a16zCan Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert
a16z
Model progress drivers 0:00
Model progress comes down to compute, data, and talent. The big closed labs have had the most of all three, and they have been better at organizing them. The org chart matters because building models is hard, and some companies have had to reset from the ground up before they started moving again.
Open models and catching up 4:00
OpenAI and Anthropic are described as the clearest frontier labs because they took the biggest bets earliest. The Chinese labs are technically strong and have comparable talent, but they are more compute starved. Open releases are a practical way for them to get into the conversation, while distillation and other fast-following methods can narrow the gap by a few months.
Distillation and access 13:32
Reasoning traces from closed models are easy to strip out through APIs, so long as the companies keep offering them. That makes distillation hard to stop. If the models are kept inside products and enterprise deals, it gets much harder. The speaker is sympathetic to labs protecting themselves, but less sympathetic to government action over what looks like a terms-of-service breach.
Open and closed paths 19:30
Open and closed models move on different clocks. Closed labs have the capital, the inference systems, and the first shot at high-value markets such as coding. Open model builders split into builders, inference companies, and post-training shops, so feedback and revenue move more slowly. Still, open models are already good enough for fine-tuning in many company settings, where private data and internal evaluation matter more than raw frontier scale.
Open data quality 27:30
The discussion stays on how model capability depends on more than scale. Bought environments and synthetic data can help, but the quality is uneven. Labs want to reduce reward hacking, while data vendors often focus more on building worlds and selling volume than on behavior quality. The result is a rough supply chain that still needs better filtering and tighter work between labs and data makers.
Open environments work 31:31
The speaker says the nonprofit will mostly make open-source environments for high-value tasks. The example is T-Max for terminal work, where frontier models help generate problems, then rollouts and bug checks remove bad cases. These environments are expensive to build well, but they matter because open releases let researchers study model behavior over time and give PhD students real infrastructure to work with.
Why openness matters 33:02
Open work is described as slow to reward but important over the long run. The speaker argues that industry captures the value later, while the open community supplies the basic tools and standards. He wants a place where more people can contribute, ask better questions, and share code and findings without needing to sell directly to labs. He also says AI agents make open work noisier, since people now submit more false or careless issues, even though the problem is still manageable.
RSI and open stacks 38:00
The talk turns to RSI, or recursive self-improvement, and the fear that it could let frontier labs pull ahead. The speaker says the best way to make it legible is to show the full stack: the tools used, the training setup, the tasks, the compute, the trajectories, and what worked or failed. He wants an open record of how much automation is really happening so people can judge capability more clearly, including how it compares with harder work like alignment.
Policy value of examples 41:01
Showing a model fail in a concrete case can help policy makers understand what is happening. The point is to demystify the work, not to claim the models are already solving every office task. The bugs and catchups will shrink over time, but some rough edges will stay for a while.
Open research and constraints 42:31
Open AI groups can document their tasks, build useful evaluation data, and release work with rare details that others can use later. There is also room for different open efforts to take different paths. On pre-training, the tools are fairly open and a solid base model does not always need a billion dollars, but the hard part is still getting enough chips and choosing the right recipe.
Post-training and agents 49:00
Post-training is harder than pre-training because the cycle is messier and more specialized. The new agent setups in closed labs are also different enough to create a mystique, since the models are tuned for specific harnesses and coordination systems. Even so, the speaker expects open model teams to catch up once the shape of those systems becomes clearer and more people try to use them.
Limits and pace 54:31
He says the big labs are spending at an exponential rate, and that this sets a clear limit on how far they can pull away. He does not expect a fast takeoff where the models build themselves. The bottleneck is still human work, even as compute keeps climbing for smaller gains.
Prepare for present risks 57:01
He is more worried about weak preparation for known risks, especially public cyber security, than about a sudden new jump in capability. He wants more eyes and more science so people feel informed. Not fearing fast takeoff does not mean ignoring the risks that are already here.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
