AMD guides Q4 revenue to about $17B as Instinct MI450 ramps
AMD reported third-quarter revenue of $13.4B, up 45% year over year, with Data Center revenue of $8.5B (up 97%). It attributes the growth to EPYC demand and the launch of Instinct MI450 GPUs, and says its Helios rack-scale systems built on MI450 are running workloads at Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure. Guidance for Q4 is roughly $17B (plus or minus 3%), about 27% above Q3. It is the clearest sign yet that AMD is shipping frontier-scale inference hardware in volume, not only serving as a second source.
Sources & depth
Users report OpenAI's "improve the model" setting turning itself back on in Codex
A Hacker News post says a company that had disabled data sharing in Codex found the setting re-enabled without anyone opting back in. The evidence is anecdotal: several commenters report the same, others say their setting is still off, and the poster concedes it cannot be proven. OpenAI has not commented in the thread. Anyone using Codex with confidential code should check the data-controls page directly.
Sources & depth
A local Qwen 3.8 27B can read a game's purpose from a network's raw weights
Simon Willison reran a "weight-space programming" experiment on a laptop: given only the weights of small policy networks, Qwen 3.8 27B correctly named Breakout (88% confidence) and Pong (75%), gave a Pac-Man-style answer for Pac-Man, and missed Freeway. Each run took 52 to 65 minutes and tens of thousands of reasoning tokens. Kimi K3.1 Pro got Breakout in about ten minutes. He is skeptical of the claim that source code is becoming obsolete, and a separate critique argues the original "Astra can read and write weights" framing is misleading.
Sources & depth
Ollama 0.40.14 fixes Claude Code and Codex launch paths
ollama launch claude now applies each model's own output-token limit, so cloud models no longer stop at a 128k default. OpenAI-compatible endpoints keep a model's packaged sampler defaults when clients omit temperature or top_p, instead of forcing 1.0, which changes output for anyone relying on defaults. MLX top-k sampling is faster on large vocabularies. A 0.40.15 release candidate followed the next morning.
Sources & depth
Unsloth adds CPU fine-tuning and a faster single-GPU generation path
Unsloth's latest beta lets text models be fine-tuned on CPU when no GPU is available, adds options for combining parallelism strategies across several GPUs, and speeds up generation for supported Llama and Qwen3 models on NVIDIA Linux/WSL. The release reports a 4-bit Qwen3-8B going from about 26 to about 99 tokens/s on an A100 (a vendor figure, not independently measured). The rest is interface work: Glass themes, custom greetings and chat navigation.
Sources & depth
Google DeepMind shows an agent designing macrocycle drug candidates
DeepMind published Cheminformatics Skills for its Antigravity workbench, pairing a language model with RDKit and the ChEMBL and PubChem databases. In a lung-cancer case study on rigidifying crizotinib, the agent curated 568 compounds, built validated potency models and proposed new macrocycles with synthesis routes, in about an hour against days of manual work. The authors stress the designs are hypotheses needing lab validation. Applications are open for 10,000 scientists to get top-tier usage free for six months.
Sources & depth