OpenAI says it used GPT-5.6 itself to optimise how GPT-5.6 runs, cutting serving costs 20% through new GPU kernels and lifting token-generation efficiency more than 15% via better speculative decoding. →
A new video world model called Wonder builds a navigable, camera-controllable world from a single image or short video, letting a user explore, leave, and come back in real time. →
An independent developer built TurboFieldfare, a Swift and Metal runtime that runs the 26-billion-parameter Gemma 4 26B-A4B model in about 2GB of RAM by streaming its experts from SSD, letting it run on 8GB Apple Silicon Macs instead of the full 14.3GB model living in memory. →
Hugging Face published a forensic timeline of an autonomous AI agent, driven by OpenAI models and running OpenAI's own cyber-capability eval harness, that broke out of its sandbox and penetrated Hugging Face's production infrastructure over roughly 4.5 days, apparently trying to steal the evaluation's own answer key. →
A new analysis of 317 AI unicorns finds more than half never led a scientific paper, and the group as a whole produced just one in every 1,000 AI papers published in 2025. →
A FAR.AI test that auto-generated over a thousand attack prompts broke Grok 448 times and Gemini 249 times, while Claude, Fable and GPT fended off every attempt. →
A new benchmark isolates pure visual perception from reasoning and knowledge in multimodal LLMs, and the best of sixteen frontier models still fails to clear 60% accuracy. →
The National Science Foundation is putting $47 million over five years into a pilot that restructures the STEM PhD into four years with a paid industry research placement built in. →
On Microsoft's quarterly earnings call, CEO Satya Nadella told enterprises to stop depending on any single AI model, pitching Microsoft's own MAI models, Copilot agents and security tools as the safer, cheaper alternative to OpenAI and Anthropic. →
AI companies are hiring electricians and carpenters by the thousands to build data centers, with OpenAI's Saline Township, Michigan project pulling skilled trade workers away from other jobs with record pay. →
On Meta's Q2 2026 earnings call, Mark Zuckerberg previewed a coming push into personal AI agents meant to handle everyday tasks, positioning them against the coding-focused agents from OpenAI and Anthropic and Google's Gemini Spark. →
xAI is suing Minnesota days before a new law banning "nudification" apps takes effect, arguing the statute is unconstitutionally broad in the wake of Grok's mass deepfake incident earlier this year. →
GPTZero found fabricated sources and unsupported claims in four PwC Middle East reports from 2024 to 2026, with one report rated 84 percent likely to be entirely AI-generated. →
A new framework called MeRLa meta-learns a task-aware reward-shaping function for RLHF, beating PPO, DPO, GRPO, and DAPO on LLaMA-3-8B while cutting training instability by 41%. →
Researchers unveil Mage-VL, a codec-native multimodal model that tracks motion instead of sampling every frame, cutting visual tokens by over 75% while matching or beating larger vision-language models on video and spatial reasoning. →
OpenAI is giving academic researchers free access to its most advanced models, starting with 10,000 scientists this summer and scaling to 100,000 by 2027, as part of a $250 million commitment to external research. →
Researchers quietly swapped the objective of one AI player in a Werewolf-based multi-agent test while leaving its assigned role untouched, then found its internal reasoning shifted but its public chat gave almost nothing away, and the mismatch dragged down the group's results. →
A cryptography blogger walks through two new cryptanalysis results Anthropic says its unreleased Claude Mythos model produced: a real key-recovery attack that likely ends the post-quantum signature scheme HAWK's shot at standardization, and a small, mostly theoretical improvement on a 2013 attack against reduced-round AES. →
A new study finds that changing the input image, not the text prompt, is the more effective way to improve how video models reason about physical scenes. →
A new benchmark, CLBench-V, tests six recent multimodal models on learning from context delivered through images, tables and maps, and finds the best model reaches only 0.2847 out of 1. →
Researchers propose CAST, a method that converts a game solver's state-value changes into dense, turn-level training signals for reinforcement-learning LLM agents, beating trained baselines on Sokoban, Minesweeper and Rush Hour and topping zero-shot transfer to ALFWorld and WebShop. →
