Evi

@willdepue @FeltSteam @nickcammarata - Scaling personal sleep (auto RL per user) - Solving spatial reasoning and consistency (aka TikZ problem or “intelligent canvas”)

@FeltSteam @nickcammarata i think it’s clear there should be one to two more of these left, but it’s really hard to predict when. i think you’ll most likely get another one in the next three years for sure, and a good chance by 2028.

if openai had somehow kept strawberry secret would this lightcone belong to them instead of claude

@ColinGardiner Why do you need to talk at all?

As an early-stage investor, I don’t want to talk to your CFO co-founder about the raise. I want to hear from the CEO. If you’re an early-stage CEO, don’t outsource your fundraise.

@Dorialexander They latency is very high, so batch size is large.

Bonjour BFM. Vous avez personne pour lire les model reports ? Réponse en p. 9.

@LyalinDotCom Isn’t it obvious that Spark is much more compute dense while your M5 Mac is bandwidth winner, so you should really connect them with USB-3 and do prefill on Spark then transfer KV to M5 and do inference there. Ask Codex to do this (from scratch obviously).

@GeorgeJeffersn lol, you should have made clear in your post that your app specializes on growth, so you’re obviously not sharing the full story.

At YC we got the advice of posting on LinkedIn At first ngl I thought it was bullshit because I’ve never liked or engaged with the platform before But holy shit, Ive tried to post x5 a week and its mostly been slop like this, but it’s pure gold for driving inbound & sales

@burkov This was pre CoT.

Always remember that when an LLM prints the beginning of a text, it has no idea what the end will be. Therefore, when it says "The answer is yes, and this is why:" the text after "why" would most likely be a very elaborate lie combined with gaslighting in case "yes" was the wrong answer.

@zhaisf It is obviously not cubic. Matrix multiplication is obviously quadratic! Well, everyone knows (see Dwarkesh lecture is you decided to forget this) that data transfer is the main bottleneck. For matmul that’s quadratic, not cubic.

Maybe it’s time to address the cubic complexity of matrix multiplication.

@emollick that's a bug which has to be fixed :)

You can use AI to help with writing, but you need to actually do some writing to be helped with. "Write me a post about [topic], make it really good and interesting" is not going to cut it.

As more people come to recognize the tells of AI, which mostly happens as you start to work with AI a lot, the scales are going to fall from their eyes and they are going to realize what some of us already see: how much of this site (and blog posts, articles, papers) are AI now.

@sumjitg Cope for them missing out on coding capabilities.

Just saw Demis Hassabis push back hard on all the recent Erdős problem hype. He basically said: sure, today’s AIs are knocking out some of these tough math problems, but that’s nowhere near real AGI. It doesn’t come close to the kind of raw creative invention someone like

@MTSlive Lame! Folks you really need some better sourcing of “the situation”:)

SITUATION DETECTED: Google DeepMind’s AI agent autonomously solved 9 of 353 open Erdos problems in mathematics, at a cost of a few hundred dollars per problem.

@MWill2025 @sierracatalina Sam has nothing to do with what models do. He obviously gets funds and compute sorted. Genius who created recursive self improvement for GPT is someone else.

@sierracatalina Correct, and I'm not a tech person. However, I do look at the leadership and hope they are ethical in their developments.

so… do we need to talk about how claude & gpt are FUNDAMENTALLY different!? because I keep seeing yall compare them & I am just unsure if no one understands the math or what.

@Nick_Davidov @nikitabier These are bots obviously. There is no downside for highly valuable employe s to stay in US while waiting for EB-1 and not have to go to likely hostile “home country”.

It has been this way for some time, but @nikitabier, the algo now fully embraces the vice and only promotes posts where shit hits the fan. I don't want to be a voice on capitalism or immigration and be hated, I want my useful posts with things that might be interesting for the tech community to get some attention too. I'm now closing my account to followers only and using "following" page for the feed. Because FYP is just ragebaiting and hate

@ProbyShandilya Is that because they resell Claude and GPT at below API cost?

