@wolfejosh @satyanadella I love seeing Satya reference Hayek
@AravSrinivas Well said
One of the key architectural questions of the 21st century in business will be how you maximize your corporate IP in the form of decisions, insights, workflow patterns, and best practices in a world where so much intelligence is packed into AI models.
One might think these questions could just get bitter lessoned out of existence, but in reality they become even more germane as intelligence becomes more powerful. In a world where any firm also has access to frontier intelligence, understanding how you leverage it uniquely becomes a critical question.
That’s why so much value is left to be created between the enterprise and the underlying AI itself. Having evals for your workflows, ensuring that you can route models from different tiers of intelligence, capturing traces in a way that improve your own workflows, and making sure the value of your information compounds as AI gets better all become critical considerations.
Which is also why there’s so much opportunity right now in the applied AI layer. The companies that help figure this out for other enterprises will be in the best position to win the next enterprise workloads.
I find it ironic that the status quo is to then turn around and impose restrictive terms on distillation, and to reserve the right to learn from customer usage and interaction data. If learning flows in only one direction, economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself. Therefore, it's imperative that we distribute the learning infrastructure to every firm so that they can control their own learning loop."
Well said.
There's a major gap in this otherwise compelling vision.
Lots of interesting signals recently: token budget caps, intelligence sovereignty, increasing competition (Meta/Grok/GLM) that drives the commoditization of intelligence.
All seem to point toward a more benevolent future equilibrium where every organization owns its human-AI learning loop through which knowledge and value accrue locally. @satyanadella, @thinkymachines, @dwarkesh_sp all laid out good arguments for it.
No doubt that is a more desirable future than the oligopolistic path we are on, but no one seems to be asking about the elephant in the room:
Do we actually know how to solve the continual learning problem required to sustain such a local co-learning loop?
Right now, there are roughly three buckets of learning methods being entertained in practice:
1) Non-parametric learning, mostly in some form of memory (skills, RAG, knowledge bases, KV cache, etc)
2) Domain-level post-training (e.g., continual pre-training on proprietary organizational data)
3) Task-level post-training (e.g., RL for specific workflows)
(More research-y ones like new model architectures are omitted here because there's still a long way to go in both research and validation)
Is any of these methods sufficiently general and deep to sustain the desired co-learning loop across all organizations and job functions? The answer is likely negative:
> Non-parametric memory is often shallow and has limited control over agent behavior (talk to OpenClaw/Hermes users who struggle to get their agents to learn and follow the rules)
> Domain-level post-training remains expensive and has yet to demonstrate broad success outside a few exceptional domains (@Cursor's Composer may be an exception but coding is an exceptional domain in itself)
> Task-level RL is engineering-heavy, sample-inefficient, and difficult to apply when success cannot be objectively verified
The human-AI learning loop isn’t inevitable. It still needs to be invented. Solving continual learning may be one of the most important problems for building an AI ecosystem where expertise compounds locally instead of concentrating globally.
Our failures are our moat.
A scientific paper is a clean repackaging of a messy process of failed syntheses, dead ends, hints of success. That mess is the durable asset: what was tried, what worked, what failed, and why.
Compound it into weights you own. Frontier labs live by this principle. If tokens-in-context were enough, pre-training would have died years ago.
Make the model a cog in a machine you own.
◾ AI SDK → open model API
◾ http://Eve.dev → open Agent API
◾ AI Gateway → open ZDR inference
Startups and enterprises must own their data, evals, model choices, software layer. Don't outsource your brain. https://twitter.com/satyanadella/status/2076323181154230284
layers to this game
A masterpiece imo by @satyanadella of a fascinatingly true information paradox (which should remind you of the real premise of West World)
and
clear-thinking competitive strategy and positioning (if you read between the lines… https://twitter.com/satyanadella/status/2076323181154230284