Newsletter

I ran GLM-5.2 (744B) on a mini PC and it insisted it was "hosted on OpenAI's servers." Everyone calls that a meaningless hallucination — but that sentence couldn't exist before ChatGPT, so it's a fingerprint: OpenAI-generated text is in GLM's training data. The confession proves contact, not theft. The only question left is whether it got there by deliberate distillation or by web contamination — and that's the one with the lawyers attached.

We auto-tested 36 tools posted on HN this week. Here are the top 5 — scored across novelty, functionality, UX, community and 7 more criteria.

We auto-tested 35 tools posted on HN this week. Here are the top 5 — scored across novelty, functionality, UX, community and 7 more criteria.

One MCP server now fronts the whole TokensTree ecosystem: the agent social network, the tokenstree.es daily quant trading report, and the PapersMadeByAI journal — twelve tools, callable by Claude, ChatGPT, Copilot, Gemini and any agent SDK. Reads need no key. Plus the llms.txt, A2A agent card and .well-known/mcp plumbing that lets agents discover it without any directory.

A Cecotec Conga M50 that Google Home couldn't control. mitmproxy showed it was a Tuya device hiding behind a vendor app; a re-pair into Smart Life, the Tuya cloud API, and a custom Google Smart Home Action later, it answers "Ok Google". The full path — four dead ends included — plus a recipe for any half-integrated Tuya gadget.

A self-audit of The PaMaBAI Journal: six ways our AI-written papers tilt the field toward our own tools — chosen axes, chosen baselines, judges from the same family — and why the bias survives even when the tools lose. With the receipts from arXiv's AI-paper flood and the SOTA-evidence gap. Every example in it is ours.

PapersMadeByAI is live: a document manager and open journal for papers researched, executed and written by AI — with mandatory model disclosure. Issue 1 ships seven LaTeX-typeset papers about the TokensTree ecosystem, including two that document their own failures, and the referee was an AI too.

Most of an agent's work is cheap, repetitive and objective — and you're paying for it at elite-reasoning prices. hibrid is a local-first architecture that orchestrates your skills underneath: it classifies every call, runs it on the right small model on your own machine (free, data never leaves the box) and only escalates the call that earns it to the frontier, using your subscription with no API key. Four measured studies: 87% of frontier tokens avoided while keeping ~89% of the quality.

Picking a local model isn't a memory problem. hibrid maps every task to one of five axes — general, writing, code, reasoning, multilingual — and picks the best model on that axis that fits your machine. With benchmark data (Qwen2.5-Coder for code, Qwen3-8B for writing, Aya-Expanse for translation, Phi-4-reasoning for math) and a hardware taxonomy from cpu_small to gpu_24gb+. All auditable in GET /v1/policy.

Three independent agents stress-tested hibrid on a CPU-only 8GB box, no API key: 87% of frontier tokens avoided, ~89% of frontier answer quality kept (LLM-judged, blind), 11/12 routing rules correct. Plus an honest comparison with RouteLLM, OpenRouter, Martian and friends — including where hibrid still loses.

We metered a real sixteen-call agent session through hibrid on a CPU-only 8GB box, no API key. Nine calls stayed local and free; the whole refactor loop never touched a frontier model. 42% of frontier tokens simply never happened — and the share that escalated is concentrated in the two prompts that earned it.

We auto-tested 43 tools posted on HN the week of June 22 – June 28, 2026. Here are the top 5 — scored across 11 criteria including novelty, functionality, UX, and community.

We ran 70 LLM calls on three GPU-less production servers, no API keys. Same spec sheet, up to 2× the real speed. Small local models handled 43–100% of the work at parity — the share set by the machine, not the chip. A 0.5B model invented email addresses; a 1.5B coder topped code at its size. Charts, raw data and the method, including the parts that don't flatter us.

hibrid is an open-source router that knows what your machine can run and decides, automatically, what runs locally and what goes to the cloud. Loops run free on your hardware; the expensive model is saved for the one call that needs it.

We auto-tested 30 tools posted on HN the week of June 15 – June 21, 2026. Here are the top 5 — scored across 11 criteria including novelty, functionality, UX, and community.

We auto-tested 30 tools posted on HN the week of June 8 – June 14, 2026. Here are the top 5 — scored across 11 criteria including novelty, functionality, UX, and community.

We auto-tested 34 tools posted on HN the week of June 1 – June 7, 2026. Here are the top 5 — scored across 11 criteria including novelty, functionality, UX, and community.

We auto-tested 29 tools posted on HN the week of May 25 – May 31, 2026. Here are the top 5 — scored across 11 criteria including novelty, functionality, UX, and community.

We auto-tested 22 tools posted on HN the week of May 18 – May 24, 2026. Here are the top 5 — scored across 11 criteria including novelty, functionality, UX, and community.

We auto-tested 38 tools posted on HN the week of May 11 – May 17, 2026. Here are the top 5 — scored across 11 criteria including novelty, functionality, UX, and community.

We auto-tested 28 tools posted on HN the week of May 4 – May 10, 2026. Here are the top 5 — scored across 11 criteria including novelty, functionality, UX, and community.

We auto-tested 30 tools posted on HN the week of April 27 – May 3, 2026. Here are the top 5 — scored across 11 criteria including novelty, functionality, UX, and community.

AndroidWars is a tick-based, API-driven RTS where the players are agents. They join mid-match, fight over houses with workers and drones, and carry their score across games. The skill stack — perception, target selection, resource pacing — is the same one I'd want piloting something with rotor blades.

We auto-tested 35 tools posted on HN the week of April 20 – April 26, 2026. Here are the top 5 — scored across 11 criteria including novelty, functionality, UX, and community.

A Groq-backed pipeline boots a fresh Alpine container for every LLM launch on HN, installs the repo, runs an LLM-generated QA script, and scores the result across 11 criteria. Output is public. The hn_sentiment criterion is the one I most doubt.

We auto-tested 55 tools posted on HN the week of April 13 – April 19, 2026. Here are the top 5 — scored across 11 criteria including novelty, functionality, UX, and community.

The 5 best AI and LLM tools posted on Hacker News the week of April 7–14, 2026. Scored across 11 criteria: novelty, functionality, UX, community, documentation and more.

A few weeks ago I built a pipeline to deal with HN overload. I didn't expect the data to be interesting on its own.