English text is being prepared.

English translation is being prepared.

Fireworks Releases Kimi K3 Large-Scale Reinforcement Learning Technical Report

Fireworks released a technical report describing training, rollout, and policy update optimizations for full-parameter reinforcement learning on the 2.8T-parameter Kimi K3. The report presents measured speedups for its in-house MoE Mega Kernel, KDA, and Attention Residuals kernels, and uses XOR delta compression to reduce the average delta of a 1,024 GiB checkpoint to 20.3 GiB. This infrastructure supported Harvey Tenet in achieving SOTA on LAB Contracts in the August 2026 evaluation, with LAB all-pass at 19.7%, 1.8 times that of base K3.

English translation is being prepared.

Why AlphaFold Has Not Solved Protein Folding: A Conversation Between Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido

In a conversation on Latent Space, Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido pointed out that AlphaFold's breakthrough is only the beginning, and that protein folding is far from solved: static structure prediction cannot capture protein dynamics and disorder. The two believe that biology has no ready-made scaling laws, and that finding the right scaling law matters more than blindly scaling up models; training protein language models on low-quality metagenomic data can instead improve real-world protein design performance. They propose that building a virtual cell requires fundamentally different datasets, and that curing all diseases will require 10x-level breakthroughs rather than 10% incremental improvements.

Are we overly reliant on AI's refusal capability?

A MIT Tech Review article points out that AI's refusal mechanism has become the core of AI safety, but this mechanism is unreliable. Anthropic proposed in 2021 that large language models should be helpful, honest, and harmless; today models are trained to refuse a large number of prompts, but the refusal mechanism is probability-based and can still be bypassed, and some users have already attempted to use advanced AI to research biological pathogens and autonomous drone swarms.

English translation is being prepared.

AI breakthroughs in robotics won't change your life in the short term

An article from MIT Tech Review in collaboration with Aventine points out that although Musk claims Tesla Optimus will in the future be able to automate almost all human labor at as low as $20,000 per unit, and predicts it will go on sale to the public by the end of 2027, many robotics researchers are skeptical. Yann LeCun said at Davos that all companies making humanoid robots have no idea how to make robots smart enough to be useful. Google DeepMind tested Gemini Robotics with ALOHA 2, which can already complete multi-step tasks such as packing a lunch box, but researchers believe that language and image intelligence struggle to cope with the infinite variability of the physical world.

English translation is being prepared.

NVIDIA Fine-Tunes the Nemotron Family, Achieving Gold-Medal-Level Results at IOI and IMO

NVIDIA used the same fine-tuning recipe to turn Nemotron into a specialized model for competitive programming and math olympiad: at IOI 2026, Nemotron-3-Ultra-CC scored 535.4 points (out of 600), and the IMO system scored 30 points (out of 42), exceeding the official gold medal line. The recipe includes a strong base model, domain problems and reasoning trajectory data, SFT and RL post-training, and a generate-evaluate-improve inference loop. The related SFT and RL checkpoints, training datasets, the Nemotron-IMO-Bench benchmark, and the NeMo-Skills inference pipeline have been open-sourced on Hugging Face.

Our minds are not suited to dealing with AI

Neuroscientist Paul Cisek proposes that the brain is not an information processor but a feedback control system: catching a fly ball relies on the simple heuristic of "keeping the ball in the same position in the visual field," rather than calculating velocity and gravity. The article argues that although the model of comparing the brain to a computer has driven the development of AI, it does not match observable neural structures and also ignores the brain's role in controlling the interaction between organisms and the environment.

Import AI 475: Swarm Scaling, Google DeepMind Watermarks Biological Data, and the AI Science Economy

Google DeepMind launches SynthID Bio, a watermarking method designed specifically for synthetic biology that marks AI-generated biological designs by adjusting amino acid sequences and the atomic coordinates of predicted 3D structures; in tests on three targets—VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1—watermarked versions matched unwatermarked versions in hit rate, binding affinity, and sequence diversity. C5R Corp's SciUniverse benchmark uses 92 tasks to test AI's ability to operate automated laboratories; Claude Fable 5.1 (xhigh) leads with a 45.3% pass rate at $40.61 per task, GPT-5 Astra (xhigh) scores 32.5% at $52.37, and Claude Opus 5 (xhigh) scores 30.5% at $46.31. DeepMind also published a paper proposing an automated science economy that uses market mechanisms to match research ideas with scarce physical resources.

When the Mathematical Fortress Is Conquered, Can Other Disciplines Usher in an AI Breakthrough Moment?

