From the archive
Tuesday, 10 November
Folded from 23 sources that day.
Sorted by quality and by how many sources covered it, the same for everyone.
-
Gemini 4 Carbon is now available
Gemini 4 Carbon is now generally available via the API at $2 per million input and $10 per million output tokens. It ranks first on several agentic coding benchmarks, and the /swarm preview in Antigravity coordinates teams of agents.
-
Fable 5.2 retakes World #1 LLM spot as $ANT IPO roadshow begins
Anthropic released Claude Fable 5.2 and Mythos 5.2 at $8/$40 per million tokens, a 20% cut from 5.1. Independent tests show higher per-task costs because the model runs more checks, and it tops several leaderboards.
-
ExecuTorch WebGPU: PyTorch Models on the GPU, in the Browser
ExecuTorch's new WebGPU backend runs exported PyTorch models on the GPU in WebGPU browsers, with no format conversion. It is 4-7x faster than the CPU path on prefill.
-
Optimizing Muse Spark Open on TPU v7 with SGLang-TorchTPU
Meta's 816B-parameter Muse Spark Open runs on one four-chip TPU v7 host using a megakernel for all 62 layers. It halves time per output token at low concurrency.
-
Next-Level Benchmarking Overfitting
Argues that the claimed 22x to 61x speedups of the AI-generated GenDB-V engine over Umbra and DuckDB are infeasible, pointing to per-query precomputation and hard-coded data generator assumptions.