Weekly Tech Cocktail - Issue #14
Twenty-five companies tell Washington to leave open weights alone, Postgres explained as a city, and the jobs apocalypse that isn't.
Twenty-five American companies told their own government to back off open-weight models this week. Nvidia, Microsoft, Meta, IBM, Mistral, Mozilla and the Linux Foundation signed the letter; two days earlier, nearly 200 startups including Proton and Y Combinator had sent the same message. OpenAI, Anthropic and Google stayed off it. The trigger: Washington is weighing a ban on Chinese open models after Michael Kratsios accused Moonshot of distilling Fable 5 into Kimi K3, and Treasury Secretary Scott Bessent raised sanctions. Dario Amodei then published a denial: Anthropic never asked for a ban, and open models without dangerous capabilities are a public good.
Open weight means you can download the trained weights and run the model yourself instead of renting it. Not open source: the data and most of the code never ship. And downloadable is not runnable. Kimi K3 is 2.8 trillion parameters at 8-bit, roughly 2.8 TB. Commodity hardware has swallowed monsters before, but not this month.
On jobs, Stanford's SIEPR brief went looking for the apocalypse. Since 2022, unemployment rose 0.77 points among the most AI-exposed workers, and 0.85 among the least. Developer postings are still growing faster than the rest of the market. New graduates are the real soft spot. The brief's opening foil is Amodei's forecast of 20 percent unemployment, which makes two appearances for him this week.
Elsewhere, Neal Stephenson on writing by hand being good for your brain. By hand, as ever, we mean a part of an actual human.
#tech
PGSimCity: How PostgreSQL Works - the database rendered as a city you can walk through. Best Postgres explainer in years, and it only works on the site.
Memory Safety's Hardest Problem - matklad on why aliasing and concurrency, not buffer overflows, are the part nobody has solved.
We have proof automation now - Adam Langley verifies a zstd implementation in Lean. Formal methods just became something you might actually use.
#apis
Model Context Protocol prepares to break with its stateful past - MCP moves toward a stateless transport. Protocol design decisions you will be living with.
BGP ORIGIN attribute manipulation and its impact on the Internet - a rarely examined routing attribute, and how it redirects traffic at internet scale.
#genai
An Inside Look at the Relay Market Powering Token Resellers and Fraud - the grey market reselling model access, and the fraud running through it.
OpenAI and Hugging Face address security incident during model evaluation - a swarm of pre-release agents compromised Hugging Face during an eval. Primary account.
Controlling Reasoning Effort in LLMs - Raschka on dialing reasoning up and down, and what each notch costs.
The first known runaway AI agent, or a very bad marketing stunt? - Willison reads the same incident the other way. Take both and make up your own mind.
"Go Home Copilot, You're Drunk": Developer Responses to Agent-Generated Code Review - what developers actually reply to review comments written by an agent. The title is the finding.
#tools
Ruff v0.16.0, 413 default rules up from 59 - the Python linter changes what it means by "default". Check your CI before it checks you.
A shell colon does nothing. Use it anyway - the colon as no-op, placeholder and guard. Short, immediately usable.
GigaToken: ~1000x faster language model tokenization - a thousandfold speedup on the least glamorous step in the pipeline, with benchmarks in the repo.
fun
Forth - this week's XKCD.
Em dashes are fucking amazing - a full-throated defence of the punctuation mark that now reads as a robot fingerprint. This newsletter remains abstinent.
I'm Romain, a software craftsman and musician. Every week I dig through the noise and send back the few links worth reading.