From the archive
Friday, 16 October
Folded from 14 sources that day.
Sorted by quality and by how many sources covered it, the same for everyone.
-
Beam: a 501B open-weight MoE model for coding and agents
Reflection's first model, Beam, is a 501B-parameter mixture-of-experts model with 23B active per token, a 1M-token context and Apache 2.0 licensing. It has day-0 support in Transformers, vLLM and SGLang, and Reflection offers $10,000 in Hub credits for projects.
-
The Decoherence Trap: When AI slop drowns out the signal in an endless sea of noise
Introduces 'decoherence', the buildup of locally valid but mutually incompatible models in a system. The essay argues AI-driven output is producing it faster than it can be absorbed, risking a 'Decoherence Collapse'.
-
Is Your Organization Messing Up AI Adoption?
The author argues that organizations roll out AI as a technology project and ignore the people side, such as staff who fear for their jobs and gain little personally. He draws parallels with the digital transformation failures described in his 2014 book.
-
Why are we even doing all this?
A satirical guest post on Terence Tao's blog questions the value of mathematical research after OpenAI's large volume of claimed results. It warns that AI output could overwhelm human work and expertise.
-
Open-weight models trail closed models in cyber capabilities by 6–7 months
Z.ai's GLM-5.3 is the top open-weight model on Epoch's Cyber ECI, above GPT-5.4 but behind Claude Mythos Preview and GPT-5.5. The closed frontier now leads at 171.