Meta says GEM trains at 20-25% model FLOPs utilization end to end. The Kunlun paper it links measures 37% on a single NVIDIA B200, a chip the blog never names.

Nvidia’s launch post for Nemotron 3.5 Bolt 74B calls it “the #1 open model in its class on PinchBench, and

Bank of America’s “couple hundred million dollars plus” of third quarter remediation is a combined AML and Mythos line, and the transcript never splits it.

Google removed an auxiliary reward that was scoring Gemini 4 Argon chain-of-thought text, and its full safety report says monitorability is now harder than for past models.

White Circle’s own research page carries two numbers that the coverage of its nudity jailbreak study left out. Running its

Apple published SelfCompact on 12 October, a rubric-gated scaffold that lets an agent decide when to compact its own context. It beats fixed-interval compaction on accuracy for all three search agents tested, but the cost saving in its tables is measured against doing no compaction at all.

Apple ran 16 language models as both writer and reader of the same arithmetic word problems. The best pairing hit 92.9%, but the 12 open-weight models averaged 9.4%.

OpenAI put ChatGPT Space on every Pro, Business and Enterprise plan. The dot its demo placed inside that space ships to fewer, and on Enterprise it is off by default.