Nat Lambert: GLM-5.2 Is the 'DeepSeek Moment' for Agents
Top-tier agentic capabilities arrive in open-weight models. The researcher urges regulators to engage now.
Researcher Nat Lambert declared that GLM-5.2 represents a turning point for the agent landscape — the moment at which the highest tier of autonomous reasoning crosses into freely available open-weight models. "If you care about open, now is the time to inform regulators on how we should build a world with safe, frontier, open intelligence," Lambert wrote. Independent benchmarks place GLM-5.2 third overall on GDPval-AA, a real-world agentic work evaluation, where it leads all open-weight competitors and rivals closed frontier systems.
Altman Details Cyber Strategy as Codex Security Goes Live
The full GPT-5.5-Cyber model is now operational with state-of-the-art CyberGym scores. Altman emphasized the shift from detection to automated remediation: "Patch The Planet and Codex Security will help solve security problems instead of just finding them."
Perplexity CEO: GLM-5.2 Revives Open-Source AI, Beats Frontier Models Blind
Arav Srinivas praised GLM-5.2 as the kind of model that reignites serious interest in open-source AI. The sub-trillion-parameter model passes blind tests against frontier competitors on median production-grade knowledge worker tasks while remaining affordable to serve. Srinivas also revealed that multiple trillion-parameter open-source releases are imminent, predicting a Jevons-paradox-driven boom in token consumption and price competition.
Cursor Partners with SpaceX to Train New Model at Compile
At its Compile conference keynote, Cursor AI unveiled three announcements including a partnership to train a new model with SpaceX. Details remain limited, but the collaboration signals intensifying convergence between aerospace infrastructure and AI model development. Cursor's move positions it at the intersection of two of the most compute-intensive industries.
Human intelligence is fundamentally a collective intelligence. We solve complex problems by participating in a vast cultural network that builds upon ideas across generations. I believe the strongest AI systems will become a collective intelligence, too.
David Ha, Founder & CEO of Sakana AI
Sakana AI Launches Fugu: Multi-Agent Orchestration via a Single API
Sakana AI introduced Sakana Fugu, a full multi-agent orchestration system accessible through a single model API call. Fugu itself is an LLM trained to call various models in an agent pool — including recursive instances of itself — enabling autonomous coordination across complex, long-horizon tasks without external orchestration frameworks. The Fugu Ultra variant represents the system's most powerful configuration, rivaling industry-leading engineering benchmarks from both Fable and Mythos.
NVIDIA: Data Center Water Use Is 0.2% of US Daily Total
Citing data from the Manhattan Institute, NVIDIA pushed back against claims that AI data centers are draining water resources. Daily water consumption across all US data centers accounts for only 0.2% of national usage — a figure that has dropped dramatically thanks to closed-loop liquid cooling systems achieving near-zero marginal water consumption. Perplexity CEO Arav Srinivas corroborated: properly implemented liquid cooling has negligible water needs, and common confusion conflates power-plant water usage with on-site data center cooling.
MIT-Licensed GLM-5.2 Beats GPT-5.5 on Agentic Benchmarks
Released under the permissive MIT license, GLM-5.2 outperforms GPT-5.5 (xhigh) on real-world agentic tasks and is freely available on Hugging Face. The release has drawn comparisons to the democratizing impact of LLaMA, with twenty providers now offering API access to the model alongside its availability on AWS Marketplace.
Patch the Planet: Frontier AI Defends Critical Open-Source Projects
Co-founded by Anthropic leadership, Patch the Planet deploys frontier AI models alongside professional security researchers to find and fix vulnerabilities in the most critical open-source software projects under active threat.
Codex Security: Deep Scans, Attack Path Tracing, Patch Generation
The Codex Security plugin combines deep code scanning, vulnerability validation, attack path tracing, threat modeling, and codebase-specific patch generation — all exportable into existing security toolchains for human review.
Grok Connects to Interactive Brokers for Real-Time Portfolio Intel
xAI launched Grok's integration with Interactive Brokers, giving users conversational access to high-quality, up-to-date portfolio information as part of Grok Build's expanding financial tool suite.
TMax: Open RL Recipe for Terminal Agents Published
A new paper introduces TMax, an open reinforcement learning formulation for terminal agents. Researcher Nat Lambert, who contributed to the work, notes that RL research in mid-2026 has already diverged sharply from the paradigms that dominated just months ago. AI2 separately released TMax 27B on Hugging Face, scoring 42.7% on Terminal Bench 2.0 and rivaling models ten times its size.
