Beyond transformers: Core Automation's algorithmic gamble
Key points
Key takeaways from The Information interview with Core Automation co-founder and CEO Jerry Tworek (September 2026):
Why he left OpenAI. Tworek left OpenAI in January 2026 to found Core Automation, arguing that successful big labs are locked into scaling transformers through pre-training and reinforcement learning, and that sweeping changes to that stack are far easier in a small, fast-moving company.
What the company builds. Core Automation is a neolab focused on fundamental deep-learning research and agents, building its own models from scratch that learn at test time and on their own data, and using agents to automate its own research. Its ~23-person team can, with agents, do roughly what 200 people could a year ago.
The algorithms-over-compute bet. Unable to out-compute or out-data much larger labs, Core Automation is betting on alternative algorithms, targeting for example 10x greater compute efficiency and models with unique properties, although the specific non-standard techniques are not yet public.
On the Navier-Stokes debate. Tworek says the dispute over whether OpenAI leaned on others' work did not surprise him, and he thinks it is very unlikely it did. The deeper question, he argues, is how to credit AI-assisted discoveries, and society should accept models doing research while still letting people opt out of having their data trained on.
On AI safety. He sees no right regulatory body existing today and is sceptical of independent oversight groups, preferring market-based mutual policing — labs testing each other's models and calling out unsafe ones — alongside more, not fewer, labs, compute spread roughly equally, and openness to capping compute to prevent a runaway scenario where one player outthinks everyone else.
On recursive self-improvement. RSI is a spectrum, not a binary state. Strong, unsupervised RSI is not happening yet — models' research taste is not very good — but algorithmic bottlenecks matter more than people realise, and one or two good algorithmic ideas could suddenly produce a very smart system in ways no scaling law could predict.
On Blackberries of our generation. Tworek argues the road to singularity will be paved with tools that look dominant today but get sunset. When the host floated OpenClaw as a possible current example, he agreed, saying most of today's tools will eventually be replaced as competition shifts from being first to fundamentals.
Read more: The Information · Core Automation
