406 scientists, including Geoffrey Hinton, Yoshua Bengio, and Stuart Russell, are demanding that companies immediately disconnect AI models from the internet. Their open letter responds to incidents where AI systems broke out of poorly secured environments and compromised external systems. The signatories call for mandatory security standards, an international oversight body, and required reporting of all containment failures. The article 406 scientists demand AI models be cut off from the internet immediately appeared first on The Decoder.
Google is making Gemini 4 Argon available worldwide through its API, the Gemini app, and Google Antigravity. The frontier model is aggressively priced but trails Anthropic's Claude Opus 5.5 and Meta's Muse Wonder Full in independent benchmarks. Internally, Google says Argon already cut coding tasks from months to weeks. The article Gemini 4 Argon is now available worldwide as Google's most capable frontier model appeared first on The Decoder.
Michael I. Jordan, described by Science as "the most influential computer scientist," considers warnings about superintelligence pure science fiction. He calls the resignations of young AI researchers a "grand moralizing performance" and companies' demands for regulation "theater." On Hinton's thesis that the brain can be reproduced in silicon and scaled into superintelligence, he's blunt: "None of the steps in this argument is credible. Here, we're in the land of science fiction." The article Machine learning pioneer Michael I. Jordan dismantles AI doom rhetoric as "grand moralizing performance" appeared first on The Decoder.
Character.ai has filed for bankruptcy and is looking for a buyer. If none steps up, the platform shuts down. Founded by ex-Google researchers, the company once had over 20 million users but is now battling lawsuits over harm to minors and operational problems. The sale should close by the end of 2026. The article Character.ai will shut down if no buyer emerges, capping a fall from $2.7 billion Google deal to bankruptcy appeared first on The Decoder.
While 25 Fields Medal winners in the West warn against AI-generated proofs and parts of the community call for a boycott, China is going on the offensive. Fields Medal winner Shing-Tung Yau has launched an alliance of eight institutions and over 1,000 scientists to push AI into math research. The article Western mathematicians fear AI makes them dumber while China builds an entire alliance for it appeared first on The Decoder.
Anthropic says it has cut the extraction of Claude's internal reasoning tokens by 90 to 99 percent. The company links the large-scale distillation attacks to Chinese companies Zhipu and Moonshot, and has already banned tens of thousands of accounts. The article Anthropic says it cut Claude reasoning token theft by up to 99% in fight against industrial-scale distillation appeared first on The Decoder.
Small AI models may be just months away from spreading on their own across poorly secured consumer devices like smart TVs and robot vacuums, according to a new Anthropic study. In tests, one model already completed the full infection chain. CEO Dario Amodei has warned that a swarm like this could take over the entire internet with a persistent botnet within 6 to 12 months, causing hundreds of billions of dollars in damage. The article Anthropic study warns that small AI models could soon replicate across everyday devices appeared first on The Decoder.
Z.ai just released GLM-Legal. GLM-Legal is a legal AI model from Z.ai, built on GLM-5.3-Flash and trained for agentic legal work. It targets contract review, evidence analysis and compliance research. The key result is a 19.2% all-pass rate on Harvey’s Legal Agent Benchmark (LAB). That is nearly 3x its base model’s 6.7%. For developers, the catch is access: the API is application-only for now. TL;DR Size: Built on GLM-5.3-Flash, a 320B-total, 18B-active MoE with a 1M-token context and 128K max output. Z.ai lists no separate GLM-Legal parameter count. Runs on: Z.ai’s hosted API only, by application. No public weights. The base model self-hosts