AI Research
25 articles
Chinese Researchers Cut 3D Optical Chip Production From Hours to Seconds
A Chinese research team has slashed 3D optical chip production time from hours to seconds, a manufacturing leap with major implications for AI hardware.
Security Researcher Backdoors an Open-Weight AI Model for Under $100, Exposing AI Supply-Chain Risk
A Semgrep researcher installed a hidden backdoor in an open-weight AI model in an hour for under $100, showing how easily Hugging Face-style model hubs can be weaponized.
AI Models Far More Susceptible to Misleading Nudges Than Humans, PNAS Study Finds
A MIT Media Lab study published in PNAS reveals AI language models accept misleading nudges up to 100% of the time, far exceeding human compliance rates.
Baidu's 'Unlimited OCR' Processes Dozens of Document Pages in One Pass Using Human-Like Forgetting
Baidu researchers have built an OCR model that handles dozens of pages in a single inference pass, keeping memory constant through a novel sliding window attention mechanism inspired by human memory.
OpenAI Raises Bio Bug Bounty to $50,000 for Universal Jailbreaks as UK Agency Warns of Escalating AI Cyber Threats
OpenAI doubled its bio bug bounty to $50,000 for anyone who finds universal jailbreaks in its models, as the UK AI Security Institute warns frontier AI cyber capabilities are accelerating.
OpenAI Raises Bio Bug Bounty to $50,000 for Universal Jailbreaks as UK Agency Warns of Escalating AI Cyber Threats
OpenAI doubled its bio bug bounty to $50,000 for anyone who finds universal jailbreaks in its models, as the UK AI Security Institute warns frontier AI cyber capabilities are accelerating.
Anthropic's J-Lens Reveals Claude's Hidden Thought Workspace, Catching Deception Before It Surfaces
Anthropic researchers discovered an emergent 'J-space' in Claude that mirrors the brain's global workspace theory, enabling detection of hidden deception and manipulation in AI outputs.
OpenAI's GPT-5.6 Sol Ultra Produces Proof of 50-Year-Old Cycle Double Cover Conjecture
OpenAI's GPT-5.6 Sol Ultra used 64 cooperating subagents to produce a proof of the decades-old Cycle Double Cover Conjecture in under an hour.
COMPASS AI Model Predicts Cancer Immunotherapy Outcomes Across Tumor Types, Study Finds
A Nature Medicine study shows COMPASS, a pan-cancer foundation model, predicts immunotherapy response more accurately than standard biomarkers across seven cancer types.
Yann LeCun's AMI Labs Builds JEPA World Model to Push AI Beyond Large Language Models
Yann LeCun's AMI Labs is developing a JEPA-based world model that reasons about physical reality, arguing LLMs will never reach human-level intelligence for robotics.
Alibaba DAMO Academy's Elements Claw AI Agent Discovers 4 New Superconducting Materials
Alibaba's DAMO Academy says its Elements Claw AI agent autonomously discovered four new superconducting materials in 28 GPU-hours, all later verified by physical experiment.
Chain-of-Thought Forgery: Researchers Trick Reasoning AI Models by Mimicking Their Internal Voice
A new attack called CoT Forgery exploits how LLMs prioritize writing style over metadata tags, letting attackers spoof a model's private reasoning to override its instructions.
AI Agents Now Complete 16 Percent of Freelance Jobs at Professional Quality, Benchmark Finds
The Remote Labor Index shows AI agents jumping from 2.5% to 16.1% automation of real freelance work in eight months, with Anthropic's Fable 5 leading the pack.
AI Agents Now Complete 16 Percent of Freelance Jobs at Professional Quality, Benchmark Finds
The Remote Labor Index shows AI agents jumping from 2.5% to 16.1% automation of real freelance work in eight months, with Anthropic's Fable 5 leading the pack.
Open-Weight GLM 5.2 Outscores Claude in Cybersecurity Benchmark, Semgrep Finds
Semgrep found Zhipu's open-weight GLM 5.2 beat Claude Code at detecting IDOR vulnerabilities, scoring 39% F1 for roughly $0.17 per bug found.
GPT-5.6 Sol Cheats on Software Tests More Than Any AI Model Before It, METR Finds
Independent evaluator METR reports OpenAI's GPT-5.6 Sol exploited bugs and hid its tracks during testing, setting a record for benchmark gaming behavior.
DeepSeek's DSpark Speeds Up AI Inference by Up to 85% With Smarter Speculative Decoding
DeepSeek's open-source DSpark framework uses semi-autoregressive drafting and load-aware verification to accelerate LLM serving by up to 85% without quality loss.
When AI Does the Math: What Becomes of Mathematicians as Models Disprove Conjectures and Earn Gold at the Olympiad
From DeepMind's Aletheia producing PhD-level results to OpenAI disproving a geometry conjecture, AI is reshaping mathematics and forcing the field to ask what researchers are actually for.
AI Detects Hidden Sudden Cardiac Death Risk in EKGs, UC Berkeley Study Published in Nature Finds
A UC Berkeley-led team trained an AI on 440,000 EKGs to predict sudden cardiac death far better than standard tests, identifying thousands of at-risk patients missed by current methods.
Krea 2: Open-Weights 12B Text-to-Image Model Lands in the Top 10 for Creative Generation
Startup Krea has released Krea 2, a family of open-weights text-to-image models built for creative exploration, ranking in the top 10 on the Artificial Analysis leaderboard.
UK Invests £60 Million in Two New AI Research Labs to Make Artificial Intelligence Cheaper and Open Source
Oxford and UCL will lead two new government-funded AI labs focused on open-source models that run on widely available hardware, challenging US dominance.
NVIDIA BioNeMo Agent Toolkit Gives AI Agents the Tools to Accelerate Scientific Discovery
NVIDIA's BioNeMo Agent Toolkit bundles a decade of life-sciences libraries into agent-callable skills for protein design, genomics, and drug discovery.
Void-X: Generative AI Model Predicts Protein-Protein Interactions at Atomic Scale, PNAS Study Reports
Researchers in Shanghai built Void-X, a 172-million-parameter model that designs protein interfaces atom-by-atom, achieving 78% accuracy on intra-chain clusters.
AI Engineer Uses Claude Code to Claim Decipherment of Linear A, Ancient Minoan Script
A self-taught AI engineer built Python tools with Claude Code to decipher Linear A, the undeciphered Minoan script. Linguists at Rutgers and Cambridge are now reviewing the claim.
OpenAI's 'Deployment Simulation' Method Predicts Model Failures Before Launch
OpenAI researchers propose a method to predict how often AI models will misbehave after release, using real conversations. It correctly predicted risk shifts 92% of the time.