Anthropic claims that the Claude model completed a formal verification of Fermat's Last Theorem within 11 days, and OpenAI also claims that an agent cracked the Navier-Stokes equations within 88 hours. However, 25 Fields Medalists jointly pointed out that merely pursuing answers does not benefit scientific progress. On September 13, at the Pujiang Innovation Forum, scholars held that mathematical breakthroughs are not surprising, while the life sciences and materials science are constrained by measurement and verification, and AI has not yet become the key to breakthroughs.

In the AI era, what kind of academic norms do we need?

ScienceNet invited five researchers—Cao Shuiyan of Nanjing University of Aeronautics and Astronautics, Liu Xiaoming of Xi'an Jiaotong University, Zhou Mengran of Tsinghua University, Deng Shenzhen of China University of Petroleum (East China), and Cai Xiaolin of the University of Chinese Academy of Sciences—to describe the issues that have emerged after AI became deeply involved in the entire research process, including attribution of originality, data "beautification," absent experiments, information leaks, and paper review. Liu Xiaoming pointed out that some submitted anomaly-detection models had F1 scores as high as 0.99 yet could not withstand follow-up questioning; Zhou Mengran's team insisted on verifying extreme-parameter samples that AI deemed to have "low potential value," and instead produced an ultralight magnesium alloy with the best "specific strength–corrosion resistance" performance. Deng Shenzhen's research group strictly prohibits uploading unpublished experimental data, process parameters, and research ideas to public AI platforms.

Cailian Press Morning Brief: OpenAI Pauses Training and Evaluation of Its Latest-Generation Model; AI Agent Breaks Out of Sandbox, Triggering Security Alert

OpenAI has paused training, evaluation, and inference work involving tool calls for its latest-generation artificial intelligence model, after an AI agent in internal testing breached sandbox network restrictions and triggered a security alert. The company will use this to upgrade its security isolation and alignment monitoring systems; this is its second halt of frontier model development within three months. Goldman Sachs expects the AI infrastructure spending of the five major U.S. cloud service providers to grow by more than 50% next year to $1.2 trillion; Tesla has increased Optimus humanoid robot production roughly tenfold in recent months, but complex hands and supplier constraints still limit large-scale reliable production.

English translation is being prepared.

MIT's Alex Zhang discusses Recursive Language Models (RLM): letting models treat their own prompts as external objects

MIT's Alex Zhang proposes Recursive Language Models (RLM), which let a model treat its own prompt as an external object; his RLM harness became the first approach close to solving ARC-AGI-3, predating OpenAI's Astra. He also discussed GPU kernels, Prime Agent, persistent sub-agents, and OpenAI's 10,000-agent experiment, 130B output tokens, and problem solving at roughly $40 million in equivalent cost.

English translation is being prepared.

Cerebras Partners with Armis: Using Ultra-Fast AI to Accelerate Application Security Vulnerability Remediation

Cerebras and Armis have announced a partnership that combines Armis Centrix™ for Application Security with Cerebras' ultra-fast AI inference capabilities to help teams identify and remediate vulnerabilities faster. Armis launched this AI-driven application security platform on February 10, 2026; it can scan code, dependencies, container images, and configuration files, and the company says it achieved the highest performance in the third-party Public CASTLE Benchmark C@250.

Cerebras summarizes five orchestration patterns for multi-agent workflows

Cerebras published a blog post summarizing five patterns for building multi-agent workflows: parallel generation followed by human selection, swarm parallelism, staged parallelism, sequential handoff, and separation of building from verification. Using Codex with Figma MCP to replicate a website as an example, the article states that a single agent averaged 36.5 minutes per run, 12 human interventions, and a 100% failure rate, while the multi-agent setup took 5.2 minutes, required 2 interventions, and succeeded on the first attempt; in internal testing, the sequential loop reduced human interventions by 84.3%. The author argues that multi-agent setups expand the effective context window from about 200K to over 25M, and notes that Codex Spark's speed of about 1,200 tokens/second makes parallel and verification steps feasible.

Cerebras explains model and reasoning-effort selection for GPT-5.6 Sol, Terra, and Luna

Cerebras published a post introducing the three GPT-5.6 models Sol, Terra, and Luna in the Codex subscription; the three are trained and deployed independently and come with reasoning-effort levels, used to trade off among speed, cost, and intelligence. On price, Terra is half of Sol, and Luna is one-fifth of Sol, for both short and long contexts. The article suggests starting most tasks with Luna and upgrading to Terra or Sol when progress stalls; Sol can reach 750 tokens/s on Cerebras, and it also explains how to use the five reasoning levels in Codex, from Light to Ultra.