Mistral Large 4 is in public preview, with 1 trillion total parameters and 52 billion active parameters; its weights are planned for release at the end of this month. Amazon found that PHP2Shell exploitation attempts began four days before public disclosure. Additional proof-of-concept code appeared within hours of the patch release and was quickly used to attack its sensors. SuperMimic researchers report training motion control for 129,514 motions in three hours on a single GPU, using over 1,000 times fewer GPU-hours than the cited approach, though that reduction applies only to this training setup.
Mistral is using its own infrastructure for both training and preview serving, while Google is lowering the barriers to moving PyTorch workloads onto TPUs.
Mistral has launched a public preview of Large 4, a natively multimodal model with 1 trillion total parameters and 52 billion active parameters. Its largest model to date, it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in company-owned data centers in Europe, with preview serving running on the same infrastructure. The preview API is now available through Mistral Studio, and the weights are planned for release at the end of this month.
Read original →Google open-sourced TorchTPU at the 2026 PyTorch Conference, enabling native training, deployment, and inference on TPUs. Developed jointly with Meta's PyTorch team, the stack requires only minor code changes for existing workloads. Developers can debug in eager mode, then optimize performance with torch.compile. It already integrates with tools including vLLM and SGLang, and Transformers v5 can deploy models by setting device_map="tpu".
Read original →🔍 Analysis: Mistral Large 4 uses 3,800 NVIDIA Grace Blackwell GPUs across training and preview serving, while TorchTPU connects developers to another compute option through native PyTorch interfaces. They address two different constraints: control over infrastructure and software portability. For development teams, access to model weights and workload portability are separate questions: the former hinges on Mistral's planned weight release at month-end, while the latter depends on how well existing code and tools such as vLLM and SGLang actually work on TPUs.
Exploitation attempts preceded public disclosure, and attacks spread within hours of the patch release.
Amazon found attempts to exploit PHP2Shell in its MadPot honeypot logs as early as October 17, 2026—four days before public disclosure. Within hours of the October 21 patch release, additional proof-of-concept code appeared and was almost immediately used to attack its sensors. Amazon says the vulnerability is now being widely exploited, but whether the attempts observed in its honeypots would succeed against real systems remains uncertain.
Read original →SuperMimic reduces GPU-hours for motion-control training, while JOLT reduces the training tokens needed to reach comparable performance.
SuperMimic researchers report completing humanoid motion-control training covering 129,514 motions in three hours on a single GPU. The method combines a learned differentiable world model with backpropagation through time. Compared with a cited approach that used 128 GPUs for seven days, the researchers report using over 1,000 times fewer GPU-hours. This reduction applies to that motion-control training setup, not to the overall cost of robot learning.
Read original →Perplexity and NVIDIA introduced JOLT, which jointly trains a teacher and a student. The authors report matching GRPO's performance on MATH using roughly 1/6.5 as many training tokens. It trains the teacher with outcome rewards while penalizing excessive divergence from the student, addressing the problem of teachers exploiting shortcuts unavailable to students. JOLT+, which adds student outcome rewards, outperforms GRPO across five evaluations.
Read original →🔍 Analysis: SuperMimic connects a differentiable world model to backpropagation through time, while JOLT constrains divergence between teacher and student. Both change how learning signals are generated and passed along rather than simply adding training resources. For R&D teams, a reduction of more than 1,000-fold in GPU-hours and training with roughly 1/6.5 as many tokens do not belong on the same cost-reduction leaderboard: the former is specific to a motion-control setup, while the latter refers to matching GRPO's performance on MATH. Both must be compared alongside their tasks and results.
Claude Duet explores background execution, Uber expands employee access, and V7 organizes enterprise files with a Context Graph.
Anthropic has opened the research-preview waitlist for Claude Duet, powered by Opus 5.5. It inherits the memory, skills, and connected apps from a Claude account, checks apps every few hours, and continues tasks on a separate virtual machine without requiring the user's device to be on. The preview is included in Pro and Max plans and shares usage allowances with Claude; background checks also consume usage.
Read original →Uber is rolling out ChatGPT Work to its entire workforce, covering more than 30,000 employees. Teams can use it for research, data analysis, and creating presentations. The rollout also covers completing complex workflows, rather than being limited to individual content-generation tasks.
Read original →On V7's hardest graph-query test spanning thousands of documents, GPT-6 Astra achieved 89% accuracy. V7 Go organizes enterprise files into a Context Graph that agents can query, rather than storing them solely as standalone files. This structure is used to screen investment deals and support insurance underwriting.
Read original →🔍 Analysis: Uber is rolling out ChatGPT Work to more than 30,000 employees, extending its use to complex workflows. Claude Duet and V7 address two prerequisites for executing such workflows: continuing to run unattended and organizing the context needed for cross-document queries. For enterprises, rollout numbers are no substitute for validating task results. The 89% accuracy on V7's hardest graph-query test concerns result reliability, while Duet's background checks drawing from shared plan allowances make continuous operation a usage cost that needs to be accounted for.
Three sampling parameters have been removed, and even explicitly passing their default values triggers a 400 error.
Claude API version 2026-10-21 removes temperature, top_p, and top_k. Messages requests containing any of these parameters return a 400 error, even when their default values are supplied. The new major SDK version defaults to this API version, with a sampling temperature of 1.0. Requests that previously used a low temperature to improve output consistency lose that effect, with no replacement setting available. Requests that depend on non-default values can remain on 2023-06-01.
Read original →🔍 Analysis: The new SDK defaults to version 2026-10-21, making an SDK upgrade a change to both API compatibility and output behavior, not merely a dependency update. Even after developers remove temperature, top_p, and top_k to eliminate 400 errors, they cannot restore the output consistency previously provided by a low temperature. Request success and behavioral consistency therefore need to be validated separately.