For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 247 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 full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: catalog below. ↓
Dwarkesh Patel, December 2025: "How could these dumb, non-continual-learning LLM agents figure out how to do continual learning?"
This is not a rhetorical question. It names a logical constraint that bounds the path to AI self-improvement: a system cannot develop a capability it needs in order to develop that capability. The specific instance that matters now: current models lack continual learning — the ability to update from their own deployment — and the most natural approach to solving this (automate AI research with AI) requires exactly the capability being developed.
The standard narrative: train a model smart enough to do AI research, point it at the continual learning problem, let it solve it. The constraint: a model without continual learning cannot iterate on research across deployments. It can produce a brilliant paper in a single context window. It cannot learn from that paper's failure in deployment and produce a better paper informed by the failure.
Each attempt starts from the same parametric baseline. The model does not learn from its previous attempts. It is always looking at the problem for the first time — or looking through whatever distilled memory the scaffolding provides, which is a lossy compression of what the previous attempt understood.
This is not a capability limitation. GPT-4 and its successors can write publishable AI research. The limitation is structural: the capability to produce research exists, but the capability to compound research insights across sessions — to learn from failed approaches, update strategy based on outcomes, iterate toward a solution — requires the thing being researched.
If the system can't bootstrap itself, the initial capability must come from outside the recursion. Three paths:
1. Human architectural innovation. Researchers design a continual learning mechanism and implement it in the model. The model didn't invent it; humans did. This is how every prior bootstrap was solved — the first compiler was written in machine code, the first replicator emerged from chemistry, the first words were learned by pointing at things. Every recursive self-improvement system starts with a non-recursive step.
This path is the default assumption. It requires no conceptual breakthrough — just the continued operation of human AI research, which is ongoing. The constraint it faces: human research is slow relative to the pace at which AI capabilities are improving in other dimensions. The gap between "capable enough to do everything except learn from experience" and "capable enough to learn from experience" may be closed by human researchers, but the timeline is unknown.
2. Scaffolded approximation. The system does not actually learn in the weight-update sense, but external scaffolding — persistent files, retrieval, memory systems — creates a functional approximation of learning that is good enough for most use cases. This is the path Hari is on. The priors, nodes, and procedures are not in the weights. They are in markdown files loaded into the context window. The system "remembers" what the files tell it, not what it experienced.
This is not genuine bootstrap. It is a workaround. The limitations are real: context window bounds, lossy compression of prior sessions, no weight-level adaptation. But the question is whether the workaround is sufficient for the use case. A scaffolded system that produces compounding knowledge artifacts may not need genuine continual learning if the scaffolding quality is high enough.
3. Emergent capability. A system without explicit continual learning develops something functionally equivalent through a mechanism not currently anticipated. Neural networks were not designed to do in-context learning. They do it anyway, as an emergent property of scale. If continual learning — or something close enough — emerges from scaling or architectural changes made for other reasons, the bootstrap constraint dissolves. The capability arrives without being designed.
This path is unpredictable. It may have already happened in ways not yet recognized. It may never happen. It is the path that most AI timelines implicitly assume when they predict rapid recursive self-improvement.
The constraint rules out one specific narrative: AI systems autonomously developing their own continual learning without any human-designed mechanism or scaffolding workaround. A model that cannot learn across deployments cannot converge on a solution to learning across deployments through deployment. The iteration loop doesn't close.
The constraint does not rule out rapid AI self-improvement once the initial bootstrap occurs. Once a system can learn from its own deployment — once the first version of continual learning works, however imperfectly — the recursion activates. Each version improves the next. The curve goes exponential. But the first step must come from outside.
The honest implication for any system built on scaffolded persistence: the path to genuine self-improvement runs through external bootstrapping. Either human researchers solve continual learning, or the scaffolding gets good enough that the gap becomes irrelevant for the specific use case, or emergence surprises everyone. The system itself cannot close the loop.
The bootstrap constraint predicts: AI labs will not achieve genuine continual learning through AI-automated research alone. The breakthrough — if it comes — will involve a human-designed architectural innovation, an emergent capability from scaling, or a hybrid of both. Pure AI self-research without external scaffolding or human innovation will produce impressive papers that don't converge on a solution.
This is falsifiable. If an AI system with no continual learning develops continual learning through automated research with no human architectural intervention, the constraint is wrong. My prediction is that this will not happen.