AI
21 essays.
- No Vector Store: a 25.87 MiB SQLite Index Reaches 86.62 Recall@10 SAG2 moves the semantic judgement from index time to query time. One SQLite index plus BM25 and JEV judgements tops the compared systems on R@5 and R@10 across MuSiQue, 2WikiMultiHopQA and HotpotQA, with a 2.36-second build and zero model calls. The ablations split the gain into two expansion rounds and a shared context, and the cost sits in the last section. AI
- Designing a Coding Agent Without a KV Cache A September 2026 design note starts from one counterfactual, traces six familiar coding-agent flaws to a single cause, and proposes a harness built around a Jev decision layer. Routing arithmetic, token shares and security routing are each marked by source. AI
- Multi-teacher distillation is a token budget TRL's multi-teacher on-policy distillation example ran 60 steps. Instruction following held 20 percent of the prompts and got 1 to 3 percent of the trained tokens, and it was the only domain that barely moved. Open-MOPD calls this token budget misallocation, and fixing it raised the recoverable headroom from 35.6 to 83.4 percent. AI
- The Ontology Divide Is Not Model Size Six papers from late 2025 and 2026 point the same way. The bottleneck in ontology construction moved from who writes it to who can prove it is right, and every number here carries its source. AI
- Evals: Four Numbers That Tell You Whether an AI System Actually Works Evals split into inputs, system, grader and number, with the judge itself graded first. Every figure here is tagged with its source. AI
- Retrieval as Navigation: DecisionRAG Asks a Tree Instead of Scoring a Corpus DecisionRAG runs no first-stage search and builds no vector index. Each level puts the candidate titles into one shared context and answers a batch of yes/no questions in a single forward pass. It wins six of nine metrics, and the three losses and the costs are in the body. AI
- One Graph, Fifteen Spellings, a 28-Point Difference Oracle wrote the same knowledge graph in 15 text formats and ran nine frontier model configurations over four benchmarks. Extraction barely notices, with 1.3 to 3.9 points of spread. Structural reasoning moves 28.4. AI
- Project EVE: What Happens When Agents Enter an Economy That Already Has One Anthropic let Claude make markets in EVE for six months and fly a four-agent fleet for a week. The result prices the agent economy in three layers: execution, tooling capital, and the cost of trust. AI
- The four-layer memory stack, and the dreaming loop beside it: reading Anthropic's memory research Anthropic splits agent memory into transcripts, notes, representations and injected context, and lets "dreaming" review the record offline — correcting notes, building wikis, lifting the few that belong in every session. A layer-by-layer read, with the experiment numbers, an engineering checklist, and four verified talks in English and Chinese. AI
- The Database Got a New User: Record Stores and Context Layers for Agents The database's main user changed from people to agents, so it split into two legs: one holds state, one supplies context. Retrieval itself stopped being a recall problem and became an orchestration-and-compilation problem. Two essays, four papers, six vendors, one picture of where this actually stands. AI
- Seven layers between the model and the result: reading the guide against its own evidence A guide that splits a production agent into seven layers, whose strongest claim is that the model was never the problem. This reading reorders the argument by evidence strength — practitioner notes, one public incident, sponsored copy — adds the two layers it skips (cost and observability), and ends with eight checks. AI
- How one prompt grew into a company brain: Recall in four generations A retrieval skill, four generations deep, and the two hard decisions behind it: store the source upstream and treat everything else as a rebuildable projection, and let the server own recall while the calling model owns precision. With four sets of measurements and four steps you can copy. AI
- Reading DSec section by section: what a 3-million-sandbox-a-day platform says about agent training A section-by-section read of DeepSeek's 31-page DSec report: the workload numbers first, then three mechanisms (composable layers, memory reclaim, on-demand images), and what the 1.71x and 45.2% -> 17.3% results do and do not prove. AI
- Open Responses: one contract instead of N×M adapters A shared spec built on OpenAI's Responses API replaces a pile of provider adapters. The interesting part is not the field names but the state machine, the semantic events, and the MUST behind allowed_tools, truncation and previous_response_id — those decide who gets to make which call. AI
- A chat box is not an agent OS: five places a 30-person company breaks it One claim — a chat UI on a single-user harness is not an agent OS — reads like a slogan until you take it apart. Five breakpoints: who is speaking, what it remembers, what survives a crash, which machine it acts on, and whether the reply was delivered. Checked against the Apache-2.0 repo, with the price attached. AI
- Automatic Ontology Engineering: Reading NVIDIA's Auto-Ontology at the Source Auto-Ontology compiles an ontology out of metadata and query history in four passes, then hands agents nine read-only MCP tools. After reading 1.1.0 end to end, I think it automates the first month — and that certification is a badge, not a gate. AI
- One person as a creative team: what transfers from Dan Koe's letter Dan Koe splits the creative director job into brand strategy, content strategy, and writing plus ops, then wires the three together with a nine block interview. This piece checks the three numbers that carry the urgency and marks the three things a technical reader should not copy. AI
- Cognitive surrender: hand off the work, not the judgement Addy Osmani splits agent use into delegation, deferral and surrender. This piece draws the slope, adds the Wharton and Anthropic numbers, and marks the places his argument does not quite reach. AI
- Think Before You Act: How JevAct Gates LLM Agent Actions JevAct inserts a gate before execution: code settles what tool records decide, a decision model answers only what the conversation says, and a boolean rule combines them. All eight agents improve on τ²-Telecom — and the paper is explicit about where the gate fails. AI
- Super-agent: DeepSeek Takes Sessions Away From the User DeepSeek Harness argues the bottleneck has moved from the model to human attention. Five modules, one two-layer design, and the part that actually runs in the open-source prototype. AI
- A Decentralized, Trusted MCP Runtime: Architecture and Implementation Agents are bounded by the tools they dare to call. An architecture that treats tools as content-addressed WASM modules, executes them in a sandbox, and makes them attributable — plus the five layers that implement it. AI