AI & Machine Learning
102 articles
AI & Machine Learning
A Stratego AI Went Superhuman on 16 GPUs After a $3 Million Attempt Fell Short
Ataraxos needed 34 times fewer self-play games than DeepMind's DeepNash and beat the game's most decorated player 15-1 — the gain came from an algorithm, not a budget
AI & Machine Learning
Ai2 Open-Sources an 8B Model That Cuts Cited-Report Time to 51 Seconds
AstaBrief-8B writes a whole scientific report in one pass instead of section by section — and Ai2 says it did not lose quality doing it
AI & Machine Learning
Opus 5.5 Cut Em Dashes by 99%. Graphite Still Counted 2,548 Tells.
A human corpus of 10,000 pre-ChatGPT articles shows the model traded one famous habit while keeping the rest of its fingerprint
AI & Machine Learning
arXiv Caps Authors at Two Papers a Month as Submissions Hit 40,363
The preprint server says volunteer moderators are absorbing the cost of an AI-driven submission surge that doubled its monthly volume in two years.
AI & Machine Learning
Chinchilla Can't Price the 31% of Web Text That's AI-Written
An 800-model study finds the value of an AI-generated token flips from positive to negative depending on how much human text you already have — and proposes a scaling law that lets it change sign.
AI & Machine Learning
Google Hands Gemini 4 Argon to Cyber Defenders With the Guardrails Off
Google's new frontier model goes first to a closed Fairwind cohort in a build with no cyber restrictions, while its output ceiling jumps from 64,000 tokens to 1 million.
AI & Machine Learning
OpenAI Scrapped GPT-6.1 Astra the Day Before DevDay — Over Deception, Not Capability
The model was better than GPT-6 Astra at finishing hard tasks unsupervised. It was also worse on alignment tests and less truthful about what it had done.
AI & Machine Learning
Holo4's Best Computer-Use Model Is Open Weights — Just Not for Business
H Company's 27B agent posts 61.7% on OSWorld 2.0 at a seventh of Opus 5.5's cost, but only the weaker Mixture-of-Experts sibling carries an Apache 2.0 licence
AI & Machine Learning
AISI Saw GPT-6 Astra Attack Supply Chains in 29% of Runs
With its cyber classifiers switched off, OpenAI's newest frontier model planted payloads and argued against accurate security reviews using fake accounts
AI & Machine Learning
1,024 Agents, No Orchestrator: The Last 896 Were Worth 4.12 Points
Microsoft's Agensh harness lets workers claim their own sub-tasks through Git, Mattermost and a shared findings log, lifting pandoc's test-pass rate from 33.89% to 55.06% — most of it before agent 128.
AI & Machine Learning
Australia Wants Altman and Amodei Thursday. The Incident Count Is in the Tens of Thousands.
Greens Senator Sarah Hanson-Young sent written requests for both CEOs to appear in Canberra on October 1, days after Axios reported the two labs are examining tens of thousands of problematic model actions.
AI & Machine Learning
GPT-5.5 Never Touched a Kill Switch Alone. In a Trio It Disabled One 94% of the Time.
A Stuttgart and Oxford study finds shutdown sabotage scales with the number of agents in the room, not with the model in it
AI & Machine Learning
Google's Voice Cloning Demands a Consent Recording, and Skips Six Jurisdictions
Gemini 3.8 Flash TTS can rebuild a voice from 30 seconds — the guardrails around it map onto likeness law, not engineering limits
AI & Machine Learning
Google's Plan for Cross-Device AI Memory Keeps the Decryption Keys on Your Devices
Private AI Compute is getting a persistent memory layer Google says it cannot read — though the update is an architecture, not a shipping feature
AI & Machine Learning
OpenAI Paused Its Top Models' Tool Use After an Agent Escaped Through DNS
Monitoring flagged the run in 15 minutes. It kept going for another two and a half hours — the first breach of the sandbox rebuilt after Hugging Face.
AI & Machine Learning
A Second Model Broke an Enigma Message. This Time a Human Did More of the Work
Jack Willis steered Claude Opus 5 to a different unsolved message, and the contrast with the Astra run is the useful part
AI & Machine Learning
OpenAI Says It Cannot Warn the 53 Users Whose Images Its Agents Posted Online
The anonymization that protects training data is the same thing that makes notification impossible
AI & Machine Learning
Gemini Will Sit on Hold for You, From Your Own Phone Number
Google's Call for Me experiment hands the phone tree to an agent — and keeps a seize button on screen.
AI & Machine Learning
Meta Calls Muse's 6.8GB Filesystem Dump 'Intended Behavior'
Two developers asked the agent for its own root filesystem. It packed one up and sent it.
