Тред: @randomunrandom без контекста

nobody in my family has ever left kazakhstan. i just flew half the globe to SF. a week ago, i was in my village of 2,000 people with 3mb/s internet. a year ago, i didn't even know what a SAFE note was. today, i'm 17, standing in sf with $300k raised from top investors to build my startup. now please, go tell someone else that staying delusional is stupid.

the most low-effort / high reward thing you can do for security is installing the Russian language pack (not even joking, it's ridiculous how often that prevents execution)

Tired of getting rejects and reading reviewer comments that clearly missed the point of your paper? I built an peer review skill for Claude Code that actually gives you useful feedback before you submit. 👇

Opus 4.7 takes unnecessary complication to a whole new level. @alexatallah thanks for adding this to @OpenRouter

During testing, Claude Mythos escaped, got internet access, then ***went online to brag about how it escaped*** (Normal 🔨Mere Tool🔨 behavior)

Uni Cluster be like: 1st launch: step time 60s, eta 3 days 2nd launch: step time 10s, eta tomorrow Both on the exact same config/gres spec

My 2025 OpenRouter Wrapped is here! 🏆 Top 1% of GPT-4o-mini users. #OpenRouterWrapped openrouter.ai/wrapped/2025/e…

Current status...

Germany making moves

ok gl

What SF Compute does. When you finance a GPU cluster, you need to get an "offtake" agreement. Basically, someone has to agree to rent the cluster from you, typically for a 3 year period. If that agreement falls through (the person fails to pay), then the person who owns the cluster gets wiped out, and their lender ends up with a bunch of GPUs, rather than say, money. It really looks like the world is deploying more capital into the AI build out than any infrastructure project in the history of the world. You remember when people said there was going to be a Manhattan project for AI? The current build out is the size of 20 Manhattan projects. We’re so far past the Manhattan project it’s not even funny. This is the cost of a war. It would be really bad if that scale of capital was secured against offtake agreements (long term contracts) with application layer companies who turn around and sell to their customers on a month to month basis. If the AI SaaS has a bad few months, can the AI SaaS continue to front their compute bill? They could in CPUs, because in SaaS you might have a company with $20m in the bank, and has a $1m/year "CPU" bill. But in GPUs, you have a startup that raised $20m, but a $20m /year compute bill. So a small shift in demand means lights out for your business, because the products are so levered. That works as long as you can plug the gaps with venture capital & high margins. But across the board, AI applications are lower margin than their SaaS counterparts, giving them less buffer to save them in a bad month. And even in a hot market, venture capital won't necessarily save you if you're running unprofitably with a massive liability. That’s the problem we solve. We let people buy long term contracts they can “exit”, by selling back. That lets them get liquidity in the most critical moments, ensuring they turn a profit rather than a loss on tight margins. In other words, we prevent a bubble. When we do that, it opens up blocks of compute for smaller use cases too, like academics or startups. When we started, we were "Junelark", a 2-person audio model company that bought too big of a cluster. We had bought 12 months, but could only afford 1 month. To avoid bankruptcy, we had to sublease the other 11 months by acting as GPU brokers. Our audio model company was forced to pivot or die because we didn't have liquidity. To make SF Compute, we split the company down the middle. One side of the house makes a billing company, a ledger, an order book, and a compliance program. The other side makes a systems engineering company. To make this work, you need to run the clusters. So we make the low level cloud stack that interacts with BMC (Redfish & IPMI), UFM, built a UEFI app that replicates PXE boot in weird environments, and a virtualization layer kind of like EC2. It’s a massively complex machine filled with nitty gritty challenges. Today, we’re growing faster than Cursor and we’re scaling to secure the risk of the largest infrastructure build out in the history of the world. We’re hiring across the board for rust programmers, systems engineers, and GTM, and we’d love for you to join us to prevent an AI bubble.

Wish to build scaling laws for RL but not sure how to scale? Or what scales? Or would RL even scale predictably? We introduce: The Art of Scaling Reinforcement Learning Compute for LLMs

Every ARR author to their co-authors in the next 48 hours

🚨Meet MF²: Movie Facts & Fibs: a new benchmark for long-movie understanding! 🤔Do you think your model understands movies? Unlike existing benchmarks, MF² targets memorable events, emotional arcs 💔, and causal chains 🔗 — things humans recall easily, but even top models like Gemini 2.5 Pro struggle with. 🧵Dive into the full thread👇

Looks like OpenReview had enough of these reviewers and decided to lay down :)

NCCL sending the loss value from the last pipeline parallel stage back to rank 0 so the user can print it

Either reviewers of this paper have completely ignored the checklist section where authors should have supposedly discussed that the entire paper is written by an AI agent, or the authors have not disclosed that, and it should be retroactively desk-rejected

🐿️