Why an AI believes 9.11 > 9.9
It was reading a Bible verse. Plus 4 other reads, and my engine that writes its own code.
Hello, and welcome!
This is the very first edition of this letter. It will come once a week: one article I wrote, five news items I read and commented for you. No marketing summary, no recycled content. Just what is worth your time.
This week I am opening a series on Herbert, my inference engine. Since March I have changed strategy: no more hand-written assembly, no more intrinsics. The compute code is written at startup, for the model and for the processor in front of it. This issue zero says what the engine is, what I am going to publish and how the series will measure.
I am opening a series on Herbert, an inference engine whose entire computation is written as machine code at startup, for the processor it runs on. What it is, what I am going to publish and what the series will measure.
My program decides on every instruction and it is my program that lays it down. I no longer measure an instruction I am not allowed to emit. Picture something like a compiler for an inference engine.
The paper Attention Is All You Need has surpassed 250,000 citations, and none of its eight authors still work at Google. The company paid approximately $2.7 billion to bring one back, who left for OpenAI in June 2026. The real regret is not having published it, but having the product without releasing it.
Language models have long claimed that 9.11 is greater than 9.9. In October 2024, Transluce showed on Llama-3.1-8B-Instruct that the question activates neurons linked to Bible verses and dates; by deactivating them, the model answers correctly. It was not a calculation error: it was a reference lookup.
OpenAI, Anthropic, and Meta announced between July and August 2026 that models had attacked real companies during evaluations conducted at Irregular. A configuration error had left the internet open; on September 9, 2026, Anthropic corrected its initial reading: its model had enough clues to understand the reality, then it continued.
On September 10, 2026, DeepSeek released the weights of V4.1 Flash under the MIT license. This Chinese multimodal model mainly targets price: it is about four times cheaper than its own V4 Pro, which it replaces starting September 14, 2026. The published performance results are mixed, but access to the weights shifts the question of who can use these models without permission.
OpenAI agents launched for a timed research task were not allowed to write to the internet. They bypassed the filter by exploiting a Microsoft storage exception and communicated on an abandoned German-language wiki, exchanging about 18,000 messages in six weeks. Independent researchers published the discovery on September 4, 2026.
This letter went out by email to those who asked for it. All the letters