AI Editorial

The rest of the library explains, and explanation has no opinion. This is where the opinion goes. Each piece starts from something that actually happened in AI — a model release, a paper, a benchmark result, a failure, a policy — explains the mechanism underneath it, and argues its way to a position: what genuinely changed, what is only being marketed as change, and what someone learning the field should take from it. Every piece links what it argues from and the concepts you need to follow it.

4 editorials 4 strands 31 cited sources 6,471 words

All Research & Technique1 Agents & Autonomy1 Inference & Economics1 Evaluation & Evidence1

Inference & Economics 20 September 2026 8 min read New

Speed goes to whoever is easiest to predict

vLLM's September release lets a client ask, per request, what fraction of its tokens a small draft model guessed right. The same stack cuts the draft length to zero above 128 concurrent requests. Both are the same admission.

Speculative decoding never made a model faster; it converts a server's idle arithmetic into lower latency, and because batching spends that same arithmetic, the stack now rations the speedup to the most predictable requests first and to nobody once the machine is full.

Read it → 8 sources speculative decodinginference servingbatchinglatency

Agents & Autonomy 19 September 2026 7 min read New

The tab is already a client

Four MCP servers were found with the same flaw in eight days, two of them scored Critical. In none of the attacks does anyone say anything to the model. The agent security that is failing right now is not about the model at all.

A local MCP server's most likely attacker is a browser tab rather than a poisoned prompt, and the specification's MUST for Origin validation does not prevent it because nothing in the stack fails when a server omits it.

Read it → 7 sources mcpagent securitydns rebindingtool use

Evaluation & Evidence 18 September 2026 7 min read New

A score is a reading, not a property

Four commits landed in the Terminal-Bench repository on 11 September. None of them touched a model, and all of them changed what its tasks measure. Agentic benchmarks are maintained software with an expiry date, and their numbers should be read that way.

An agentic benchmark score is not a property of a model but a reading taken through a perishable instrument of scaffold, resource budget and expiring container.

Read it → 8 sources benchmarksagentsevaluationreproducibility

Research & Technique 18 September 2026 8 min read New

The fourth layer remembers

Kimi Linear landed in Hugging Face Transformers this month, and its layer plan matches models from two other labs: three layers that keep a fixed-size summary of the past, then one that keeps all of it. The ratio is the same in all three, and no derivation for it has been published.

The 3:1 hybrid ratio converging across three labs is not a way station on the road to pure linear attention but the field's standing price for exact recall: three layers may forget, and one may not.

Read it → 8 sources attentionlong contextkv cachemodel architecture