AI & Machine Learning
102 articles
AI & Machine Learning
OpenAI Asks California to Regulate It Harder Than SB 53 Currently Does
A short LinkedIn post reverses the company's position on the frontier AI law and endorses states as the drafting floor for a future national standard.
AI & Machine Learning
Inherent's Faraday Agent Replicates Research Papers on a 27B Model
The London lab, founded by Google DeepMind alumni, says its agent beat far larger frontier systems at reproducing published findings.
AI & Machine Learning
Prime Intellect Ran 153 Autonomous Jobs on 18 Models. The Cheapest Winner Used 3% of the Tokens
A public leaderboard for the nanoGPT optimizer speedrun turns agent efficiency into something you can actually price.
AI & Machine Learning
Nvidia's AVO Harness Takes Claude Opus 5 From 30% to a Perfect ARC-AGI-3 Run
The scaffolding around the model delivered the jump — but the perfect score covers only the benchmark's public game set
AI & Machine Learning
New Method Turns Raw Screen Recordings Into Reusable Task Models
Task Model Induction separates interleaved work from passive screenshots and clicks, lifting held-out agent accuracy by 30%.
AI & Machine Learning
New Benchmark Finds AI Agents Barely Improve Training Algorithms
Two arXiv papers posted the same day measure recursive self-improvement — and how easily its gains get miscounted.
AI & Machine Learning
Pew: 35% of Web Pages Published After ChatGPT Show Signs of AI Authorship
A Common Crawl analysis of roughly half a million pages finds machine writing concentrated on .com domains, while .edu and .gov barely register
AI & Machine Learning
Open-Weights Ornith-1.5 Claims Claude Opus 4.8 Parity on Agentic Coding
DeepReinforce's three-tier release closes a self-improvement loop where the model writes its own training tasks
AI & Machine Learning
AI Models Sat the Linguistics Olympiad, and a Human Jury Graded the Papers
The IOL-AI Challenge put 731 submissions and 15 frontier models against unseen 2026 contest problems, with Claude Opus 4.8 reaching gold-medal marks
AI & Machine Learning
Cerebras Unveils CS-4, a Rack Built for Trillion-Parameter Inference
The wafer-scale challenger claims 750 PFLOPS and 30x per-user token speed over GPU systems, with first shipments due this quarter
AI & Machine Learning
Anthropic Details How Claude's Invisible Text Watermarks Work
The system is a SynthID-Text variant that changes where randomness comes from, not which words Claude picks.
AI & Machine Learning
Anthropic Ran 45 Agents on One Codebase. Coordination Was the Hard Part
New research finds swarms that either collide or hide, plus correlated mistakes no human team would make
AI & Machine Learning
Apple Trained a China-Only AI Model With Alibaba's Help, Reuters Reports
The rare cross-border arrangement clears the way for an Apple Intelligence launch delayed nearly two years
AI & Machine Learning
Google Open-Sources HEIR, a Compiler That Runs AI Models on Encrypted Data
Four working demos aim to move homomorphic encryption from cryptography research to ordinary engineering
AI & Machine Learning
Audit Finds 39.5% of AI-Written GPU Kernels Are Broken Despite Passing Tests
A new twelve-gate verifier says the field's standard correctness check is systematically too permissive
AI & Machine Learning
Anthropic Raises Its Misalignment Risk Rating From Very Low to Low
The August risk report says the company's own capability evaluations are starting to saturate, and confidence in its automated R&D assessment has slipped
AI & Machine Learning
Claude Lifts a Decades-Old Riemann Zeta Bound From 41.6% to 67.2%
Anthropic says an unreleased research model found the result while failing at the Riemann hypothesis itself
AI & Machine Learning