Cognition hit $445M run rate in their first 18 months of business. I don't think that's talked about enough.

@tunguz Totally! That whole LeCunn thing is a psyop

word model >>> world model

@michaelzixizhou @ycombinator @zfellows Ask ChatGPT to read the actual NDA and check legality. It is very useful to know that fresh grads (level 3 at Google, Meta, OpenAI etc) get 2M+/year TC offers (8-12M over 4y). That’s more than Ilya’s comp was 10 years ago at OpenAI (when it was non profit and data was public).

@geteviapp @ycombinator @zfellows NDA unfortunately i can’t even namedrop

I got into @ycombinator S26! The past 6 months have felt like 2 years: - Turned down 7-figure new grad offers at top companies - Bootstrapped until my bank account hit zero - Pivoted 5 times, and did door-to-door sales for a couple of months - Joined @zfellows and flew out to SF last month - Cancelled my return flight to Toronto - Had wonderful neighbours I'd just met crowd-fund my apartment deposit - Got the call from @aroraharshita33 6 hours after my YC interview ...all without telling my mom. Codag is the crystallization of everything I picked up across infrastructure, systems, and platform engineering. Our thesis is that within two years, agents will be the primary consumer of logs, and soon the primary producer of them. Today's observability stack and dashboards were built for humans to look at. Agents are a different reader entirely, and the infrastructure for understanding systems at scale has to be rebuilt for them. That is what we're building. We start with log compression: turning millions of lines into a compact, cited capsule an agent can actually read, so it stops drowning and starts debugging. I cannot believe I get to work on this full time - I wake up every day looking forward to building more. To everyone who made it real: @carterkev, @cory, @aroraharshita33, @amiklas, @nemild for taking the bet, the neighbours who backed a stranger with a laptop, and the countless friends that support me. Thank you. We just published an open source CLI and algorithm for log compression, fully free at codag.ai. If your agent debugs production, I'd love for you to try it, or just to talk. Got a lot more coming soon, and most of it will be open source. Now I should probably call my mom 😰

@dharmesh Cope

The harness matters more than the model. Models have gotten really good. Great reasoning, large context windows, better instruction following. But, what makes *use* of those capabilities is actually the harness. It's what provides tools, memory, skills and context to the model. ChatGPT is a harness. Claude Cowork is a harness. Without the harness, the model is just an engine with no car. You don't get anywhere.

@johnennis It is usually people who use Claude think (wrongly) that humans are somehow the key and model intelligence is not strictly superior to human intelligence every way. Claude is trained in sloppy way so its frontier is much more jagged.

There's really no substitute for actually doing things if you want to understand them Among other things this weekend, I've been using a number of AI tools to work on Erdos problem #3 on arithmetic progressions, it's been a worthwhile experience, and I've even proved a few new results (although still far from solving the main problem - it is very hard) I also learned a lot about additive combinatorics, a subfield that was born after I moved on from math to computational neuroscience, so that's been fun too However, my takeaway from all of this is that, while AI is going to be an awesome tool for professional mathematicians to work with, the idea that you are just going to turn the keys and the AI will "solve math" is horribly wrong It's just not going to go like that, and not because the models aren't smart enough The models are plenty smart and incredibly useful But human and machine intelligence are different, and I think you have to be in trenches to get a feeling for how these two types of intelligence are just not the same (And happy to share my note with anyone who wants to read it, just comment or DM, I put an appendix in there on my workflow for those who are interested)

@LLMJunky @mark_k Continue learning will be fully solved this year via: RL for very common tasks DB (files + embeddings/soft tokens) for knowledge Obviously it will start with personal DB, RL will be done later using data from DB.

@geteviapp @mark_k That is not continual learning.

So what happened to the Continual Learning breakthrough we were promised in 2026? AI labs have been fairly silent about it lately... Is Continual Learning still going to become a thing or will it remain a pipe dream for AI?

@_arohan_ Which model from Antropic is good and Jensen said they used moderate amount even for sloppy Mythos.

This is very true. This is why the best models are from Anthropic earlier this year as they had most compute compared to everyone else.