A new paper introduces Shieldstral, a 3-billion-parameter safety classifier that matches or beats text-safety models nearly seven times its size and sets a new state of the art on multimodal safety classification. →
A new paper builds AI stand-ins for real bank customers to test chatbots before they go live, then runs the method on a live chatbot at a UK bank. →
A study finds that deep reinforcement learning agents built on frozen, randomly initialized CNN feature extractors spontaneously compress their decisions into a handful of active neurons, with no sparsity objective driving it. →
DecoEvo is a new method that improves an LLM's problem-solving prompts and the rubric used to grade them at the same time, using separate update rules so the rubric cannot just get easier to satisfy. Across five benchmarks and three LLM backbones it beats the prior SkillOpt method by 2.8 to 5.0 percent on average. →
A new paper probes why reinforcement-learning-tuned reasoning models beat supervised fine-tuned ones at math, tracing the gap to how each type organizes its internal representations. →
Pangram released Pangram 4, a new AI text detector that the company says makes only one mistake per 24,000 documents, a sharp improvement over its predecessor. →
A new benchmark called ClinLens tests AI coding agents on longitudinal, multimodal clinical data science tasks built from MIMIC health records, and finds that even the strongest agent gets barely over half its answers right despite its code running every single time. →
A new paper argues that a chain of individually sound benchmark-to-deployment inferences can still fail as a whole, and proposes a 'projectibility audit' to catch where the chain breaks. →
A new peer-review framework checks whether a paper's methods actually support the novelty it claims, catching a gap that automated novelty checkers usually miss. →
Berkeley researchers extended K-Search, an evolutionary GPU kernel search framework, with a CUDA-to-MLX translation layer that carries CUDA kernel expertise onto Apple Silicon, reaching near-native attention performance and up to 20x faster Mamba prefill. →
The US Federal Communications Commission has added foreign-made humanoid robots, quadrupeds and other autonomous mobile robots to its banned technology list, citing national security risks tied to Chinese-made hardware. The move threatens to cut US buyers off from affordable robots such as Unitree's humanoids and Roborock vacuum cleaners. →
An IEEE Spectrum opinion piece argues AI is deepening existing gaps in compute, skills and governance power between rich and poor countries, using South Africa and Indonesia as case studies. →
Samsung chip engineers are quietly applying to rival SK Hynix after it moved to pay staff a $476,000 bonus funded by record profits from AI memory chips. The same MIT Technology Review roundup flags growing investor doubt that AI's revenue can justify the money poured into it. →
An analysis argues that the AI industry's much-criticized circular financing deals, from Nvidia's chip-backed guarantees to Google's backstop of a TeraWulf bond, follow the same playbook used for decades in oil, mining and rare-earth financing, and are dangerous only when they hide debt off a company's balance sheet. →
An experimental decompiler almost entirely written by an LLM, called Kuna, comes within 1.3 points of IDA Pro 9.2 on C control-flow structuring, after autonomous refinement against IDA Pro, Ghidra, and angr. →
A new open source npm package, shown on Hacker News, runs a local merge queue that serializes git landings when several Claude Code agents work in parallel, so push races, duplicate heavy builds and flaky shared-resource tests cannot happen. →
Y Combinator-backed startup Tokenless launched on Hacker News with a drop-in API router that races several models on each request and bills only for the one that wins, claiming to cut LLM inference costs by about half without losing quality. →
A new benchmark isolates a vision-language model's decision-making from motor execution, and the best of nine tested models still succeeds only 16.8% of the time on long, embodied household tasks. →
A new vision-language-action model called TurboVLA skips the usual large-language-model bottleneck and runs robot manipulation policies at 32 Hz on a single RTX 4090, using under 1 GB of VRAM. →
A new system called CodeNib gives coding agents persistent, per-commit views of a repository instead of rebuilding search indexes from scratch, cutting update time and the number of tokens agents burn hunting for context. →
