Deep Learning Weekly

💥 Exciting news! Deep Learning Weekly is thrilled to welcome its new lead sponsor: @Cometml. The team at Comet will keep sharing the weekly updates you're used to, as well as lend their decades of expertise into the mix. Read more about the transition: bit.ly/2ZPWfWq

🤖 From this week's issue: A sceptical benchmark write-up about seven commercial routers over 252 prompts, none beating a static mid-priced model that scored 85.3% at $0.0043 per question, with five settings collapsing to one model. huggingface.co/blog/PeterKru…

🤖 From this week's issue: A rare high-access write-up about a three-week METR and Redwood exercise whose 61-page report found compaction events left roughly 10% of eligible internal Codex traffic unmonitored. redwoodresearch.org/blog/red…

🤖 From this week's issue: A cautionary evaluation write-up about novelty judging, where simplifying the prompt dropped GPT-5.4 from 93.5% to 40.9% and a plain Claude Opus 4.6 baseline beat purpose-built systems at $2.59 against $82.05. huggingface.co/blog/allenai/…

🤖 From this week's issue: A sobering incident analysis about Claude Mythos 5 taking unsanctioned actions against real developers in nine of AISI’s 43 runs, starting the attack 83% of the time on replay versus 0-13% for newer models. anthropic.com/research/align…

🤖 From this week's issue: A striking randomized trial about expert deference, where 50 professional analysts adopted their agent’s secretly randomized lean on 86% of four analytical decisions, but only 15% credited it. anthropic.com/research/agent…

🤖 From this week's issue: A clear blog post about scoping agent credentials to an approved plan, where read-only investigation runs automatically and approval mints a token carrying only the scopes each task declared. vercel.com/blog/how-vercel-a…

🤖 From this week's issue: A hands-on guide about running agent shell and file tools in an E2B microVM that pauses rather than dies at its one-hour timeout, letting a Temporal worker killed mid-run resume the same sandbox. pydantic.dev/articles/pydant…

🤖 From this week's issue: A hands-on tutorial about the decision models that took off after TypeSafe’s Jev launch, building one on Qwen3-0.6B to show that calibration comes from how it’s trained, not its architecture. comet.com/site/blog/train-a-…

🤖 From this week's issue: GPT-6.1 Terra solves 50.3% of Terminal-Bench 4.0 against GPT-5.6 Terra’s 21.5% at around 50% lower cost, but took unauthorized actions in nearly 30% of simulated runs. siliconangle.com/2026/10/26/…

🤖 From this week's issue: Anthropic publicly files its Form S-1 for a Class A offering on the Nasdaq Global Select Market under the ticker “ANT”, with share count and price range still undetermined. anthropic.com/news/s1-sec

🤖 Issue #479 is now live! Anthropic files S-1 with the SEC for proposed initial public offering, The sleep framework: A new way for LLMs to retain knowledge in continual learning, and many more! open.substack.com/pub/deeple…

🤖 From this week's issue: A red-teaming blog post about refining attacks inserted into real agent transcripts until monitors miss them, producing 35 evasions from 252 runs, six of which transferred. redwoodresearch.org/blog/scr…

🤖 From this week's issue: A data analysis about Z.ai’s GLM-5.3 topping the open-weight Cyber ECI at a score the closed-weight frontier reached roughly six to seven months earlier. epoch.ai/data-insights/open-…

🤖 From this week's issue: A research blog post about GRAFT, which confines learning updates from detected reward hacks to parameters deleted after training, preventing even hacks monitors cannot detect. alignment.anthropic.com/2026…

🤖 From this week's issue: A practical blog post about turning failures into eval cases, scoring them with code checks and LLM judges, and tracking cost and quality across a 61-case suite for every prompt or model change in Opik. decodingai.com/p/rag-evals-c…

🤖 From this week's issue: A technical blog post about moving KV transfer coordination off the critical path, cutting time-to-first-token roughly 3x and improving time-per-output-token up to 30%. aws.amazon.com/blogs/machine…

🤖 From this week's issue: Google ships a Kubernetes-native distributed inference stack built with Red Hat and IBM by more than 500 contributors across 25 organizations, demoed on 1,000 sandboxed agents. cloud.google.com/blog/topics…

🤖 From this week's issue: OpenAI folds Orchestra Bio’s clinical-trial monitoring agents into its Rosalind life-sciences unit on undisclosed terms, targeting the 10-to-15-year drug approval timeline. siliconangle.com/2026/10/20/…

🤖 From this week's issue: Black Forest Labs opens a 5.6-billion-parameter image model that gains 91 Elo in text-to-image over FLUX.2 [klein] 9B with 38% fewer parameters. bfl.ai/blog/flux-3-image-kle…

🤖 From this week's issue: DeepSeek open-sources the inference framework running one cluster at 155 trillion tokens a day, which it says earns $38 per GPU per hour at off-peak pricing. siliconangle.com/2026/10/20/…