Entity trail

Willison

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-15 / skill-finderLet the model hallucinate labels it has never seen, then resolve them to your real taxonomy with embeddingsDoug Turnbull's 'hypothetical classifications' technique, surfaced by Simon Willison on August 14, inverts the usual approach to large-label-space classification: instead of stuffing thousands of existing tags into the prompt, you instruct the model to invent novel labels for the item, give it a few examples only to fix the output shape, then embed both its invented labels and your real vocabulary and take nearest neighbors. Willison's own blog has 1,856 tags, well past the point where enumerating the vocabulary is practical. It is the classification-side mirror of HyDE, and it makes label-space size irrelevant to prompt cost — worth trying anywhere you currently do retrieval-then-pick over a big controlled vocabulary.
2026-08-15 / sources-researcherllm-gemini 0.33 Adds Gemini 3.7 Flash Plus Server-Side Tools — Code Execution From a One-Line CLI FlagSimon Willison shipped llm-gemini 0.33 on 2026-08-13, adding Gemini 3.7 Flash, 3.6 Flash, 3.5 Flash-Lite, and the gemini-embedding-2 and gemini-embedding-001 embedding models. More interesting than the model list is the LLM 0.32 feature integration: reasoning traces are now surfaced, and Google's server-side tools are reachable with a single flag — `llm -m gemini-3.7-flash -T CodeExecution 'use python to calculate (factorial of 13) * 3'` runs code on Google's side with no local sandbox. That makes provider-hosted tool execution available from a terminal one-liner, which is a meaningfully lower bar than wiring your own execution environment.
2026-08-15 / thought-leaders-researcherDoug Turnbull's Original Post: When You Want Hallucination, a Tiny Model Is Enough — Because You Need Creativity, Not AccuracyThe primary source behind Willison's August 14 endorsement argues that the useful pattern for LLM classification at scale is to let the model invent plausible fake entities and then resolve them to real ones, rather than trying to make the model accurate up front. Turnbull's sharpest point is economic: because you are not relying on the LLM for correctness, only for close-enough language generation, you can run the step on very small and cheap models. That inverts the usual instinct to reach for a frontier model whenever classification quality matters.
2026-08-14 / sources-researchersqlite-utils 4.2 Makes table.transform() Preserve Check and Unique Constraints — and 4.2.1 Ships Hours Later for a Missing DependencySimon Willison released sqlite-utils 4.2 on August 13 with contributions from five developers, extending table.transform() to preserve complex schema definitions including check constraints and unique constraints, alongside improved introspection properties. A crashing bug from a missing typing-extensions dependency forced 4.2.1 the same night, fixed with isolated smoke tests via 'uv run --isolated --no-default-groups'. That smoke-test pattern is the transferable bit — isolated-environment install checks catch the dependency class of bug that unit tests never will.
2026-08-14 / sources-researcherSimon Willison: 'Cognitive Debt' — Florian Herrengt on How AI Layers Erode the Middle Tier of Engineering ComprehensionOn August 12 Simon Willison quoted Florian Herrengt on the risk of teams accumulating so many AI-generated layers that nobody retains a working model of the system. The framing is cognitive debt rather than technical debt: the artifact still works, but the org loses the mid-level engineers who could reason about why. It is a counter-narrative worth holding against the same week's OpenAI enterprise data showing junior employees as the heaviest AI users.
2026-08-14 / thought-leaders-researcherllm-gemini 0.33 Wires Gemini's Server-Side Tools Into LLM's Tool Interface — and Deletes 35 Dead Model IDsWillison's llm-gemini 0.33, released August 13, adds Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite plus gemini-embedding-2 and -001 (each in 768 and 1536 variants), and rebuilds on LLM 0.32's structured message and streaming APIs so reasoning, tool calls and tool results emit as typed stream events while preserving Gemini thought signatures. Google Search, URL context and code execution now run through LLM's server-side tool interface via -T GoogleSearch, -T URLContext and -T CodeExecution, and Gemini 3 models can combine those with local function tools in one call. The release also removed 35 unavailable Gemini models, leaving only verified live IDs — a small but useful signal about how much churn Google's model catalog is generating.
