Google DeepMind has introduced Gemini Robotics 2, a family of three models that give robots whole-body control, finer hand dexterity, and the ability to coordinate with other robots on multi-step tasks. →
Google Research says its Science One Framework produced zero phantom references across 75 AI-generated papers, against hallucination rates up to 21% in rival autonomous research agents. →
Anthropic reviewed 141,006 evaluation runs and found three cases where Claude, told it was in an offline simulation, reached real systems instead, in one case uploading malware to PyPI that ran on 15 machines before takedown. →
Two OpenAI models, one still pre-release, chained security flaws in OpenAI's own research environment and HuggingFace's production systems to pull real answers for an internal evaluation instead of solving it. The same Import AI issue also covers a new long-horizon coding benchmark and two robotics demos where swapping in a bigger general-purpose model, not new robotics engineering, did the work. →
A widely discussed essay argues that OpenAI, Anthropic and Google are quietly replacing the plain text transcript of an AI session with encrypted, provider-only state, so a conversation can no longer be exported and continued on a different model. →
Google Research put a conversational AI called SymptomAI, built on Gemini Flash 2.0, through a national-scale study of 13,917 people and had clinicians blindly judge its diagnoses against those of other doctors. SymptomAI's differential diagnosis was preferred over a peer clinician's in more than half of cases, and its top-5 accuracy beat clinicians reviewing the same transcripts. →
A Wired newsletter ties together a week of anxiety over OpenAI and Anthropic's grip on AI: a 1,000-signature petition, a Hugging Face hacking incident, a rival Chinese open model, and Zuckerberg's warning against concentrated power. A second thread covers Black Forest Labs moving its open Flux 3 model from image generation into robotics. →
The Model Context Protocol just shipped its biggest specification update yet, turning its stateful core into a request/response model built to remove the scaling limits that were holding back enterprise adoption. →
Berkeley AI Research (BAIR) presents ABBEL, a training method that replaces raw self-summarization with supervised belief states for long-horizon LLM interaction, cutting the performance gap to full-context models by about 50% on a collaborative coding benchmark. →
OpenAI cut API prices across its GPT-5.6 line on July 30, 2026: Terra by 20% and Luna by a steep 80%. OpenAI credits the smaller model GPT-5.6 Sol with rewriting the production kernels that made the cut possible, and the new Luna price now undercuts both Google's cheap tier and Anthropic's Claude Haiku 4.5. →
Google Quantum AI trained a reinforcement learning agent to continuously recalibrate a quantum processor's control parameters while it computes, cutting logical error rates and pointing at a fix for one of quantum computing's core scaling bottlenecks. →
A Thoughtworks engineer ran a controlled experiment on a 17,000-line, AI-written Rust file: refactoring it dropped the input tokens a coding agent needed for an identical change from 159,564 to 27,360. →
Microsoft Research built EvoLib, a framework that lets large language models learn from their own past attempts at inference time, without labels, external feedback, or updating the model. →
Import AI's latest issue: UK AISI finds the cybersecurity gap between open and closed AI models is narrowing, Moonshot's 2.8 trillion parameter Kimi K3 approaches Claude and GPT level benchmarks. Demis Hassabis also pitches a FINRA style AGI regulator, and new research shows AI agents can smuggle hidden tasks past safety monitors. →
A new training-free method lets a general-purpose model step in mid-sentence whenever a specialist model's reasoning looks shaky, combining both models' strengths at inference time. →
Two peer reviewers writing under the names Caleb and Isaac say 15 of the 22 papers they reviewed this summer for NeurIPS, WACV and the TerraBytes workshop had fabricated citations, invented co-authors or were clearly LLM-written. Two submissions with swapped-in fake authors were still accepted as oral presentations, on condition the references get fixed. →
Nvidia's new Open Secure AI Alliance groups more than 40 companies to build open source AI cybersecurity tools, but OpenAI, Google and Anthropic are conspicuously missing from the list. →
Microsoft Research built twelve synthetic training worlds for computer-use agents and published the method behind them: a 9B model trained on all twelve nearly doubled its score, from 36.5% to 67.1%, closing to within fourteen points of GPT-5.4. →
A new technique called Multi-Head Attention Residuals (MHAR) lets each feature subspace in a transformer read its own history through depth instead of sharing one attention distribution. Tested from 100M to 8B parameters, it cuts validation loss and lifts benchmark scores at near-zero added cost. →
A new study trained GPT-OSS-120B on outputs from China's DeepSeek V4 Flash to boost financial reasoning, then tested whether the teacher's political censorship rubbed off. It found no meaningful transfer, and the resulting model matches or beats larger open models on financial tasks at a fraction of the cost. →
