TabLOID

The Decoder, citing The Information, reports that Meta and Microsoft are trying to reduce employees' reliance on Anthropic's Claude. The reported reasons include costs and competition as Anthropic expands into workplace software. We have not independently verified the internal budgets or headcounts, so those precise figures stay off this page.

The useful question for builders: how dependent is your product on one supplier? A second model option and a clear budget are less glamorous than a launch video, but considerably more useful when the invoice arrives. Read the reporting.

Researchers introduce Queen, a four-billion-parameter model that connects a chess expert encoder to a language model. Their paper reports an estimated playing-strength rise from 1782 to 2697 Elo over seven improvement rounds, alongside explanations of moves and plans. That is a research evaluation, not an official tournament rating.

The interesting part is the partnership: a specialized system supplies the expertise; a language model makes it understandable. The authors suggest this pattern could extend to other domains. Your next chess opponent might finally explain the humiliation. Read the paper.

OpenAI says API customers can opt into text watermarking for selected models, while eligible ChatGPT and Codex output in the EU will receive watermarks over the coming weeks. Its textGrain system puts a statistical signal into word choices. Initial detector access is limited to approved researchers and expert organizations.

A detected watermark does not identify the user or establish whether the text is accurate. An absent watermark does not prove human authorship. Translation, editing and short passages complicate detection. In other words, the name tag is not a lie detector. Read the announcement.

Reflection describes Beam as a mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, aimed at coding, reasoning and agentic work. The company says it is completing evaluations and red-teaming, with weights, a technical report and developer artifacts planned for later in October.

Treat the announcement's performance and efficiency comparisons as company claims until outsiders can test the released model. For now, it is an early-access signup, not a model you can already download and run. The robots can queue like the rest of us. Read Reflection's announcement.

Quanta explains how holographic duality can connect a gravity theory in one space with a quantum theory on its boundary. This has given physicists powerful ways to investigate spacetime and information. It does not mean a laboratory has proved our universe is a projection.

The best-developed examples use special theoretical settings, and translating them to our universe remains an open problem. Still: a universe whose physics can be described on a boundary is a much better conversation starter than asking someone what they do for work. Read Quanta's explainer.