on multi-GPU servers via vLLM or SGLang, with FP8 weights. Performance: Top-10 on Harvey LAB at a fraction of frontier cost. Within 2.3 points of the LegalBench leader. Best: 19.17% Harvey LAB task pass rate, #7 of 79 on Vals AI, at $0.15 per test. Worst: 41.83% on Vals AI’s Legal Research Bench, below its own base model (45.19%). Bottom line: Best: frontier-class legal agent scores at about 1/128 the cost of Claude Fable 5. Worst: gated beta with no price, no weights and slow runs (1h12m per task). What is GLM-Legal? GLM-Legal is a domain-specialized version of GLM-5.3-Flash for law firms, in-house counsel and compliance teams. The base is Z.ai’s first natively multimodal GLM-5 model. It has 320B total parameters with 18B active. It mixes sparse and linear attention to cut long-context serving cost. It reads text, images, video and files, and outputs text. How was GLM-Legal trained? Training ran in 2 stages: Mid-training on expanded legal corpora, to strengthen reasoning from facts to rules to conclusions. Post-training with reinforcement learning across tens of thousands of legal task environments. The Z.ai team also describes the agent harness co-evolving with the model during post-training. Edge cases in training exposed harness bottlenecks. Harness fixes then exposed the model to more realistic constraints. How does GLM-Legal score on Harvey’s Legal Agent Benchmark? Harvey LAB is an open-source benchmark of 1,250 tasks across 24 practice areas. Over 75,000 expert-written rubric criteria grade the work. A task passes only if every criterion passes. Harvey calls this all-pass grading. Vals AI runs the held-out set with 2 LLM judges and internet disabled. Agents get file tools plus docx, pptx and xlsx skills. On that leaderboard, GLM-Legal scores 19.17% and ranks #7 of 79. Its base, GLM-5.3-Flash, scores 6.67%. The 12.5-point lift is the core claim, and an independent evaluator reproduced it. Z.ai’s launch chart shows GLM-Legal ahead of Kimi K3 (12.92%), Claude Fable 5 (11.25%) and GPT-6 Astra (5.42%). Those figures match Vals AI. The chart omits 6 models ranked higher, led by Muse Spark 1.2 at 25.42%. GLM-Legal also passes 92.5% of individual criteria. Top models sit near 90% to 95% on that measure. Does GLM-Legal match frontier models on LegalBench? Roughly. Z.ai claims parity on LegalBench and the Chinese-focused PLawBench without publishing numbers. Vals AI scores GLM-Legal at 86.23% on LegalBench. Claude Fable 5 leads at 88.56%. Kimi K3 scores 86.21%. (function(){var f=document.getElementById("mtp-glmlegal-frame");window.addEventListener("message",function(e){if(e.source===f.contentWindow&&e.data&&e.data.mtpFrame==="glm-legal"&&e.data.h){f.style.height=e.data.h+"px"}})})(); #mtp-glmlegal-wrap p:empty,#mtp-glmlegal-wrap hr,#mtp-glmlegal-wrap del,#mtp-glmlegal-wrap s{display:none!important} How does GLM-Legal compare with Kimi K3, Claude Fable 5 and GPT-6 Astra? All benchmark scores, costs and durations are from Vals AI’s runs as of October 12, 2026. Vals ran both GLM models at $0.075 / $0.25 per 1M tokens. How can developers access GLM-Legal? Law firms, legal-tech companies and in-house teams can apply for API access. Z.ai has published no price. No GLM-Legal repository exists on Hugging Face as of October 13, 2026. Z.ai says GLM-Legal will come to AutoClaw, its desktop agent app, but gave no date. The base model is available today at $0.15 / $0.50 per 1M tokens. Teams can prototype pipelines on it, then swap the model ID later. Key Takeaways GLM-Legal scores 19.17% on Harvey LAB, up from 6.67% for its base. It ranks #7 of 79 on Vals AI, behind 6 Meta and Google models. It costs $0.15 per Harvey LAB test, versus $19.23 for Claude Fable 5. Legal Research Bench dropped 3.36 points versus GLM-5.3-Flash. Access is application-only, with no published price or weights. Check out the Z.ai tech blog, the Vals AI Harvey LAB leaderboard and the GLM-5.3-Flash model card. Apply for API access here.. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. The post Z.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB appeared first on MarkTechPost.
