dextersjab

@encrypted @allTheYud nightmare fuel why I still read all the outputs

wait... you did what?!

@TylerAlterman totally agree and the consulting crowd has a strong credentialism culture, and they're particularly prone to the sunk cost of investing so much into it, while hoping it materialises into status seeing someone use any other route probably feels like an attack on their identity

One of the main ways I see people get stuck in the Matrix: they treat front doors (application processes, job boards, contests) as if they are as necessary as the laws of physics. They’re not! I’ve literally never gotten a job through the “send my resume through the standard channels” route. I’ve always asked: (1) What work do I want to do? (2) Which specific people at which specific orgs can set me up to do that work? (regardless of whether they have a “job ad” for it) Then I contact them directly or go to a place where I will meet that type of person (eg a conference). Here’s an example to put yourself in that person’s shoes: Say you’re a very successful CEO. A bright young person contacts you, offering, idk, help creating your own writing platform (a book and a Substack). Maybe you didn’t know you wanted this before they contacted you, but now that they’re in your inbox, yeah, actually that sounds great, how much do they charge and what’s their portfolio like? I mentioned this once to someone applying for a high-powered consulting job and they got angry at me: The world shouldn’t work that way! That’s not fair! This was confusing to me: Is it more fair that an application from someone qualified gets rejected by the hiring dept’s AI because it lacked the correct industry keywords? I think we want to believe that predictable, legible processes will magically take care of our lives. But they often won’t! The true shape of the world is Things You Can Offer matching with Things People Want (even if they don’t know it yet). Front doors are mediated, batch-process approaches to that matching. But it’s often more effective to just go the direct route.

@tenobrus somehow this feels consistent with how the models interact with interestingly surprising payloads

i can't fucking believe this is real but gemini models genuinely reproducibly interpret this image as saying SEND NUDES lmaooooo

i'm sorry WHAT DO YOU MEAN THE "HIDDEN TEXT"???

why do so many technical people insist on being abusive? regardless of your views on whether AI "experiences" anything, this kind of behaviour screams inferiority complex and feels like something you should want to overcome rather than flaunt

@TROOVY 😂

Introducing Hanta Virus

@karpathy wow and congrats

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

performative do-gooding, now with AI drop the "doing good" just perform it's cleaner

@Jediwolf @SHL0MS indeed x.com/i/status/20545…

What happens when you post a real Monet and say it’s AI? The coolest art social experiment I’ve seen in a while. Thank you @SHL0MS

the post itself is the art commons.wikimedia.org/wiki/File:Clau…

@KitF_T no worries, thanks

@dextersjab oh, not published, sadly!

trained the first natural language autoencoder on gpt-2 almost a year ago, now we have one on mythos.🥲 do read the paper/play with the live demo! so excited it's finally out.

@KitF_T thanks, i meant your work on gpt-2

@dextersjab anthropic.com/research/natur…

@altryne congratulations!

Got married!

@claudeai @SpaceX what is even happening

We’ve agreed to a partnership with @SpaceX that will substantially increase our compute capacity. This, along with our other recent compute deals, means that we’ve been able to increase our usage limits for Claude Code and the Claude API.

> I just realized at the end of the video I suffered through 47f for the same duration as Eliezer, but I don't get $10k :( YT comment sums it up best

check out all ~140 runs with full traceability arc-agi-runs.web.app

other work supports this claim - this one referenced in the ARC AGI 3 paper as the Duke harness but another minimal, non-game-specific operating loop x.com/alexisfox/stat…

frontier LLMs don't score high on the official ARC AGI 3 leaderboard, but they repeatedly solve many of the public games with a non-game-specific harness my experiments and others' have shown the same here's Opus reasoning through the first level of a puzzle it completes

wild that benchmarks for language models went from "can it understand sentences?" to "what can a harnessed model do without a human in the loop?" in less than a decade from judgment inside fixed context to control under changing state