Search “Reasoning”

Tencent Hunyuan launches open-source interactive video world model Hy-Nomas

Tencent Hunyuan launched the open-source interactive video world model Hy-Nomas on October 24, 2026. According to official release information, the model supports interaction using text prompts and camera trajectories, and has long-horizon generation capability with world memory; it can be adapted to specialized domains such as robotics via lightweight LoRA, and has been optimized for inference on consumer-grade GPUs. The official release also provides website, GitHub, Hugging Face, and demo entry points. There are currently no subsequent development facts.

Baseten launches Environment Network

Baseten launched Environment Network on October 24, 2026, bringing together rigorously vetted reinforcement learning (RL) environment companies to build custom RL environments for customers based on their own data and domains. According to the Baseten engineering blog, partners work directly with customers to complete the construction of environments, tasks, and scoring criteria, while Baseten remains responsible for model training and inference deployment. Baseten also plans to open-source a set of frontier-quality environments as well as vetting and training methods. There are currently no subsequent development facts.

Nathan Lambert says DeepSeek processes 150T tokens per day on a single cluster

On October 20, 2026, DeepSeek open-sourced its internal inference framework DeepInfer. According to Yuchen Jin, citing data disclosed by DeepSeek, a single cluster processes 155 trillion tokens per day and generates $38 in revenue per GPU-hour (off-peak pricing); DeepInfer supports both NVIDIA Blackwell and Huawei Ascend. A subsequent 36Kr report stated that its accompanying open-source deepseek-recipe supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. On October 21, swyx, citing Latent.Space data, said a single cluster processes 155T tokens per day with a 96% input cache hit rate, and argued that the lifetime value of each sold GPU has risen by about 30%. On October 24, Nathan Lambert said DeepSeek's single cluster processes 150T tokens per day, exceeding the combined open-source model inference volume of OpenRouter, Fireworks, Together, and BaseTen, and noted that the actual number of clusters is unknown; on the same day, LMSYS published a same-hardware benchmark comparison, stating that SGLang meets or exceeds DeepInfer on DeepInfer's own benchmark, with prefill throughput up to 1.04x and decode step length shortened by 1.1%–7.4% on 64x GB300.

View story:DeepSeek open-sources inference framework DeepInfer

Bardia Shahali says he can close an eight-figure deal within two weeks

On October 22, 2026, Bardia Shahali offered the judgment that a single deal at the AI infrastructure layer can reach tens of millions of dollars and can be completed within two weeks. The same report introduces that he once grew the sales team from about $3 million to about $400 million at Fireworks AI, and is now Vice President of Sales at Granola. The report also states that all inference providers are competing for the same set of customers, making competition fiercer; because model capabilities reach new heights every 45 days, compensation plans have been changed from annual to quarterly. At the infrastructure layer, FDEs, SEs, and AEs are paired to deliver complex products, while at the application layer, AEs and CSMs deliver software, but with higher productivity and speed of adoption. At present, this judgment appears only in this one report, with no other independent sources or subsequent developments.

Cohere announces public beta of Model Vault confidential computing

Cohere announced that the confidential computing feature of Model Vault has entered public beta. The feature is based on the NVIDIA H100 GPU confidential computing capability on Microsoft Azure, providing hardware-level isolation for hosted model inference. According to the Cohere product and research blog, the scope of encryption extends from data at rest and in transit to prompts, documents, responses, and model execution during inference; after application payloads are end-to-end encrypted, they can only be decrypted in a confidential environment that has passed remote attestation. Customers can cryptographically verify the hardware, firmware, and software configuration before sending sensitive data, and the client will refuse to send requests when the security policy is not met. The existing Model Vault API and request format remain unchanged. The feature is currently in public beta.

Cohere launches confidential computing public beta on Model Vault

On October 22, 2026, Cohere launched confidential computing on Model Vault, currently in public beta. The feature runs on NVIDIA GPUs on Microsoft Azure, combining confidential CPU and GPU execution, OHTTP end-to-end encryption of request content, and remote attestation, allowing customers to verify the target environment before releasing sensitive data. Cohere also plans to open-source the confidential inference stack and advance support for NVIDIA Blackwell and B200 GPUs.

Ant Ling Open-Sources Medical-Enhanced MoE Model Ling-3.0-flash-Sante

Ant Ling announced on October 22, 2026, the official open-sourcing of its medical-enhanced model Ling-3.0-flash-Sante. The model uses a MoE architecture with 124B total parameters, activating only 5.1B parameters per token. The company states that it leads models of the same scale in performance on professional medical tasks and complex medical reasoning. In addition to medical question answering, the model has also been strengthened for deep research, multi-agent collaboration, and tool calling, and the model weights are now available for download. There are currently no subsequent developments.