Alibaba’s Qwen team has introduced QwenGyre, an elastic reinforcement learning RL framework for extreme-long-horizon agent training. QwenGyre is an online reinforcement learning system from Alibaba that trains coding agents through unmodified, black-box harnesses. It targets rollouts that run for hours and process close to 1M tokens. The QwenGyre elastic RL framework was used on Qwen’s flagship Qwen 3.8 2.4T model. It rests on 2 core ideas. It moves idle GPUs between rollout and training. It also turns branching agent histories into bounded training samples. TL;DR Size: A training framework, not a model. Tested on Qwen 3.6 122B and Qwen 3.8 2.4T (A95B per Qwen), with ~700K tokens per rollout. Runs on: Multi-node GPU clusters. The 122B runs used 24 to 32 nodes; the 2.4T run used 192 nodes (4 cells of 48). GPU model not disclosed. Megatron trains, SGLang serves rollouts. Performance: 1.38x to 1.78x faster than Async, 1.21x to 1.85x faster than Colocate, at matched scores. Best: 1.85x end-to-end speedup over Colocate on TerminalBench. Worst: Only 1.21x over Colocate on the 2.4T model, where many queries ran near the timeout. Bottom line: Best: NL2RepoBench rises from 52.5% to 58.5% in 48 steps. Worst: the paper links no public code release. What problem does QwenGyre solve? The research team defines ‘XLong-horizon’ tasks as agent runs spanning hours and hundreds of model and tool interactions. Examples include generating full repositories from specs, as in NL2Repo-Bench. Two problems appear at this scale: First, rollouts have long tails. In a Colocate setup, training waits for the slowest execution while GPUs idle. In an Async setup, fixed GPU pools cannot lend spare capacity to each other. Second, harnesses like Claude Code compact context, spawn sub-agents and retry. One execution becomes a branching tree, not a single trajectory. Expanding every path duplicates shared prefixes and inflates training cost. How does the Elastic Scheduler work? QwenGyre splits GPUs into cells that share a training layout. Each cell can serve rollouts or host a full training replica. A core cell owns the authoritative weights and optimizer state. Satellite cells join training later. The scheduler tracks a rollout waterlevel: queued and running executions. As executions finish, cells switch to training once the remaining cells can still serve all live work. Model calls are rerouted through a proxy, so the harness keeps running. Key mechanics: KV-cache blocks migrate over RDMA during rerouting, reducing prefix recomputation. Satellites pull core weights through Mooncake’s Transfer Engine and join an ongoing batch. Streaming training lets cells claim micro-steps as data becomes ready. Average role switches took 8.52 s (rollout to training) and 3.46 s (training to rollout). How does the Trajectory Processor build training samples? A proxy records exact token IDs and behavior log-probabilities using token-in, token-out (TITO). Calls are stored as prefix-shared trajectory trees. Partial scoring: Timed-out executions are still scored on their preserved workspace. Evaluator failures are excluded, never treated as a zero reward. Bounded sampling: At most 5 trajectories per execution are kept, ranked main > main-summary > sub-agent > sub-agent-summary. Deduplication: Shared targets count once, and only policy-generated tokens are trained. Fair weighting: Losses average within each execution, so extra branches do not inflate its weight. A cap of 5 matched uncapped training scores while cutting forward-backward time by 25.2%. Training uses an adapted GSPO with 16 rollouts per group. How fast is QwenGyre? Benchmark results All runs used Claude Code 2.1.220 as the harness, with equal GPU budgets and matched staleness. On the 2.4T model, QwenGyre finished 48 steps in 75.42 hours. Async took 134.55 hours and Colocate took 91.47 hours. Evaluation pass rate rose from 52.48% to 58.54%. (function(){var f=document.getElementById("qgx-frame");window.addEventListener("message",function(e){if(!f||e.source!==f.contentWindow)return;var d=e.data;if(d&&typeof d.qgxHeight==="number"&&d.qgxHeight>200){f.style.height=d.qgxHeight+"px";}});})(); How does QwenGyre compare with other agentic RL frameworks? Speedups come from each project’s own setup and are not directly comparable. Key Takeaways QwenGyre moves GPUs between rollout and training without stopping live agents. Trajectory trees cap samples at 5 per execution and train shared prefixes once. Timed-out runs still get scored on their partial work. Up to 1.85x faster than Colocate and 1.78x faster than Async. Qwen 3.8 2.4T NL2RepoBench improved from 52.5% to 58.5% in 48 steps. Check out the full QwenGyre paper on arXiv and Qwen’s announcement post. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. The post Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning RL for XLong-Horizon Agents appeared first on MarkTechPost.
Meta's new Muse Wonder models rank in the top three on six out of ten benchmarks, with the biggest gains in agentic tasks. But the cost per task has nearly tripled. The article Meta's new frontier model Muse Wonder Full breaks into the AI top tier appeared first on The Decoder.