Elon Musk Announces Grok Build Upgrades with /goal Command
Elon Musk confirmed upgrades to Grok Build, headlined by the new /goal command that enables autonomous execution of long-running tasks with multiple rounds of subagents implementing and verifying progress. Observers note this marks a shift from coding agents that behave like enhanced chatbots toward truly autonomous task execution.
Gray Swan: Red Teaming and the Coming AI Security Crisis
OpenAI board member Zico Kolter and Gray Swan CEO Matt Fredrikson discussed the emerging AI security landscape on Latent Space, arguing that AI safety is not merely traditional cybersecurity plus AI — it demands fundamentally new approaches. Their red teaming of AI models reveals gaps that conventional defenses miss.
Vercel Flags Decouples Deployment from Feature Release
Vercel launched Vercel Flags, a platform-native feature flag system that executes server-side with zero impact on page performance. The tool enables teams to merge code when ready and flip feature toggles independently, fully decoupling deployment cadence from release timing.
GLM-5.2 Ranks Third Overall on GDPval-AA
Leads all open-weight models on the real-world agentic work benchmark, demonstrating competitive agent reasoning.
AI2 Releases Qwen 3.5 9B Terminal Agent
Trained with DPPO on the OpenThoughts dataset, the compact agent punches above its weight class.
Google DeepMind, A24 Launch AI Research Collaboration
The two companies will explore AI tools in creative fields, ensuring filmmakers help shape the technology.
Google Invests $75 Million in A24 AI Research
The investment targets AI tools for creative and film industries, deepening the tech-entertainment convergence.
Neuralink Reaches 26th Implant Recipient
A Vancouver police officer with ALS becomes the latest BCI patient, marking continued clinical progress.
AI2 TMax 27B Terminal Agent Hits 42.7%
A 27B model rivaling much larger systems on Terminal Bench 2.0, now available on Hugging Face.
Agentic RL Blog Surveys 10+ Frameworks
Comprehensive review highlights modular tool interfaces and XML-based function calling in Qwen3.
Cline Tests GLM-5.2 and Opus on Real-World Bugs
Skeptical of benchmarks, the Cline team ran head-to-head bug-fixing tests, with results favoring open models.
Nadella: Public Will Reject an AI Monopoly by a Few Labs
Microsoft CEO Satya Nadella argued the public would not tolerate a handful of AI labs doing all the learning for the world, as Microsoft accelerates efforts to broaden AI access across industries and geographies.
Chollet Slams SaaS Doomsayers as 'Staggeringly Short-Sighted'
Keras creator François Chollet dismissed the belief that Claude can one-shot any SaaS application, calling it shortsighted. Programming, he argues, is the art and science of managing complexity through layers of abstraction — AI is merely one tool in that arsenal.
Chollet: Adobe Is One of GenAI's Biggest, Most Profitable Winners
While markets treat Adobe as a legacy software company in terminal decline, Chollet points out it is among the top five most profitable and fastest-growing AI companies today — a GenAI beneficiary hiding in plain sight.
Hugging Face Nears 3 Million Models and 1 Million Datasets
Hugging Face co-founder Clement Delangue announced the platform is about to cross major open-source milestones, reflecting the explosive growth of community-driven AI and the platform's centrality to model distribution.
Perplexity CEO: Multi-Trillion-Parameter Open Models Coming Soon
Srinivas says multiple labs are preparing massive open-source releases, which stands to drive token prices down further through Jevons-paradox dynamics.
Seedance 2.0 Native 4K Video Model Released
Higgsfield AI launches best-in-class particle physics, lighting, and shadows for cinematic storytelling.
Sakana Fugu Ultra: Slow and Below Fable
Ethan Mollick reports typical coding tests take 30 minutes with average results — not matching Fable in real use.
Fable Demonstrates Creative Problem-Solving in Snake Game
Mollick built a self-aware Snake game with no design feedback — just "make it better" — showcasing Fable's judgment.
Distilling Frontier Models Costs ~$4M per Trillion Tokens
Analysis suggests distilling models like Opus 4.8 is surprisingly affordable at scale.
GLM 5.2 Impressions: Compliant but Lacks Initiative
Observers call it a "yes sir" model — strong execution but unlikely to explore new directions independently.
Chinese AI Labs Offer ¥6M Salaries, Recruit at Age 17
Core position campus recruits command millions in annual salary, with big tech dropping age limits entirely.
Huawei 950 NPU Cluster Rivals DeepSeek Infrastructure
8192-NPU SuperPOD spotted in Inner Mongolia; DeepSeek reportedly building similar clusters in Ulanqab.