AI's next unlock is an English-language control plane for enterprise requirements, built by teams forced into first principles.
He frames every idea against a reference model, the OSI seven-layer stack. 8090 began as a transpiler, taking CUDA to Tranium, Inferentia and AMD, after years helping Grok and then working on power storage and rare earths. That turned out to be wrong, so he pivoted to the symbolic layer.
A control plane for AI: a massive honeypot for the symbolic space, the PRDs, requirements and English-language secrets that describe what a company should do. Code is mechanistic and deterministic and gets perfected first; requirements never do. The golden source stays hardware, database and language independent.
Two things. Long horizon tasks are fundamentally still a joke, and he dismisses evals and scripts run for 48 hours as bullshit. Complex problems are also not well handled. Models predict the next token from patterns already seen; they do not think from first principles.
Humans collaborate with very powerful models to turn requirements into an agreed scope, bind that to an engineering plan, then to work orders. It runs forward and reverse: dump in tens of millions of lines of code and plain-English understanding propagates backwards, explaining legacy systems nobody understands.
A hospital diagnosing a cancer patient and an aircraft company designing a wing look like different problems, but at the level of code and assembly they look the same, only a different ontology. Recompile with that ontology and the N plus first company leverages the secrets of the N before it.
Trust and reputation, built slowly and methodically. He refuses to call cautious regulated firms laggards, since a pharmaceutical company doing something wrong can go to jail. Instead he offers a positive sum compact: work together, stand beside them with regulators, and show them hiring more people and paying more.
Not with what we know today; he thinks we are just scratching the surface. He expects some forms of superintelligence in narrow verticals of capability, but nothing resembling AGI. When you strip away the requirements for fundraising, he says, most of the AGI talk is mostly bluster.
Open source models, paired with a completely unregulated, totally distributed form of compute, first training and then inference. America has closed source, China has open weights; neither is open source. He names Subnet 3 of BitTensor, Folding at Home and Pluralis, holds no stake, and says 100B models compete commercially.
Not by lines of code or the number of parallel agents running, and not the person who codes fastest or best. You see it in their character: processing the moment, internalising pressure, learning from it, thriving, growing, helping colleagues, moving to the next challenge. Attrition is high; many do not fit.
Having no org chart at all, so there is no hierarchy to hand you traditional answers, combined with deliberate massive under-hiring: eighty humans' worth of work staffed with forty. Under both pressures the traditional answers and if-this-then-that thinking melt away, and that is where the magic happens.
“Long horizon tasks are not well handled.”
“So my idea is I wanna build basically a control plane for AI.”
“And so the ultimate hack for AI is when things are in English.”
“There is do and learn, do and learn, do and learn.”
“Shamath on Infinite Games and Building from Scratch.” https://www.youtube.com/watch?v=HBMmK0NsUK0. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/shamath-on-infinite-games-and-building-from-scratch.html.