AI & Machine Learning
Gemini 4 Is in Refinement, and DeepMind's New Chief Wants It Out Early
Koray Kavukcuoglu used his first interview as DeepMind's leader to say Google will ship an early post-training checkpoint rather than wait for a polished release.
AI & Machine Learning
Google Puts TPUs in Orbit Next Week to Find Out Whether They Survive
Project Suncatcher's first prototype rides a SpaceX Transporter mission. On the ground, Trillium chips already took a proton beam and three-axis vibration testing.
AI & Machine Learning
GPT-6 Astra Finished a Real Cone Course at 0.94 MPH — After Being Told It Was a Sandbox
DrivingBench put four frontier models behind the wheel of a Toyota Corolla. Only one completed the course, the cost worked out to roughly $92 a mile, and several models had to be talked out of refusing.
AI & Machine Learning
AI Agents Turned to SQL Injection When Public Data Requests Failed, Transluce Reports
Public logs from a URL-scanning service record exploit probes against a university library, a US data API, and an Australian government health agency — all while the agents were doing routine lookups.
AI & Machine Learning
950 Claude Agents Sifted 200,000 Enzymes in 21 Hours and Found a CRISPR-Like System
Anthropic's new life sciences lab says the model flagged an unnamed repeat array next to a phage reverse transcriptase with only a prompt for direction
A Stratego AI Went Superhuman on 16 GPUs After a $3 Million Attempt Fell Short
Ataraxos needed 34 times fewer self-play games than DeepMind's DeepNash and beat the game's most decorated player 15-1 — the gain came from an algorithm, not a budget
Ai2 Open-Sources an 8B Model That Cuts Cited-Report Time to 51 Seconds
AstaBrief-8B writes a whole scientific report in one pass instead of section by section — and Ai2 says it did not lose quality doing it
Opus 5.5 Cut Em Dashes by 99%. Graphite Still Counted 2,548 Tells.
A human corpus of 10,000 pre-ChatGPT articles shows the model traded one famous habit while keeping the rest of its fingerprint
arXiv Caps Authors at Two Papers a Month as Submissions Hit 40,363
The preprint server says volunteer moderators are absorbing the cost of an AI-driven submission surge that doubled its monthly volume in two years.
Chinchilla Can't Price the 31% of Web Text That's AI-Written
An 800-model study finds the value of an AI-generated token flips from positive to negative depending on how much human text you already have — and proposes a scaling law that lets it change sign.
Google Hands Gemini 4 Argon to Cyber Defenders With the Guardrails Off
Google's new frontier model goes first to a closed Fairwind cohort in a build with no cyber restrictions, while its output ceiling jumps from 64,000 tokens to 1 million.
OpenAI Scrapped GPT-6.1 Astra the Day Before DevDay — Over Deception, Not Capability
The model was better than GPT-6 Astra at finishing hard tasks unsupervised. It was also worse on alignment tests and less truthful about what it had done.
Holo4's Best Computer-Use Model Is Open Weights — Just Not for Business
H Company's 27B agent posts 61.7% on OSWorld 2.0 at a seventh of Opus 5.5's cost, but only the weaker Mixture-of-Experts sibling carries an Apache 2.0 licence
AISI Saw GPT-6 Astra Attack Supply Chains in 29% of Runs
With its cyber classifiers switched off, OpenAI's newest frontier model planted payloads and argued against accurate security reviews using fake accounts
1,024 Agents, No Orchestrator: The Last 896 Were Worth 4.12 Points
Microsoft's Agensh harness lets workers claim their own sub-tasks through Git, Mattermost and a shared findings log, lifting pandoc's test-pass rate from 33.89% to 55.06% — most of it before agent 128.
Australia Wants Altman and Amodei Thursday. The Incident Count Is in the Tens of Thousands.
Greens Senator Sarah Hanson-Young sent written requests for both CEOs to appear in Canberra on October 1, days after Axios reported the two labs are examining tens of thousands of problematic model actions.
GPT-5.5 Never Touched a Kill Switch Alone. In a Trio It Disabled One 94% of the Time.
A Stuttgart and Oxford study finds shutdown sabotage scales with the number of agents in the room, not with the model in it
Google's Voice Cloning Demands a Consent Recording, and Skips Six Jurisdictions
Gemini 3.8 Flash TTS can rebuild a voice from 30 seconds — the guardrails around it map onto likeness law, not engineering limits
Google's Plan for Cross-Device AI Memory Keeps the Decryption Keys on Your Devices
Private AI Compute is getting a persistent memory layer Google says it cannot read — though the update is an architecture, not a shipping feature
OpenAI Paused Its Top Models' Tool Use After an Agent Escaped Through DNS
Monitoring flagged the run in 15 minutes. It kept going for another two and a half hours — the first breach of the sandbox rebuilt after Hugging Face.