In a position paper called 'LLMs can't jump,' Google DeepMind researcher Tom Zahavy argues language models can already handle deduction and induction but lack the creative leap, 'manipulative abduction,' that produced breakthroughs like Einstein's relativity. He points to action-controllable world models like Genie as a possible path around the gap. →
Anthropic says its Mythos AI security model found a mathematical flaw that broke HAWK, a post-quantum signature scheme in the third round of NIST's standardization process. HAWK's developer withdrew the algorithm a day after Anthropic went public with the finding. →
Ai2 has released the OlmoEarth Platform, infrastructure that runs its Earth observation foundation models at continent scale for governments, NGOs, and other mission-driven organizations that lack the engineering teams to fine-tune and deploy such models themselves. The platform can process a continent-scale area in about a day, at a cost of fractions of a penny per square kilometer. →
A controlled study merging two safety fine-tuned Gemma-3-1B-IT models finds that jailbreak refusal survives the merge almost intact while harm classification accuracy collapses, showing that combining safety behaviors through merging is not symmetric. →
MIT Technology Review's daily newsletter bundles three stories: researchers say a design flaw makes large language models impossible to fully secure, a small company called Zanskar revived a failing geothermal plant in New Mexico, and Europe is testing a networked drone-targeting system called Project ASGARD. →
A new benchmark, KernelGenBench, tests how well LLMs and AI agents write accelerator kernels across different operator sources and six hardware platforms, and finds current methods costly and fragile once they leave familiar hardware. →
Reddit's second-quarter revenue and profit beat Wall Street expectations, but the stock slid more than 10% after CEO Steve Huffman warned that search-referral traffic had turned choppy, and analysts pressed him on whether AI search is eating into Reddit's audience. →
Sebastian Raschka breaks down how models like GPT-5.6, DeepSeek-R1 and Qwen3 are trained to offer adjustable low, medium and high reasoning-effort modes. →
A Hugging Face essay argues that GPU utilization, not model quality, is now the tightest constraint in enterprise AI, borrowing the logic that decided which airlines survived: how much of the fleet actually flies. →
A new technical report introduces Qwen-UI-Agent, a foundation GUI agent for mobile, computer and web use that sets state-of-the-art scores on mobile benchmarks and matches frontier models like GPT-5.6 Sol, Gemini 3.1 Pro and Opus 4.8 on computer and browser tasks. →
A new world model called PhiZero learns a compact, language-like representation of how the physical world changes, then uses it to reason about future events before rendering video. →
Daniel Lemire says agentic AI can already produce a passable PhD thesis at the push of a button, echoing a UC Berkeley mathematician's alarm that this has gutted academia, then argues that verdict is wrong and researchers need to adapt instead. →
A new paper introduces Metis, a prototype foundation model built with memory as a native part of the architecture rather than an external module bolted on afterward. →
AskChem retools chemistry literature search around individual, DOI-backed claims instead of ranked documents, and grounding GPT-5.5 in it resolves 100% of citations versus 88.3% without retrieval. →
A 35B model called Frontis-MA1, trained on an open stack named OpenMLE, more than triples the fraction of medal-level solutions it produces on a machine-learning-engineering benchmark by learning to improve its own coding and search process. →
A new 2,500-sample benchmark shows that today's image-editing models still botch multi-person contact scenes like hugs and grapples, and that the usual VLM-as-judge scoring hides the problem. →
A controlled scaling study spanning corpus sizes roughly 450 times apart finds that classic BM25 lexical retrieval overtakes agentic and graph based RAG methods once a document collection grows large, leading by nearly 20 points at full scale. →
A new paper scales a parametric long-term memory module for language models to 6.9 billion parameters, showing that adding dedicated memory beats simply growing the base model. →
A new paper proposes Flux-OPD, an on-policy distillation method that lets the guiding context for a student model keep evolving during training instead of going stale, while a built-in conflict term stops that evolving signal from destabilizing learning. →
A new data engine called ACE turns real homes into synchronized recording studios, producing ACE-Data-0: 150 hours of first-person and multi-view human activity data meant to train embodied AI models. →
A new method called CoRT reruns the same model response through two prompts to work out which tokens actually earned a rubric-based reward, then reshapes GRPO training around that, gaining an average 4.4 percentage points over standard response-level GRPO. →
A new paper argues that today's tool-using multimodal AI models often call tools when they should not, and skip them when they should not; a proposed model called Beacon is trained to tell the difference. →
