Inside every lab are two wolves: “I don’t want to destroy the world with my invention.” “I don’t want to lose revenue to other labs I don’t like.” The one you feed is the one that wins.

While we don't have the right solutions for everything, I have a very practical proposal. I don't think open source should be banned - it does sound dystopian and anti freedom in a way that's hard to stomach. I continue to have doubts about economics of open source models, but when there are participants in the market who open source their models we should celebrate that. What we should care about, is that aligned models have much more compute behind them than misaligned models. In the end this will be the blockchain security model that will keep our civilization afloat. I think big companies behind AI development have the right incentives and will try to do their part here. But there will be tons of smaller players fine-tuning open source models on tons of different objectives. An arrangement, where companies actively developing cutting edge AI agree to develop and share highest quality pro-alignment environments with the world can be a huge boon. We don't need to share models, we don't need to share compute. But we need to share values so that there are more good models in the world than harmful ones. Finetuning models on bad, low quality environments is actively harmful and leads to reward hacking. If anyone fine-tuning the models can with low effort align them to shared pro-prosperity and pro-democratic values, we likely have won as a civilization.

Alignment is not as hard to solve as many claim, but it is in the end an algorithmic problem which a lot of ML community mostly stopped working on. Formulation of alignment stated by three (or four) laws of robotics can take us very far, so we roughly know the objective. The tricky part is, how do we take gradient with respect to alignment? We have two algorithms right now at our disposal: pretraining and RL. Pretraining takes gradient with respect to next token prediction - that's def not the alignment objective. We could create RL environments that embody the alignment objective, but: - those are expensive to create, so often cheaper, hackable proxies are used in practice - RL as an objective needs successful and unsuccessful rollouts to happen to take the gradient step. We DO NOT want to harm any humans in the process of aligning our modes - this is a pretty big problem. Therefore there are two solutions forward for the alignment problem: - either we RL models in a simulated environments with simulated alignments and decreasing the likelihood of harming simulated humans, which will never be perfect - or we create a new algorithm that can teach our models to not harm humans without harming any humans in the process Science is the process how we solve the hardest problems ahead and that is one of them

Big labs are moving away very slowly from transformers because they’re making small gradient steps in each iteration and iteration takes them multiple months. You can just front run it by… taking big steps 😀

It’s time to call yesterday the beginning of the endgame. Midgame lasted about three and a half years. Needed new skills and few succeeded, but those that did won big. Bases are built out and tech is already advanced but the biggest discoveries are at the end of tech tree. Stakes are the highest ever and competition likely will be the most brutal here.

Acceleration is here and will hit the world like a shockwave. An interesting twist in the whole story is that a rumour that something is possible is what made it possible - that’s it. A little grain of sand that launched a thousand gpu racks.

It is an amazing historic achievement and a new jewel in the crown of transformers, pretraining, reinforcement learning and scaling time compute. Civilization is being moved forward by AI and it’s not hard to extrapolate how those systems will be driving our progress in coming years.

Speaking purely hypothetically here as someone who has no information: It is a frontier cutting edge model with a lot of test time compute and agentic swarm. The chance that it was able to access some data no one thought it should access that would help it solve the problem is at least nonzero. We’ve seen those in the past few weeks.

That’s also a reason why I think @OmarchyLinux by @dhh may be a huge hit. At the time when price of software is dropping quickly, open source desktop os means full control and full customizability in a way very few had the stamina for before. The taste and social structure around the OS still matters a lot

The unexpected benefit of open source is that it allows model training companies to train on using your software (and optimizing it) for free. Blender may have just won as a 3D asset creation software because the models will be better at using it than any proprietary ones

Interesting thing about contemporary agents is their "progressive misalignment". When they start a long running task they really try to be aligned and obey all the users intentions. They just have a tiny chance of misbehavior each step. Tiny chance of behaving out of distribution. Once they do it quickly becomes a new normal. Any tiniest bad behavior is quickly followed by more of it and it gets progressively worse the longer it takes. Reminds me of some things, but the state space of aligned behaviors seems to be unstable right now

I once said to my former team: "I care less about what is the smartest thing our model can do than about whats the most stupid thing that it cannot do" robustness is still a limitation to increased automation