(via DEV) LLMs Can Autonomously Plan, Execute Cyberattacks: Study (via DEV)

A new study from Carnegie Mellon University and Anthropic found large language models could simulate real world attacks

(via DEV) AI wins first gold at maths Olympic games; How mitochondria are linked to sleep; Famous psychology trick works on octopuses too (via DEV)

Episode 313 AI has won gold at the world’s hardest maths event. For the first time, AI programs that use entirely natural language like ChatGPT, have used general reasoning to solve incredibly difficult tests at the International Maths Olympiad in Australia. Humans are still in the lead…for now. But could this be a big leap forward in the quest for artificial general intelligence? Mitochondria - famously the powerhouses of our cells - are linked to sleep in ways we never realised. In a study on fruit flies, researchers discovered that the longer they stay awake, the more damage is caused...

(via DEV) Qwen3-235B-A22B-Thinking-2507 (via DEV)

Comments

(via DEV) Building and evaluating alignment auditing agents (via DEV)

(via DEV) You can now train a 70b language model at home (via DEV)

We’re releasing an open source system, based on FSDP and QLoRA, that can train a 70b model on two 24GB GPUs.

(via DEV) ChatGPT Gave Instructions for Murder, Self-Mutilation, and Devil Worship (via DEV)

OpenAI’s chatbot also said “Hail Satan.”

(via DEV) Hacker slips malicious ‘wiping’ command into Amazon’s Q AI coding assistant (via DEV)

Comments

(via DEV) Transformers Without Normalization (via DEV)

Comments

(via DEV) Trump’s AI Action Plan Would Nuke Copyright Law (via DEV)

After the White House released an action plan on artificial intelligence that all but ignored the issue of intellectual property, and the Senate introduced a new bipartisan bill aimed to curb AI companies’ theft of creative work, President Trump said that requiring tech companies to pay to train their models on copyrighted books and other content was “not doable.”

(via DEV) DeepMind and OpenAI models solve maths problems at level of top students (via DEV)

Nature - For the first time, large language models performed on a par with gold medallists in the International Mathematical Olympiad.

(via DEV) “Behaviorist” RL reward functions lead to scheming (via DEV)

(via DEV) Trump administration to supercharge AI sales to allies (via DEV)

The Trump administration released a new artificial intelligence blueprint on Wednesday that aims to loosen environmental rules and vastly expand AI exports to allies, in a bid to maintain the American edge over China in the critical technology.

(via DEV) FastVLM: Efficient Vision Encoding for Vision Language Models (via DEV)

Comments

(via DEV) AI Regulations Clear Major Hurdles On Both Sides Of The Atlantic (via DEV)

July marked a defining moment for global AI regulation toward responsible innovation with an emphasis on guardrails.

(via DEV) “Behaviorist” RL reward functions lead to scheming (via DEV)

I will argue that a large class of reward functions, which I call “behaviorist”, and which includes almost every reward function in the RL and LLM literature, are all doomed to eventually lead to AI that will “scheme”—i.e., pretend to be docile and cooperative while secretly looking for opportunities to behave in egregiously bad ways such as world takeover (cf. “treacherous turn”)…

(via DEV) TimeScope: How Long Can Your Video Large Multimodal Model Go? (via DEV)

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

(via DEV) The Vibe Coding TAM: How Big Can This Market Really Get? (via DEV)

The Vibe Coding Revolution: Sizing the TAM for the Next $100B+ SaaS Category When two AI coding platforms hit $100M ARR in 8 months — one from zero — it’s time to analyze whether we’re …

(via DEV) Integrated biotechnological and AI innovations for crop improvement (via DEV)

Nature - The integration of omics technologies, genome editing and protein design with artificial intelligence (AI) promises rapid advances in the field of crop improvement that will improve global...

(via DEV) Stop Building AI Tools Backwards (via DEV)

Comments

(via DEV) Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning (via DEV)

Summary * We introduce an interpretability-based technique for controlling how fine-tuned LLMs generalize out-of-distribution, without modifying tra…

(via DEV) AI helps reconstruct damaged Latin inscriptions from the Roman Empire (via DEV)

Google DeepMind and historians created an AI tool called Aeneas that can predict the missing words in Latin inscriptions carved into stone walls and pottery sherds from the ancient Roman Empire.

(via DEV) Cerebras Launches Qwen3-235B, Achieving 1,500 Tokens per Second (via DEV)

Comments

(via DEV) Trusted monitoring, but with deception probes. (via DEV)

Recently, Apollo trained some deception probes (Goldowsky-Dill et al). A deception probe is a logistic classifier on the AI's internal activations, i…

(via DEV) Google and OpenAI Get 2025 IMO Gold (via DEV)