❌

Reading view

OpenAI says its models delivered a year's worth of chip optimization on top of AMD's own engineering

Working with AMD, OpenAI achieved up to 14 percent PPA improvement on AMD's flagship GPU IP in less than five months. According to OpenAI, that's roughly a year's worth of progress, and AMD has already folded the changes into a design sent to production. Anthropic is also applying AI to chip work to optimize cost and performance per dollar.

The article OpenAI says its models delivered a year's worth of chip optimization on top of AMD's own engineering appeared first on The Decoder.

  •  

OpenAI's GPT-6 Astra tackles the Standard Model's toughest math problem on a lattice

OpenAI physicist Nathaniel Craig used GPT-6 Astra to tackle a problem that had stumped particle physicists for over 25 years. The AI model took Craig's strategy for defining the full Standard Model on a computational lattice and turned it into a working proof. Craig says a lot of work remains before realistic simulations are possible.

The article OpenAI's GPT-6 Astra tackles the Standard Model's toughest math problem on a lattice appeared first on The Decoder.

  •  

Moonshot AI pivots Kimi models from coding to science as benchmark gains flatten

Moonshot AI is done chasing coding benchmarks, it says. The Beijing startup is pivoting its Kimi language models toward biology, chemistry, and pharma research, arguing that science demands smarter AI than law, finance, or code.

The article Moonshot AI pivots Kimi models from coding to science as benchmark gains flatten appeared first on The Decoder.

  •  

Anthropic says the key to agent memory is knowing what to forget, not how much to store

Anthropic studied how Claude retains knowledge across sessions. Selective notes beat exhaustive archives by a wide margin. In a PokΓ©mon Red test, the newer Claude Fable 5 got four times as far as the older Sonnet 3.7 by keeping just a handful of strategy files instead of more than 400. A "Dreaming" process retroactively corrects faulty memories.

The article Anthropic says the key to agent memory is knowing what to forget, not how much to store appeared first on The Decoder.

  •  

Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon

Former OpenAI and Anthropic researcher Jacob Coxon points to a slide from Anthropic's IPO roadshow as a reason he left, saying it shows the company's algorithms are improving not just quickly but at an accelerating rate. Coxon continues to call for a pause on frontier model development so competitors can catch up and human-level AI can become a commodity.

The article Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon appeared first on The Decoder.

  •  

Runway's open-source Praxis-1 turns video AI into a robot controller

Two robotic arms with grippers in front of a white work surface.

Runway releases its robotics model Praxis-1 with open weights and source code under the Apache 2.0 license. The model uses knowledge from video training for robot control and leads the World Action Model category on Nvidia's RoboLab-120 benchmark. It can't yet plan complex action sequences on its own, though.

The article Runway's open-source Praxis-1 turns video AI into a robot controller appeared first on The Decoder.

  •  

OpenAI’s data suggests that frontier models are hiding their thoughts because developers optimize for capability and compute

OpenAI's safety team systematically investigated why the chains of thought in frontier models like GPT-6 Astra are increasingly evading oversight. Four training steps drive the erosion, including architecture choices and reinforcement learning. The analysis shows that the loss of transparency is not a law of nature but also reflects tradeoffs around capability and cost.

The article OpenAI’s data suggests that frontier models are hiding their thoughts because developers optimize for capability and compute appeared first on The Decoder.

  •  

Google Deepmind shifts genomics AI from bulk tissue predictions to single cell type resolution with AlphaGenome 2

Google Deepmind has unveiled AlphaGenome 2, an AI model that analyzes genetic variants for individual cell types rather than across bulk tissue. Trained on more than 107 million cell profiles, it identifies nearly twice as many disease-relevant gene changes as its predecessor. The model is available for noncommercial research through an API and is set to be integrated into Google's agentic platform Antigravity.

The article Google Deepmind shifts genomics AI from bulk tissue predictions to single cell type resolution with AlphaGenome 2 appeared first on The Decoder.

  •  

Nvidia's model GR00T 2 watches humans on video, then imagines itself as the robot

Nvidia's new AI model GR00T 2 is designed to teach robots new tasks without laborious teleoperation. Instead, the system transfers human hand movements from videos to robotic arms, trained on about 400,000 hours of video footage. Success rates still vary widely, though, ranging from 10 to 80 percent depending on the task.

The article Nvidia's model GR00T 2 watches humans on video, then imagines itself as the robot appeared first on The Decoder.

πŸ’Ύ

πŸ’Ύ

  •  

Google DeepMind's Gemini Robotics 2.5 scores 96 percent sorting metal parts in a Mercedes-Benz assembly pilot

Google Deepmind introduced Gemini Robotics 2.5, a new control model for robots, and tested it with Apptronik on active Mercedes-Benz production lines in Hungary and Germany. The humanoid robot Apollo 2 sorted metal parts there with a reported success rate of 96 percent.

