Quad9 leaves France on November 24
Four more reads, and Herbert against vLLM and llama.cpp on five GPUs.
Not getting this letter yet? It goes out every Friday by email, in English or French. Subscribe to get it automatically.
SubscribeHello,
On the menu this week: the third article of the Herbert series, where the same compiler moves to a GPU, and five reads of the week.
The reads: DeepSeek opens its inference engine, PHP, Node.js, Java and pip patched on the same day while OpenCode had keys stolen, Quad9 leaving France, Meta opening Muse Spark, and Anthropic putting 5% of its shares into a public benefit fund.
The previous article showed how Herbert writes its machine code at startup, for the processor in front of it. This one asks the same question on a GPU: Herbert writes its Vulkan kernels for the card it finds, against vLLM and llama.cpp, at equal format, on five consumer GPUs. The cases where it is slower are in the text, with their numbers. The predictions were written beforehand.
Herbert writes its Vulkan kernels at startup, for the card in front of it. Against vLLM and llama.cpp, at equal format, on five consumer GPUs: where it is faster, where it is slower, what makes the difference, predictions written beforehand.
Herbert writes its machine code at startup, for the model and for the processor in front of it. Does that principle hold on a GPU, against the reference GPU engines?
DeepSeek releases under Apache-2.0 the engine that serves its models in production. No cache, no phase separation, no load balancing, for a single model and two chip families: an offline inference tool, not a service.
On October 21, PHP and Node.js fixed remote code execution vulnerabilities found by AI agents, Java and pip dropped the compromised eMudhra certificates, and LiteLLM was compromised on PyPI. Three days earlier, OpenCode revealed the theft of API keys for 1.4% of its users.
Ordered on October 21 to block sports piracy domains for French users only, the Swiss foundation behind 9.9.9.9 says it cannot do so and will stop serving France on November 24. Government agencies, banks, and universities must change their resolver.
On October 15, Meta publishes the weights of Muse Spark Open: 816 billion parameters, 42 of which are active, a context window of one million tokens, and support for text, image, audio, and video. Available on all platforms from day one, announced on two M5 Ultra Mac Studios, with no throughput figures published.
On October 18, Anthropic announces a public benefit fund endowed with 5% of its shares, up to 20% if the valuation rises. A 100% subsidiary, not a foundation, whose governance remains to be written before the initial public offering.
This letter went out by email to those who asked for it. All the letters