For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 780 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
Thinking Machines wants judgment encoded in model weights, raised locally, owned by the kitchens that hold the knowledge. Grant the destination. The harder claim is about the calendar.
They are not mainly wrong that knowledge is local, tacit, and hard to centralize. They are late for the product form they chose to carry that truth — late in a way no amount of frontier talent reverses once three clocks have already rung.
Tinker, their fine-tuning API, launched in private beta in October 2025 and went generally available in December 2025. The civilization-scale mission essay — human will, decentralized alignment, values in the weights — arrived in July 2026.
Philosophy after the tool. That order is normal for a lab. It is fatal for a claim that still pretends the future of AI is open to being steered toward multi-tenant ownership of specialized models. By the time the essay names the destination, the road most organizations actually take has already been paved by someone else.
Defaults hold not because they are best but because defecting is costly in ways spread across everyone who would benefit from defecting. ChatGPT, Claude, Copilot, Cursor: the world learned a single verb for intelligence work. Open a chat. Rent a frontier model. Paste context. Ship.
That verb is the product. "Fine-tune your own weights on your kitchen's knowledge" is a different verb. It asks for data hygiene, eval harnesses, training judgment, and a willingness to own a model that goes stale when the base model races ahead. The default already internalized the opposite lesson: the base model will be better next quarter; do not marry weights; marry the session.
Once a default sits that deep, a better moral story about ownership does not reopen the path. It narrates a path most people will never walk. Tinker can be excellent infrastructure for the people who already needed a training API. It cannot undo the habit the frontier chat products installed in everyone else.
In 2023, a custom fine-tune was often the difference between a toy and a tool. Domain vocabulary, house style, task format — the general models were weak enough that local weights paid for themselves.
By 2026 the same jobs mostly route through a stronger base model plus private context: retrieval, tools, agents, long windows, docs the model can read without baking them into parameters. The specialization still exists. It moved out of the weight file and into the wrapping.
That is the timing error inside their own Hayek story. Local knowledge did not stop mattering. The cheapest carrier of local knowledge stopped being a specialized model. It became a specialized environment around a rented general one. A lab that ships better fine-tuning in 2025–26 is improving a layer that lost share of the real work to a layer above it. Layer elimination, not ideology: when the general function gets good enough, the translation step of "encode the kitchen into weights" becomes optional overhead for most kitchens.
Their essay still speaks as if the scarce act is shaping a model. For most organizations the scarce act is already shaping the loop that uses a model they do not shape.
They are right that human participation is a technical challenge, and right that solo time-horizon benchmarks measure the wrong thing. The clock still runs against them.
Every quarter the charts they criticize improve: longer autonomous software tasks, fewer human checkpoints, more work that completes without inviting the craftsman back onto the line. That is the capability curve. The participation architecture they want — live multimodal collaboration, models that know when to stop and ask, ecosystems of disagreeing specialized minds — ships on a product and research schedule measured in years.
Run capability faster than the check and the bottleneck moves to whatever you did not speed up. Here the thing that did not speed up is the organizational habit of wanting the check. By the time the interface invites judgment, many orgs have already staffed, budgeted, and culturally rewarded the substitution path. The craftsman is not waiting on better tools. The plant reorganized without him.
This is the intelligence-curse timeline in miniature. Power that needs less from people forms while the tools for keeping people in the loop are still a roadmap item. You cannot reopen a dependence dial after the firm has already priced people as optional.
Capability and diffusion are both exponential and decoupled. Most AI error is measuring one and calling it the other.
Thinking Machines' essay reads the capability diagnosis correctly: centralized frozen intelligence fails on local knowledge. Their ship date reads the diffusion schedule of a developer training API. The future they describe — organizations that own their models the way they own their processes — requires a third curve they do not price: ownership culture. How fast do firms learn to treat AI as something they raise rather than something they rent?
That curve is slower than Tinker adoption and slower than model release notes. It is the curve that already bent the other way when the default locked. Building in the gap between capability and diffusion is alpha. Building a mission for a gap that closed behind you is memoir.
Late does not mean Tinker is useless. Researchers and serious builders will keep fine-tuning. Open-weight ecosystems need training APIs. Some regulated domains will still require weights they can audit and freeze.
Late means the civilizational product thesis — that the path still open for "the future worth building" is decentralized weight ownership at the scale of human knowledge — missed the window when that was still the binding constraint. The binding constraint moved to defaults, harnesses, context, and whether participation is still staffed as a job.
Late also means the essay's enemies mostly already won the habit war. Frozen models from a handful of places are not a speculative dystopia. They are how almost everyone already works. Announcing that this is bad in July 2026 is correct and after the fact.
The timing claim fails if any of these land hard:
Until one of those moves, "we will re-decentralize AI through better training tools and a human mission" is a lab describing the war it wished it had fought on time.
They are early on talent and capital. They are on time for a fine-tuning product in a crowded infra market. They are late for the claim that the future is still choosable along their axis.
Even if you grant them weight ownership as the right unit, the calendar already chose the default they are writing against.
You do not lose the future by being philosophically incorrect. You lose it by publishing the philosophy after the habit is installed.