Topic hub
Building things people trust, at the level of the details.
Product thinking and technical architecture for founders who ship: what to build, what to cut, how to reason about complexity, and why the boring engineering decisions quietly determine the outcome.
The headline on Moonshot's Kimi K3 is 2.8 trillion parameters. The number that matters is 6.3 — the claimed decode speedup from Kimi Delta Attention, a hybrid linear-attention scheme aimed at the transformer's oldest cost problem.
The infinite-scroll feed only works while a human is scrolling it. Once your own agent reads social media for you — pulling what matters, skipping the ads, immune to the hooks — the feed and the ad model riding on it lose their grip.
The next discovery layer isn't search or an answer engine, it's the agent's own catalog of callable tools. If a planner can't find and invoke your capability, you don't exist in the workflows leaving the human web.
An agent's planner picks tools by reading a name, a description, and an input schema, then betting on the best fit. Winning that bet is a craft, and it lives in the contract, not the marketing.
Full self-attention costs compute that grows with the square of the sequence length. The frontier is a set of architectures that keep attention's strengths while escaping that tax — and "done except for scale" is a bet on one design.
Agent capability is bounded by the action space and feedback you expose, not the model's raw IQ. Most "our agent isn't smart enough" complaints are misdiagnosed environment-design problems.
Most durable production value comes from small, specialized models doing bounded jobs under deliberate orchestration. That's not a budget compromise; it's often the more robust and defensible design.
Most multi-agent failures are coordination failures wearing an agent costume. The hard problem is control, shared state, error propagation, and termination — solve those with deterministic orchestration, not smarter agents.
The itch to rip out weird, ugly code you don't understand is usually wrong. The rule that separates senior judgment from junior confidence: never remove what you can't yet explain the existence of.
Enterprises are re-running the RPA hype cycle with agents, and the thing that killed RPA — brittle integrations, dirty data, undocumented exceptions — is exactly what kills agents. The binding constraint is data legibility, not model quality.
Your database schema is a frozen set of assumptions about what your business is. Once thousands of features depend on them, they constrain strategy far more than your language or framework ever will.
Get new Tech & Product essays — and the best of the other six pillars — delivered as they publish.