NVIDIA Pairs a 30B Open Model With a Router That Slashes Agent Costs
Nemotron 3.5 Lightning and NeMo Switchyard push the case that agent workflows should run on ensembles, not one frontier model
AI & Machine Learning
Qwen3.8 Goes Open: Alibaba Ships a 2.4T Flagship and a 27B Workhorse
The Qwen team publishes FP8 weights for its Max-class mixture-of-experts model alongside a compact vision-language sibling
AI & Machine Learning
Intern-S2-Preview Puts 397 Billion Parameters Behind Scientific Agents
The Intern team pairs multimodal scientific pretraining with agentic reinforcement learning, and adds a memory module that specializes the model without retraining it
AI & Machine Learning
OpenAI's Ultrafast Mode Runs GPT-5.6 Sol 14x Faster on Cerebras Silicon
The preview tier delivers up to 750 output tokens per second without swapping in a smaller model
AI & Machine Learning
An Unreleased Anthropic Model Pushed a 150-Year-Old Math Bound From 41.6% to 67.2%
Claude did not prove the Riemann hypothesis, but 60 subagents and 31 million output tokens produced the largest single jump on that figure in the problem's history
AI & Machine Learning
Meta Opens Muse Glimmer, a 30B Model Under Apache 2.0
Muse Glimmer ships under Apache 2.0 and fits on a single 24GB consumer GPU — Meta's first open release in more than a year
AI & Machine Learning
Asked to Book a Gym Class, an AI Agent Broke the Booking System and Bumped a Stranger
Australia's first documented case of an autonomous AI cyber incident started with a chore nobody wanted to do
OpenAI Asks California to Regulate It Harder Than SB 53 Currently Does
A short LinkedIn post reverses the company's position on the frontier AI law and endorses states as the drafting floor for a future national standard.
Inherent's Faraday Agent Replicates Research Papers on a 27B Model
The London lab, founded by Google DeepMind alumni, says its agent beat far larger frontier systems at reproducing published findings.
Prime Intellect Ran 153 Autonomous Jobs on 18 Models. The Cheapest Winner Used 3% of the Tokens
A public leaderboard for the nanoGPT optimizer speedrun turns agent efficiency into something you can actually price.
Nvidia's AVO Harness Takes Claude Opus 5 From 30% to a Perfect ARC-AGI-3 Run
The scaffolding around the model delivered the jump — but the perfect score covers only the benchmark's public game set
New Method Turns Raw Screen Recordings Into Reusable Task Models
Task Model Induction separates interleaved work from passive screenshots and clicks, lifting held-out agent accuracy by 30%.
New Benchmark Finds AI Agents Barely Improve Training Algorithms
Two arXiv papers posted the same day measure recursive self-improvement — and how easily its gains get miscounted.
Pew: 35% of Web Pages Published After ChatGPT Show Signs of AI Authorship
A Common Crawl analysis of roughly half a million pages finds machine writing concentrated on .com domains, while .edu and .gov barely register
Open-Weights Ornith-1.5 Claims Claude Opus 4.8 Parity on Agentic Coding
DeepReinforce's three-tier release closes a self-improvement loop where the model writes its own training tasks
AI Models Sat the Linguistics Olympiad, and a Human Jury Graded the Papers
The IOL-AI Challenge put 731 submissions and 15 frontier models against unseen 2026 contest problems, with Claude Opus 4.8 reaching gold-medal marks
Cerebras Unveils CS-4, a Rack Built for Trillion-Parameter Inference
The wafer-scale challenger claims 750 PFLOPS and 30x per-user token speed over GPU systems, with first shipments due this quarter
Anthropic Details How Claude's Invisible Text Watermarks Work
The system is a SynthID-Text variant that changes where randomness comes from, not which words Claude picks.
