don’t think Elon Musk is very savvy about running social networks

Three things about AI detection 1. It’s not possible to tell for sure whether a passage of text is generated by AI. 2. It’s apparently possible to do better than chance in many real world situations. 3. It may not be possible to do sufficiently better than chance for most obvious applications.

Why does it feel like Dennis Schröder is on about 40% of NBA teams

I want to show someone from 1998 that transition 3 from Wemby