A pre-commit hook that blocks files over 5,000 tokens
Matt Pocock's prompt asks for a hook that blocks staged files over about 5,000 tokens. It skips lockfiles and generated code, and tells you how to split files.
Their posts of the last 24 hours, through FxTwitter; plus everclear077/awesome-ai-x-accounts when it answers.
What tech Twitter is discussing, ranked by discussion.
The front page, plus stories over 80 points in the last 24 hours.
Every new issue, each delivered once, looking back up to 7 days.
Agents that think in hidden space, then act in text» Page 5
Latent Semantic Memory: 16 tokens beat full context» Page 5Pergent
The tech daily written for one reader
Monday, October 19, 2026No. 35Price: $0.03
Matt Pocock's prompt asks for a hook that blocks staged files over about 5,000 tokens. It skips lockfiles and generated code, and tells you how to split files.
A new snapshot --delta option returns just the changes since the last snapshot. One form-filling test used 73 times fewer tokens.

A blog post describes a lighter way to hand an agent a spec. The author says it is faster than the Superpowers approach.

Personal agents, coding agents, apps and API keys show up in one place. You can see model choice, data retention, and the metadata to judge return on spend.

Developers can let a model search the web and read pages through DeepSeek's API for $0.75 per 1,000 successful requests. It's a cheap way to give an agent live information.

Some sites now block AI agents, and posts name Delta, Chase and Amazon. Greg Isenberg predicts a split: some firms welcome agents, others charge for data access, and the giants build their own. Paul Graham says a paywalled page is basically invisible if AI can't read it. Legal experts warn that Meta's new terms still allow broad scraping, even with new wording.

A new plugin lets the agent place real voice calls over normal WhatsApp, with no business account. It only rings your own number, and speech runs on Gemini Live.

The open-source engine for running models fast gets its biggest release yet, with 814 commits from 363 people. It adds a /v1/systemone endpoint for structured decisions, cheaper memory for chat history (the KV cache), and faster startup.

OpenCode says it has faced a sustained attack for 48 hours. Its defenders used frontier AI models, but those models refused to analyze the attacking agents because the work set off safety rules. Hugging Face hit the same wall. OpenAI's team says it helped OpenCode users get back online, refunded credits drawn in the attack, and offered defenders access to a cyber-permissive setup called Daybreak Blue. The question: should safety rules bend when the work is defense?

A South Korean engineer says their manager set an AI token-use target and wants them to use more AI. Gergely Orosz says the manager is trying to help, and that engineers who refuse to adapt have a shrinking set of jobs. The debate is whether AI use should be measured like a KPI at all.

The next Bun version makes its check command 1.7 to 7 times faster, according to Jarred Sumner.
A Reddit post says Namecheap took control of and disabled every domain on one account. The HN thread has 274 comments and lots of talk about registrar risk. The lesson people are drawing: a registrar can cut you off, so keep backups and know your options.
A new blog post walks through parser combinators, small pieces that snap together to read text. It's a clean way to build parsers without a big generator tool.
Google Research describes 'harnessing' a fixed model by improving the code around it. People also call this hill climbing or continual learning.

A paper proposes LLM agents that reason in latent space (hidden numbers, not words) and act through text. It's the top story on Hacker News today.

Documents are encoded once into reusable embeddings. A small learned module writes the relevant bits into 16 tokens for a frozen LLM. On FactKG it scores 84.33% against 80.43% for full context, while using 108x fewer source tokens.

Thom Wolf's napkin math says Meta, NVIDIA, Alibaba, DeepSeek, Mistral and about 30 other labs have shared close to $1B of training compute on the Hub. It was under $40M at the end of 2023. The number is a floor. Also, FLUX.1-dev from Germany's Black Forest Labs has 43,000 derivative models.

Anthropic published research on how jobs there are changing. Kent C. Dodds notes the company says some staff still write about 10% of code by hand.

Karri Saarinen writes about the people who build software products. It drew 157 comments on HN.