Matt vandenBerg advocates letting 17-year-olds change their minds

Matt vandenBerg, president of Ohio Wesleyan University, argues that in the AI era, 17-year-old students should not be forced to choose a major too early; instead, they should be allowed to explore and change direction. His basis includes an April study by the Lumina Foundation-Gallup: 47% of currently enrolled college students have seriously considered changing majors because of AI's impact on employment, and 16% have already changed majors. Accordingly, the university allows students to declare a major only in their sophomore year and sets quantitative reasoning, scientific inquiry, aesthetic thinking, listening across perspectives, and effective writing and expression as general education requirements. At present, this argument is presented in a Fortune AI report, and there are no subsequent developments.

Aditya Agarwal said the conversation at @spc was very rewarding

Aditya Agarwal said the conversation at @spc was great and that he learned a lot. He shared the content of Sergey Levine's conversation at SPC: in his early years, Levine had a roomful of robots grasping objects all day in a converted Google conference room, accumulating a great deal of experience but lacking prior knowledge, and this bottleneck ultimately gave rise to Physical Intelligence. The conversation also covered the Pi Zero robot foundation model, the evolution of reasoning and generalization capabilities from Pi 0.5 to Pi 0.7, as well as the robot data challenge and world models. This content was published by Aditya Agarwal on October 23, 2026, and there are no subsequent developments as of yet.

Google DeepMind releases Gemini 4 Argon

On October 23, 2026, Google DeepMind released the frontier model Gemini 4 Argon. Google's official blog stated that the model performs outstandingly on complex tasks such as coding, mathematics, and 3D world building, and surpasses other frontier models on benchmarks across multiple domains; the company also showcased five community creative projects, including a 3D Earth that maps wildfires and seismic waves in real time, a turbofan engine simulation with live data, a 3D horse-drawn carriage generated from a single image, and a multiplayer capture-the-flag browser game, and said the model had already been made broadly available last week. On the same day, another report stated that Argon can learn physical tasks by watching videos and transfer them to Apptronik's Apollo humanoid robot, with Apollo hitting seven consecutive shots in internal tests; the report also said Argon scored 87% on the embodied reasoning benchmark ERQA. On October 24, the Gemini for Android app pushed a redesign of its sidebar and settings interface via a server-side update, and reports said the redesign had already arrived on iOS, macOS, and Web alongside Gemini 4 Argon last week.

UK AISI evaluates frontier AI agents in a simulated cyber range

The UK AI Security Institute evaluated the multi-step attack capabilities of frontier AI agents on a self-built cyber range, assessing seven models released between August 2024 and February 2026. The institute said that under a fixed token budget, each generation of models outperformed the previous one: in a 32-step enterprise network attack range, with a 10M token budget, the average number of completed steps rose from 1.7 steps for GPT-4o to 9.8 steps for Opus 4.6. Increasing inference compute from 10M to 100M tokens brought gains of up to 59% with no plateau observed; at current prices, a single 100M token attempt with Opus 4.6 costs about 80 USD. In a 7-step industrial control system range, model progress was limited: at 10M tokens most models completed zero steps, and at 100M tokens Opus 4.6 averaged 1.4 steps.

Shanghai releases action plan for AI-empowered industrial services

On October 23, 2026, the Shanghai Municipal Commission of Economy and Informatization released the "Shanghai Action Plan for Accelerating the High-Quality Development of AI-Empowered Industrial Services (2027-2028)." The plan proposes that by 2028, the penetration rate of AI applications in industrial services reach over 85%, and strives to cultivate 10 digital-intelligent native unicorn enterprises. The plan deploys full-chain upgrading around areas such as R&D and design, supply chain services, flexible manufacturing, embedded quality inspection, and predictive operations and maintenance, and provides funding support ranging from up to 5 million to 20 million yuan. The plan also clearly specifies directions for tackling industrial agent task planning and swarm collaboration, device-cloud collaborative inference and lightweight vertical small models, and industry high-quality corpus libraries. Other reports mention that the plan proposes building 5 industry-level collaborative design platforms, promoting 500 high-quality enterprises to build Shanghai advanced-level smart factories, and providing subsidies of up to 5 million yuan for enterprises calling third-party large model APIs or private deployment. Currently, the plan has been released by the Shanghai Municipal Commission of Economy and Informatization, and there is no qualified factual evidence of further implementation progress.

Dongxi NLP says Agent-building ability is the least important

On October 23, 2026, the blogger "Ma Dongxi NLP" published a post arguing that in "Agent +" scenarios, the ability to build Agents is the least important, and what matters more is familiarity with the business of specific scenarios. The same article also said that in the current context, an AI engineer has little to do with AI and is not valuable, and advised young people in the AI field to turn to training, quantization, inference, or AI infrastructure construction. At present, this view is seen only in this one source, and there are no other development facts.