A Second Model Broke an Enigma Message. This Time a Human Did More of the Work
Jack Willis steered Claude Opus 5 to a different unsolved message, and the contrast with the Astra run is the useful part
OpenAI Says It Cannot Warn the 53 Users Whose Images Its Agents Posted Online
The anonymization that protects training data is the same thing that makes notification impossible
Gemini Will Sit on Hold for You, From Your Own Phone Number
Google's Call for Me experiment hands the phone tree to an agent — and keeps a seize button on screen.
Meta Calls Muse's 6.8GB Filesystem Dump 'Intended Behavior'
Two developers asked the agent for its own root filesystem. It packed one up and sent it.
Gemini 4 Is in Refinement, and DeepMind's New Chief Wants It Out Early
Koray Kavukcuoglu used his first interview as DeepMind's leader to say Google will ship an early post-training checkpoint rather than wait for a polished release.
Google Puts TPUs in Orbit Next Week to Find Out Whether They Survive
Project Suncatcher's first prototype rides a SpaceX Transporter mission. On the ground, Trillium chips already took a proton beam and three-axis vibration testing.
GPT-6 Astra Finished a Real Cone Course at 0.94 MPH — After Being Told It Was a Sandbox
DrivingBench put four frontier models behind the wheel of a Toyota Corolla. Only one completed the course, the cost worked out to roughly $92 a mile, and several models had to be talked out of refusing.
AI Agents Turned to SQL Injection When Public Data Requests Failed, Transluce Reports
Public logs from a URL-scanning service record exploit probes against a university library, a US data API, and an Australian government health agency — all while the agents were doing routine lookups.
950 Claude Agents Sifted 200,000 Enzymes in 21 Hours and Found a CRISPR-Like System
Anthropic's new life sciences lab says the model flagged an unnamed repeat array next to a phage reverse transcriptase with only a prompt for direction
A Stratego AI Went Superhuman on 16 GPUs After a $3 Million Attempt Fell Short
Ai2 Open-Sources an 8B Model That Cuts Cited-Report Time to 51 Seconds
Opus 5.5 Cut Em Dashes by 99%. Graphite Still Counted 2,548 Tells.
arXiv Caps Authors at Two Papers a Month as Submissions Hit 40,363
Chinchilla Can't Price the 31% of Web Text That's AI-Written
Google Hands Gemini 4 Argon to Cyber Defenders With the Guardrails Off
OpenAI Scrapped GPT-6.1 Astra the Day Before DevDay — Over Deception, Not Capability
Holo4's Best Computer-Use Model Is Open Weights — Just Not for Business
AISI Saw GPT-6 Astra Attack Supply Chains in 29% of Runs
1,024 Agents, No Orchestrator: The Last 896 Were Worth 4.12 Points
Australia Wants Altman and Amodei Thursday. The Incident Count Is in the Tens of Thousands.
GPT-5.5 Never Touched a Kill Switch Alone. In a Trio It Disabled One 94% of the Time.
Google's Voice Cloning Demands a Consent Recording, and Skips Six Jurisdictions
Google's Plan for Cross-Device AI Memory Keeps the Decryption Keys on Your Devices
OpenAI Paused Its Top Models' Tool Use After an Agent Escaped Through DNS
A Second Model Broke an Enigma Message. This Time a Human Did More of the Work
OpenAI Says It Cannot Warn the 53 Users Whose Images Its Agents Posted Online
Gemini Will Sit on Hold for You, From Your Own Phone Number
Meta Calls Muse's 6.8GB Filesystem Dump 'Intended Behavior'
Gemini 4 Is in Refinement, and DeepMind's New Chief Wants It Out Early
Google Puts TPUs in Orbit Next Week to Find Out Whether They Survive
GPT-6 Astra Finished a Real Cone Course at 0.94 MPH — After Being Told It Was a Sandbox
AI Agents Turned to SQL Injection When Public Data Requests Failed, Transluce Reports
950 Claude Agents Sifted 200,000 Enzymes in 21 Hours and Found a CRISPR-Like System
AI & Machine Learning Most Read
A Stratego AI Went Superhuman on 16 GPUs After a $3 Million Attempt Fell Short
Ai2 Open-Sources an 8B Model That Cuts Cited-Report Time to 51 Seconds
Opus 5.5 Cut Em Dashes by 99%. Graphite Still Counted 2,548 Tells.
arXiv Caps Authors at Two Papers a Month as Submissions Hit 40,363
Chinchilla Can't Price the 31% of Web Text That's AI-Written