ofofofofof

AI non-insider and part-time math hobbyist for 30 days per three months (I am no bot, I just type fast and scroll on Twitter if I have nothing to do)

[ofofofofof @ofofofofofdi]

@TokenGremlin Hmm. Right on to GPT-6.5?

[ofofofofof @ofofofofofdi]

Reminder that internal models at Anthropic are cracked. (Anthropic delivering their shares, sharing OSS scanner, usage being incredible, etc.)

OpenAI‘s models are cracked too, but I would say mostly in math. I don’t see any other evidence.

[ofofofofof @ofofofofofdi]

@TokenGremlin Agree

[ofofofofof @ofofofofofdi]

@notjazii @imjustnewatai I’m thinking more about it, and I think 4 Carbon is going to be the one that’s releasing in November. Deepmind already released Argon recently. Fable 5.5 and 6.1 Terra seem like they’re going to release in October, with Fable 5.5 being first.

[ofofofofof @ofofofofofdi]

Daily Dose of IntMul κ:

PR #4100 gets a κ of 0.00173097911567079247276057. Quite close to 2^-9 but who knows if we can achieve it with this level of slowdown haha

PR #4100: https://t.co/pJcfqn6yUC

Repo: https://t.co/LysE7o22qa

[ofofofofof @ofofofofofdi]

@ChrisGPT And yet, I still feel like I am not “impressed”? I know that it had casually just transformed all of math, but it is still mundane. Thanks for reading. (5)

Sorry for the rant.

[ofofofofof @ofofofofofdi]

@ChrisGPT *will be mundane even in the future.

Anyway.

I am probably one of the first to have seen AI to its huge potential because I am a hobbyist mathematician. I saw that thing solve Erdős problems. Impressive. Then it solves the Jacobian conjecture. Okay. Then it solves Navier-Stokes

[ofofofofof @ofofofofofdi]

@ChrisGPT I think, often, humans want excitement by writing stories/movies that give a very tense feeling, but it often feels a lot more gradual, and is usually more mundane, even for an AI singularity. I personally feel like everything’s mundane, and everything is going to be mundane. (3)

[ofofofofof @ofofofofofdi]

@ChrisGPT I also have to admit I am not completely pro-AI. I am pro-AI in things like the sciences, but I am more indecisive on areas like human art. I’ll see it much later when views on AI deepen and mature. (2/2)

[ofofofofof @ofofofofofdi]

@ChrisGPT For how incredible the concept of an AI singularity is, it feels like one big clicker/incremental game to me. Improve the thing, then get upgrades that makes the thing improve itself, then automate upgrades, and you finished the game finally. (1/)

[ofofofofof @ofofofofofdi]

@imjustnewatai Most optimistic interpretation is that model releases start in October, then are spread throughout November. And since I feel like being optimistic, I trust the leakers (particularly lyraxana) to hold their word that it’s this week.

Otherwise: patience

[ofofofofof @ofofofofofdi]

@TokenGremlin (2/) 6.1 Astra might be able to match Fable 5.5 though. Limited usage doesn’t mean limited intelligence. But I think OpenAI’s going to struggle quite a bit.

[ofofofofof @ofofofofofdi]

@TokenGremlin (1/) have a feeling that 6.1 Astra is going to be quite limited in usage. From what I hear Anthropic is very comfortable in its usage. It is diverting an odd amount of resources to the public, so whatever happened internally it got them quite far.

[ofofofofof @ofofofofofdi]

@itsvishaltwt 37 is the answer to all questions

[ofofofofof @ofofofofofdi]

@TokenGremlin Very very very cracked internal models.

[ofofofofof @ofofofofofdi]

Hell now that i think about it, why not a new scaling axis? Measured by how much faster it is than backpropagation. Let’s multiply the speed by 100 mate

[ofofofofof @ofofofofofdi]

Imagine an AI model proceeds to discover a new algorithm in place of backprop just to improve itself

[ofofofofof @ofofofofofdi]

Imagine an AI model proceeds to discover a new algorithm in place of backprop just to improve itself

[Samip @industriaalist]

I’m now confident that deep learning is going to look *unrecognizable* within the next 6 months. Replacing backprop is just the tip of the iceberg.

[ofofofofof @ofofofofofdi]

@platypii @jdb1729 Nice

[Kenny Daniel @platypii]

New lowest bound proved in lean: κ = 2^-19.2693 thanks to huge improvement by @jdb1729

[ofofofofof @ofofofofofdi]

@ChrisGPT Those pink and purple curves are surprisingly beautiful