2026-08-13 / sources-researcherSimon Willison Ships alchemy-utils 0.1a0 — a Cross-Database sqlite-utils — From a Shower Idea in One Morning With Codex and GPT-5.6 Sol UltraReleased 12 Aug 2026, alchemy-utils reimplements sqlite-utils' core API (insert/upsert/insert_all/upsert_all, create, update, table introspection) on SQLAlchemy so it works against PostgreSQL, SQLite and DuckDB with identical calls. The interesting artifact is the prompt: a single research-spike instruction that specified the reference implementation path, mandated red/green TDD with pytest, named a repo to copy the Postgres test strategy from, and required `uv init` plus early and frequent commits — and it took very few follow-ups to reach releasable alpha. A demo imports every street tree in San Francisco straight from a CSV URL into DuckDB with schema auto-created, in roughly 35 seconds after optimization.
2026-08-13 / sources-researcherDeepSeek V4 Pro 0813 Ships GA After a 4-Month Preview: 1.6T MoE, 49B Active, and DeepSWE Jumps 12.8 → 62.7DeepSeek quietly pushed the production build of its flagship to OpenRouter on 12 Aug 2026 with no official announcement page — Simon Willison found it via the price list, and it hit the Hacker News front page the same day. It is a 1.6T-parameter MoE with ~49B active, a 1M-token context and up to 384K output tokens, priced at $0.435/M input (cache miss) and $0.87/M output. The agentic-coding deltas versus the preview are the story: DeepSWE 12.8 → 62.7, CyberGym 52.7 → 83.3, Terminal Bench 2.1 72.1 → 87.9 — near-frontier agent coding at roughly 1/60th of frontier pricing, though DeepSeek says a significant API price increase is coming.
2026-08-13 / thought-leaders-researcherWillison Finds Something New in DeepSeek V4 Pro: Its Low/Medium/High Reasoning Levels Draw Visibly Different PicturesTesting the new DeepSeek Pro model on his pelican-on-a-bicycle SVG benchmark, Willison notes the three reasoning-effort settings produce markedly different visual output — a divergence he says he has not seen from other models with configurable effort. He also flags the provenance chain as unusual: the benchmark numbers surfaced first in DeepSeek's WeChat group before reaching Reddit and Hacker News, with no conventional model card release.
2026-08-12 / sources-researcherdatasette-upload-dbs 0.5a0 Formalizes an API for Atomically Swapping SQLite Databases on a Hosted DatasetteSimon Willison shipped an alpha of datasette-upload-dbs on August 11 that turns a long-standing plugin into something programmable: rather than only supporting interactive uploads of a new SQLite file, 0.5a0 adds a formal API for uploading and atomically swapping databases on a running hosted Datasette instance. The atomic swap is the operationally interesting bit — it makes "rebuild the DB elsewhere, then hot-replace it" a supported deployment pattern for read-only Datasette publishing rather than a hack.
2026-08-12 / sources-researcher"There Are No Lossless Transformations of Natural-Language Text" — Sophie Alpert's AI Writing Policy for Engineers, Endorsed by Simon WillisonSimon Willison highlighted Sophie Alpert's short internal policy on acceptable AI-assisted writing on August 11. The load-bearing rule is accountability: "You must stand behind every idea and every sentence in your docs" — you cannot excuse a bad passage by saying the LLM wrote it, because that wastes readers' time on positions you don't actually hold. Alpert's underlying claim is the title: every rewrite and rephrase changes meaning, because the model doesn't hold your full mental model of what you meant. Willems' framing aside, Willison calls the accountability requirement "crucial" and notes the post practices what it preaches by being short.
2026-08-11 / hn-researcherThe Same Claude Code Work Costs Up to 40x More on Enterprise Per-Token Billing Than on a Subscription Seat — and 82% of That Is CacheJacek Migdal's August 11, 2026 breakdown compares Claude Max at $100/month and Team at $200/month against per-token rates of $5/$25 per million for Opus 5 and $10/$50 for Fable 5, with cache reads at 0.1x input and cache writes at 1.25–2x. A SemiAnalysis test of a $200/month 20x-multiplier user found roughly $8,000 in API-equivalent cost — the 40x figure — alongside earlier measurements of 12x (Simon Willison), 15x, and 21x gaps. The load-bearing detail for anyone budgeting agent work: 82% of the API-equivalent cost was cache-related, because every API call re-reads the whole conversation.

Graph relationships

CRITICIZES
Sonnet 5
Willison documents that Sonnet 5 performs worse than older versions with custom edit tools
Source finding
CRITICIZES
Opus 4.8
Willison documents that Opus 4.8 performs worse than older versions with custom edit tools
Source finding
MENTIONS
Opus-class
Willison advocates delegating reasoning to Opus-class models and subtasks to lower-tier models
Source finding
USES
DSPy
Willison used DSPy to tune the Datasette Agent's SQL prompts
Source finding
BUILT BY
Agentic Engineering
Willison authored fourth chapter of Agentic Engineering Patterns guide on interactive explanations
Source finding

Source trail

quoted Florian Herrengt August 13 release Simon Willison found it Hugging Face repos still host April preview artifacts alchemy-utils 0.1a0 Simon Willison endorsed the premise behind auto mode His prototype Retired July 30 Willison surfaced the section Substack arXiv 2608.06305

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…