Blog

Notes, essays, and architecture articles.

Architecture over models

Why model selection is never the core problem — and why systems that last are built above engines.

Read →
Why LLMs are not intelligence — they are an interface

Language models generate text. Intelligence manages goals, constraints, memory, and truth conditions.

Read →
The “Hello problem”: minimal input should produce minimal output

If “hello” yields a short response, that’s not coldness — that’s system health.

Read →
Why “chatbot” is a dead end

Chatbots optimize for conversation. Runtimes optimize for correctness, control, and usefulness.

Read →
Determinism vs. probability

Use probabilistic generation for language — and deterministic logic for truth.

Read →
Memory is not a vibe — it’s an implementation

“I remember” is meaningless unless the system can retrieve, verify, and update memory.

Read →
Why multi-provider support matters (even if you love your provider)

Vendor lock-in is not a business plan. It’s a failure mode.

Read →
Why prompt engineering doesn’t scale

Prompts are fragile contracts. Roles and rules are architecture.

Read →
Cognitive roles: the next step after prompts

Roles are not “act like a lawyer.” Roles are enforceable behavioral modules.

Read →
Why English and Russian can feel like different “personalities”

Different languages create different registers. That’s normal — and useful.

Read →
“It feels like mood” — because you didn’t force a script

When you remove forced greetings, you get human-like mirroring.

Read →
Why AI systems should sometimes respond with less

Over-explaining is not intelligence. It’s noise disguised as helpfulness.

Read →
Why “model wars” are a distraction

The best model won’t save a bad system. A good system makes models interchangeable.

Read →
Real usage lessons: where AI products break first

The first failure is not “wrong answer.” It’s inconsistent behavior under pressure.

Read →
Why we write Notes instead of ads

Architecture cannot be sold with slogans. It can only be understood through clarity.

Read →
Why the Russian language exposes the limits of next-token intelligence

Meaning in the Russian language often lives in states, ambiguity, and non-local context. This article explains why that structure breaks probabilistic next-token models — and why using the Russian language as a stress test reveals architectural limits that are invisible in simpler linguistic regimes.

Read →