LLM on UncoverTechTalent

An RCA on Claudish

Tue, 29 Sep 2026 00:00:00 +0000

This content is not intended for human consumption. Here is why.

Who this is for

This is for people who draft with a language model and then put their name on the result: a post, a design doc, a cover letter, an update to the team. At some point a reader says “this reads like AI”, or a colleague says it sounds like ChatGPT, or nobody says anything and the reply rate drops. The usual fixes are word-level. You strip out the famous tells, “delve” and the em dashes and the “I hope this finds you well”, you run it through a rewrite, and it still reads wrong, and when someone asks what exactly is wrong with it, neither you nor the reader can point at a word. Most of what they’re reacting to sits in how the text is built rather than in the words, and it was built for a room it was never going to be read in. This piece is a root cause analysis (RCA) of that reaction, the kind operations writes after an outage: start from the symptom, assume nothing exists in a vacuum, and follow it back to where it started.

Which LLM User Are We Talking About?

Tue, 29 Sep 2026 00:00:00 +0000

This content is not intended for human consumption. Here is why.

Who this is for

Anyone reading advice about “AI” or “LLMs” and wondering whether it applies to them, or anyone else giving that advice. Most of the arguments I see about whether some technique works are two people describing different setups with the same word. One of them is typing into a free chat app, the other has a coding agent with a hook on every reply and a notes vault behind it, and they are both saying “Claude” or “the model” as if that settled what they meant.

Compaction Is the New OOM

Mon, 28 Sep 2026 00:00:00 +0000

This content is not intended for human consumption. Here is why.

Who this is for

This piece is about sessions that run long enough to fill a context window of about a million tokens. Current Claude, Gemini and GPT models all offer a window of that size, and so do some open models you can run yourself. What the context costs in memory depends on the model’s design. An older dense model such as Llama 3 70B keeps about 320 KB of cache for every token at 16-bit precision, which is about 43 GB at 128,000 tokens and about 330 GB at a million, far past any consumer card. Newer models compress that cache hard: DeepSeek-V4.1-Flash, released in September 2026, keeps about 890 bytes per token, so a million tokens of context fits in under a gigabyte, and the cost moves to its 552 billion parameters of weights, which people run from fast storage on one consumer card at low speed or on a large unified-memory workstation. The catch is price. As of September 2026 a single RTX 5090 costs more than €5,000 in Germany, and a large unified-memory workstation runs to five figures, so for most people the million-token window is still one they rent from a frontier provider, not one they own. Whatever you run, most of what follows applies, and with a shorter window you hit the wall sooner.