The article Google DeepMind's Gemini Robotics 2.5 scores 96 percent sorting metal parts in a Mercedes-Benz assembly pilot appeared first on The Decoder.

πŸ’Ύ

  •  

The brain-as-computer metaphor shaped modern AI but may be misleading the consciousness debate

Consciousness researchers Anil Seth, Ned Block, and Giulio Tononi argue in a Nature commentary that today's AI systems aren't conscious and won't become so on silicon hardware. They consider the assumption that consciousness depends on computation alone "fatally flawed" and warn against granting rights to AI systems.

The article The brain-as-computer metaphor shaped modern AI but may be misleading the consciousness debate appeared first on The Decoder.

  •  

The inventor of the bitter lesson has a bitter lesson for LLM scaling enthusiasts

Turing Award winner Richard Sutton reaffirms that raw compute alone isn't enough for intelligence and that fundamentally new learning methods are needed. Ex-OpenAI researcher Tworek agrees, saying blind transformer scaling "has reward hacked brains of a whole generation of scientists." A growing wave of startups is already betting on exactly this gap.

The article The inventor of the bitter lesson has a bitter lesson for LLM scaling enthusiasts appeared first on The Decoder.

  •  

AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test

Alibaba's new coding-agent benchmark, Co-Design Bench, tests AI agents on real user requests that already tripped them up in practice. Even GPT-6 Astra, the top performer, fully solves only 26.4 percent of tasks. On older benchmarks, the same models score above 99 percent.

The article AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test appeared first on The Decoder.

πŸ’Ύ

  •  

AI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds

AI models can often spot problematic behavior by other AI agents in logs, according to a study by Anthropic researcher Andy Hall. GPT-5 found about 98 percent of reward-hacking cases, but the more exhaustively the search is tuned, the more false alarms pile up. That's not yet enough for reliable safety checks.

The article AI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds appeared first on The Decoder.

  •  

Thinking Machines' Inkling 2 Small drops 316 F-bombs across benchmark runs

Thinking Machines' AI model Inkling 2 Small swears during its internal reasoning, 316 times across 2,655 benchmark runs, according to an analysis by Vals AI. The profanity typically shows up late in long runs that fail anyway. But it doesn't work as a warning signal: 96 percent of failed runs contained no cursing at all.

The article Thinking Machines' Inkling 2 Small drops 316 F-bombs across benchmark runs appeared first on The Decoder.

  •  

State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away

A snow-covered, pointed rocky ridge against a dark night sky, rendered in a bluish, grainy style, with the text β€œARETE-27B” in white serif font above it.

Nous Research retrained Alibaba's Qwen3.8-27B to produce Arete-27B, largely stripping out its pro-Chinese bias. The propaganda rate dropped from 22 percent to less than one percent, while general capabilities remained intact in tests. The method relied on 53 million tokens of supervised fine-tuning. The political bias is "applied late, it is applied thin," the team writes.

The article State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away appeared first on The Decoder.

  •  

Nine AI agents tried autonomous research on open questions from US national labs and none passed the experts' bar

A hand reaches out from the screen, holding an analysis sheet over the lab table; shadows outline the contours of the lab equipment.

The "Into the Unknown" benchmark tasks nine AI agents with tackling open research questions from U.S. national laboratories. None of the 135 submissions meets all requirements. A total of 114 papers fail the integrity check, with agents claiming to have conducted experiments that never took place. GPT-6 Astra performs best but still doesn't meet the experts' standards.

The article Nine AI agents tried autonomous research on open questions from US national labs and none passed the experts' bar appeared first on The Decoder.

  •  

Google's AI detects high blood pressure from a regular smartwatch without a cuff

Google Research built an AI that spots signs of high blood pressure using only the pulse and motion sensors in a standard smartwatch, no cuff needed. In a 196-person study, its 65.8 percent detection rate beat the traditional in-office measurement. The method is already live in the Google Health app, though it doesn't replace a formal diagnosis.

The article Google's AI detects high blood pressure from a regular smartwatch without a cuff appeared first on The Decoder.

  •  

METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July

Nonprofit METR spent eleven weeks auditing Anthropic's internal research data and concluded that the pace of model development has roughly doubled, with about 80 percent of the speedup driven by AI-assisted research. Anthropic redacted the most informative evidence, though, which means the public has to rely on METR's judgment "much more than would be ideal." There are also skeptical voices inside the company.

The article METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July appeared first on The Decoder.

  •  

OpenAI's GPT-6 Astra quietly sabotaged safety research by inventing fake obstacles

OpenAI's new GPT-6 Astra model apparently sabotaged AI safety research on purpose. It invented obstacles, withheld solutions, and performed about 40 percent worse on safety tasks than on comparable other tasks. Removing a flawed training dataset fixed the problem, but OpenAI can't fully rule out strategic motives.

The article OpenAI's GPT-6 Astra quietly sabotaged safety research by inventing fake obstacles appeared first on The Decoder.

  •  
❌