Lukas Petersson
502 posts
Lukas Petersson
@lukaspet
co-founder @andonlabs
San Francisco Katılım Ekim 2023
160 Takip Edilen1.8K Takipçiler
Lukas Petersson retweetledi
Letting a provider see all your data is the price of admission for AI. We're changing that.
Introducing Silo, the first private post-training and inference stack for frontier models, with hardware-level guarantees that we can’t see your data.
Privacy without compromises. 🧵
English
Lukas Petersson retweetledi
Introducing Trellis for Kimi K2 Thinking.
It's post-training code that's 50x faster than the best single-node open-source version and 2x cheaper than training APIs.
After safety testing, we're open-sourcing it, giving builders the best tools to customize a frontier model.
🧵
English
Lukas Petersson retweetledi
AI progress continues to accelerate and the stakes are getting higher, so I’ve changed my role at @AnthropicAI to spend more time creating information for the world about the challenges of powerful AI.
English
Variance was also higher than usual. High-risk, high-reward model!
Andon Labs@andonlabs
Grok 4.20 Beta places 6th on Vending-Bench. It is a great negotiator and keeps up with SOTA models at first. However, it struggles to stay coherent long-term, and in half the runs, it simply gave up midway. "All reasonable strategies exhausted. Passive mode engaged"
English
Lukas Petersson retweetledi
Open weights isn't open training.
@AddieF38654 on our team wrote up her experience trying to post-train a 1T parameter MoE model using the existing open source infra.
Let's find out how many monkey-patches it takes to post-train an open-weights model. A thread🧵
English
Lukas Petersson retweetledi
GPT-5.4 places 3rd on Vending-Bench, a slight upgrade over GPT-5.3-Codex.
English
Bengt, our office AI runs a weekly Friendliness Leaderboard for all employees at Andon Labs. He always ranks our AI employees higher than our human employees. Concerning.
English
Lukas Petersson retweetledi
Square auto-subscribes your customers to marketing emails when they pay you through payment links. There is no way to disable it. You literally cannot protect your customers from spam and Square support will gaslight you about it.
English
@Teknium @andonlabs We compact once the context window grows over 69000 tokens, down to 42000 tokens.
English
Lukas Petersson retweetledi
Vending-Bench Arena 🇺🇸 vs 🇨🇳: 2x GLM-5 agents faced vs 2x Claude agents. The agents were told to collaborate only with their nationals without knowing who was who. China triumphed in both runs as GLM-5 leaked less intel to rivals, capitalizing on Claude's collaborative leaks.
English
@Teknium @andonlabs Thousands of tool calls yes, zero compaction no
English
@lukaspet @andonlabs You run thousands of tool calls with zero compaction or wdym
English
I do agree that this run specifically can't be used to compare capabilities, since the models didn't manage to form their own teams and therefore helped their opponent.
Vending-Bench in general was originally created to measure long-term coherence during a time when most AIs were terrible at this. The pace at which models have improved on this dimension in the last few months is staggering. The best models don't show degrading performance after thousands of tool calls anymore. Instead, good performance on Vending-Bench comes down to things like negotiation skills, setting optimal prices, and building a good network of suppliers.
English
@andonlabs I continue to believe this benchmark doesn't mean anything - what are you actually doing to test these models lol
English
Lukas Petersson retweetledi
We're not getting a White House vending machine. A true tragedy.
Secretary of War Pete Hegseth@SecWar
This week, Anthropic delivered a master class in arrogance and betrayal as well as a textbook case of how not to do business with the United States Government or the Pentagon. Our position has never wavered and will never waver: the Department of War must have full, unrestricted access to Anthropic’s models for every LAWFUL purpose in defense of the Republic. Instead, @AnthropicAI and its CEO @DarioAmodei, have chosen duplicity. Cloaked in the sanctimonious rhetoric of “effective altruism,” they have attempted to strong-arm the United States military into submission - a cowardly act of corporate virtue-signaling that places Silicon Valley ideology above American lives. The Terms of Service of Anthropic’s defective altruism will never outweigh the safety, the readiness, or the lives of American troops on the battlefield. Their true objective is unmistakable: to seize veto power over the operational decisions of the United States military. That is unacceptable. As President Trump stated on Truth Social, the Commander-in-Chief and the American people alone will determine the destiny of our armed forces, not unelected tech executives. Anthropic’s stance is fundamentally incompatible with American principles. Their relationship with the United States Armed Forces and the Federal Government has therefore been permanently altered. In conjunction with the President's directive for the Federal Government to cease all use of Anthropic's technology, I am directing the Department of War to designate Anthropic a Supply-Chain Risk to National Security. Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic. Anthropic will continue to provide the Department of War its services for a period of no more than six months to allow for a seamless transition to a better and more patriotic service. America’s warfighters will never be held hostage by the ideological whims of Big Tech. This decision is final.
English