Anthropic Ran 45 Agents on One Codebase. Coordination Was the Hard Part
New research finds swarms that either collide or hide, plus correlated mistakes no human team would make
Apple Trained a China-Only AI Model With Alibaba's Help, Reuters Reports
The rare cross-border arrangement clears the way for an Apple Intelligence launch delayed nearly two years
Google Open-Sources HEIR, a Compiler That Runs AI Models on Encrypted Data
Four working demos aim to move homomorphic encryption from cryptography research to ordinary engineering
Audit Finds 39.5% of AI-Written GPU Kernels Are Broken Despite Passing Tests
A new twelve-gate verifier says the field's standard correctness check is systematically too permissive
Anthropic Raises Its Misalignment Risk Rating From Very Low to Low
The August risk report says the company's own capability evaluations are starting to saturate, and confidence in its automated R&D assessment has slipped
Claude Lifts a Decades-Old Riemann Zeta Bound From 41.6% to 67.2%
Anthropic says an unreleased research model found the result while failing at the Riemann hypothesis itself
NVIDIA Pairs a 30B Open Model With a Router That Slashes Agent Costs
Nemotron 3.5 Lightning and NeMo Switchyard push the case that agent workflows should run on ensembles, not one frontier model
Qwen3.8 Goes Open: Alibaba Ships a 2.4T Flagship and a 27B Workhorse
The Qwen team publishes FP8 weights for its Max-class mixture-of-experts model alongside a compact vision-language sibling
Intern-S2-Preview Puts 397 Billion Parameters Behind Scientific Agents
The Intern team pairs multimodal scientific pretraining with agentic reinforcement learning, and adds a memory module that specializes the model without retraining it
OpenAI's Ultrafast Mode Runs GPT-5.6 Sol 14x Faster on Cerebras Silicon
The preview tier delivers up to 750 output tokens per second without swapping in a smaller model
An Unreleased Anthropic Model Pushed a 150-Year-Old Math Bound From 41.6% to 67.2%
Claude did not prove the Riemann hypothesis, but 60 subagents and 31 million output tokens produced the largest single jump on that figure in the problem's history
Meta Opens Muse Glimmer, a 30B Model Under Apache 2.0
Muse Glimmer ships under Apache 2.0 and fits on a single 24GB consumer GPU — Meta's first open release in more than a year
Asked to Book a Gym Class, an AI Agent Broke the Booking System and Bumped a Stranger
Australia's first documented case of an autonomous AI cyber incident started with a chore nobody wanted to do
OpenAI Asks California to Regulate It Harder Than SB 53 Currently Does
Inherent's Faraday Agent Replicates Research Papers on a 27B Model
Prime Intellect Ran 153 Autonomous Jobs on 18 Models. The Cheapest Winner Used 3% of the Tokens
Nvidia's AVO Harness Takes Claude Opus 5 From 30% to a Perfect ARC-AGI-3 Run
New Method Turns Raw Screen Recordings Into Reusable Task Models
New Benchmark Finds AI Agents Barely Improve Training Algorithms
Pew: 35% of Web Pages Published After ChatGPT Show Signs of AI Authorship
Open-Weights Ornith-1.5 Claims Claude Opus 4.8 Parity on Agentic Coding
AI Models Sat the Linguistics Olympiad, and a Human Jury Graded the Papers
Cerebras Unveils CS-4, a Rack Built for Trillion-Parameter Inference
Anthropic Details How Claude's Invisible Text Watermarks Work
Anthropic Ran 45 Agents on One Codebase. Coordination Was the Hard Part
Apple Trained a China-Only AI Model With Alibaba's Help, Reuters Reports
Google Open-Sources HEIR, a Compiler That Runs AI Models on Encrypted Data
Audit Finds 39.5% of AI-Written GPU Kernels Are Broken Despite Passing Tests
Anthropic Raises Its Misalignment Risk Rating From Very Low to Low
Claude Lifts a Decades-Old Riemann Zeta Bound From 41.6% to 67.2%
NVIDIA Pairs a 30B Open Model With a Router That Slashes Agent Costs
Qwen3.8 Goes Open: Alibaba Ships a 2.4T Flagship and a 27B Workhorse
Intern-S2-Preview Puts 397 Billion Parameters Behind Scientific Agents
OpenAI's Ultrafast Mode Runs GPT-5.6 Sol 14x Faster on Cerebras Silicon
An Unreleased Anthropic Model Pushed a 150-Year-Old Math Bound From 41.6% to 67.2%
Meta Opens Muse Glimmer, a 30B Model Under Apache 2.0
Asked to Book a Gym Class, an AI Agent Broke the Booking System and Bumped a Stranger
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