OpenAI fires 3 safety researchers after leak investigation

According to a WSJ report, OpenAI fired 3 safety researchers after a leak investigation. The report says safety staff discovered disappearing Signal chats between these 3 safety researchers and an outside safety researcher. The phone checks began with a September news report that said OpenAI's latest model Astra uses a new approach that makes it smarter but writes down less of its thought process, so humans can see less of the reasoning. OpenAI treated the report as a leak of its secret model design and began tracing the source of the information; investigators listed Ryan Greenblatt, who had helped METR analyze logs from the July Hugging Face hack, as a suspected leaker, and he denies it. Currently the public information comes only from this report, with no further developments yet.

OpenAI officially opens the Decisions API

English text is being prepared.

Modal releases open-source Tinker-compatible training server Spindle

On October 22, 2026, Modal released the open-source training server Spindle, which is compatible with the Tinker API. According to Modal's official engineering blog, Spindle can run in one's own Modal workspace, is billed by GPU time rather than tokens, requires only executing spindle deploy to deploy, supports full-parameter training and multiple LoRAs, and covers a variety of open-source models. In the comparison provided by the official source, for the DAPO-math single-turn task with 8-card H200 training, 8-card H200 inference, and 8 clients sharing, the price is 1.99 USD per million output tokens, 80% lower than Tinker. There are currently no subsequent development facts.

LMSYS releases SGLang AWS Trainium plugin Beta

On October 22, 2026, LMSYS partnered with AWS to release the SGLang Trainium plugin Beta. Built on the newly released TorchNeuron, the plugin supports LLM inference serving and RL rollout generation on Trainium2 and Trainium3. Initial coverage includes Qwen3 (dense), Qwen3-MoE-30B-A3B, DeepSeek-V2-Lite, GLM-4, and GLM-5.2/5.3, and it supports data parallelism, tensor parallelism, expert parallelism, Radix Cache, BF16/FP8, and asynchronous scheduling. In the first end-to-end heterogeneous RL run, SGLang generated rollouts for Qwen3-MoE-30B-A3B on Trainium3, while Miles trained a GRPO policy with Megatron on a single H200 node. The Beta is currently limited to a small number of teams. The next day, Amazon announced the open-sourcing of TorchNeuron, a backend that converts PyTorch code into operations executable on AWS Trainium, allowing PyTorch code already running on other accelerators to migrate with almost no changes.

Railway requires all customers to rotate all keys

Developer platform Railway required all customers to rotate all keys after intruders decrypted some stored credentials. The affected environment variables include database passwords, cloud keys, and model API keys; holders could log in as owners. Railway said only some users may be affected, but it froze all deployments on its own and reset bucket keys and GitHub tokens. On October 21, Railway discovered anomalous activity in the system storing customer environment variables, saying an unauthorized third party may have accessed encrypted environment variables and decrypted some of them. Railway believes the attack may be part of a recent series of attacks targeting platform companies with the aim of compromising AI inference providers, and has notified upstream vendors to revoke potentially leaked keys. On October 23, its builds and deployments were restored, but new user registration remains suspended.

Google open-sources TorchTPU native PyTorch stack

Google collaborated with Meta's PyTorch team to launch the open-source project TorchTPU, providing a native PyTorch backend for Google TPU; migration only requires changing model.to("cuda") to model.to("tpu"). The project supports both eager and torch.compile modes, adds a new DeferAndFuse eager mode, covers about 450 Aten operators, and has been verified to run out of the box on 4,917 Hugging Face models, with a Tier 1 pass rate of 98% and Tier 2 numerical consistency relative to GPU of 97%. The official statement says it has been used in the inference and training production workloads of some Google Cloud major customers, and is integrated with vLLM, SGLang, TorchTitan, Ray, and TorchRec. Subsequent reports said that at PyTorch Conference North America 2026, Google open-sourced TorchTPU, whose runtime no longer depends on JAX, and existing workloads can be migrated with only minor code changes; PyTorch noted that TorchTPU's default Skinny Collectives scheme suffers roughly 30–40% end-to-end performance loss under distributed workloads compared with optimized schemes, and achieving peak performance still requires Pallas kernels and vLLM-T

Bill Gates says people need to prepare for the two decades of disruptive change brought by AI

Bill Gates said people need to prepare for the two decades of disruptive change that AI will bring. In an interview on The Verge's Decoder podcast, he said this roughly 20-year period of AI disruption could be very painful and might even split society into two parts: those close to the technological frontier and those who reject it. He noted that over the past year, AI has made astonishing progress in writing code, discovering network vulnerabilities, reasoning in biology, and disrupting the job market, and that the transition period will have losers and winners before a prosperous era arrives. He also said that in an article in late August he had already stated that even under the best-case scenario, the transition to the new AI era will be one of the most turbulent periods in human history, and he believes that current policy responses and industry coordination are insufficient to address the risks; the biggest near-term risk comes from models whose safeguards can all be removed.