AI research, industry, tools, and the debate over where machine intelligence is taking us
A blog post tests a recent paper's claim that language models can infer a neural network's function from raw weights without tools. Running Qwen 3.8 27B locally, the author reports it identified Pac-Man, Pong, and Breakout from prompts but failed on Freeway, while noting skeptics dispute parts of the paper.
Drew Breunig has embedded a functioning emulator for the Nintendo Entertainment System's 6502 CPU directly into the weights of Qwen3.5-4B, an approach inspired by Divyansh Agrawal's qwenputer project. NES titles such as Super Mario Bros can be run inside the model, though slowly.
Five MIT faculty members and two alumni or former postdocs were named to the 2026 AI2050 Fellows cohort from Schmidt Sciences. Their projects range from turning AI-detected patterns into testable theories to training methods, public input for lawmakers, enzyme function and effects of AI reliance.
IBM released Granite Switch 4.1.1, a backwards-compatible patch with an updated 10-language Granite Speech ASR model and an improved Decision adapter. The modular design swaps adapters without retraining the base model, and an Adapter SDK is planned.
NVIDIA released Nemotron 3.5 Transcribe, an open-weight promptable model combining transcription, speaker attribution, translation and audio Q&A across 40 languages. It reports top open-model results on several leaderboards, and the release includes the Granary v2 data pipeline and fine-tuning recipes, with a medical case study showing a 45.8% relative WER cut.
— via Hugging Face
Google announced Gemini 4 Carbon, a frontier model it says leads the Artificial Analysis Intelligence Index at 59 and posts state-of-the-art cybersecurity benchmark results. Access begins with a subset of cyber defenders via the Fairwind Program, with broader release planned later.
— via Google DeepMind
MIT professor Paris Smaragdis is building machine-learning systems that can hear and interpret sound, from separating audio mixtures to supporting musicians. His path ran from Berklee and the MIT Media Lab through Mitsubishi, Adobe and Illinois, including Beatles archival audio cleanup that contributed to an Emmy win.
— via MIT News: AI, Adam Zewe | MIT News
Liquid AI introduced LFM3, a family of six roughly 2B-class models covering text, vision, audio, omni and decision tasks on one shared backbone. The 1.7B text model is reported to beat LFM2.5-2.6B with 35% fewer parameters and is built for on-device use and single-GPU fine-tuning.
— via Hugging Face
Google announced Genie 4, a real-time interactive world model that adds native audio and dialogue and is guided by Gemini models for multi-step events. It is a research preview for trusted testers, with known limits such as drift and object permanence issues.
— via Google DeepMind
Google DeepMind has revealed Genie 4, a successor to its Genie 3 world model that is said to simulate physics better. It uses Gemini 3.8 Flash as a guiding agent to evaluate frames, though worlds stay consistent for only a couple of minutes.
— via Ars Technica: AI, Ryan Whitwam
Researchers in China built Pinal, an AI tool that turns plain-language prompts into designed protein sequences. In tests only a few dozen candidates were made per target and few worked, with one of 29 fluorescent designs glowing. Hugging Face also released Carbon X, a locally runnable model that handles both DNA and English.
— via Ars Technica: AI, John Timmer
LitPivot, built with University of Pennsylvania collaborators, helps researchers revise early ideas by organizing literature around a proposal's problem, contribution and evaluation facets. In a 17-researcher study, domain experts rated LitPivot-developed ideas as better grounded in the literature than those made with a text editor and AI chat tool.
— via Hugging Face
Microsoft and Hugging Face describe using ThinkingBox's executable database-state checks as GRPO rewards to post-train Qwen3.8-27B. After 50 steps, average success rose from 51.70% to 60.78% and fully repeatable tasks went from 38 to 107, with a small drop in at-least-once coverage.
— via Hugging Face
A historian's paper in Past and Present argues that early AI pioneers deliberately injected randomness into deterministic algorithms to mimic creativity, and that AI hallucinations trace back to this research. He suggests there are unavoidable tradeoffs between creativity and predictability.
— via MIT News: AI, Adam Zewe | MIT News
MIT CSAIL and Tianjin University researchers built X-Hinges software that lets users design 3D-printed objects with embedded sensors on up to three motion axes. It embeds two different materials so sensor signals stay clear, enabling items like a robot-teaching glove, an origami lamp and a shark-shaped game controller.
— via MIT News: AI, Alex Shipps | MIT CSAIL
A plain ReAct agent running GPT-4.1, an older non-reasoning model, solves 67.5% of τ²-bench retail tasks on all four attempts, beating every retail entry in an October 12, 2026 leaderboard snapshot. The gain comes from a short list of rules the same model wrote from its own past attempts and added to the system prompt, with no weight updates.
— via Hugging Face
A model can pick the stronger candidate whether that person is Black or white and still fail a fairness test. FairStress, a new benchmark of 300 two-candidate scenarios, found that the same persuasive argument can make models more or less likely to abandon a well-supported choice depending on which demographic group it favors.
— via Hugging Face
NVIDIA's Cosmos3-H-Dreams generates surgical video that reacts live to instrument commands, trained on clinical gallbladder-removal recordings from CMR Surgical's Versius platform. By the team's account it is the first real-time generative surgical simulator of its kind, and it runs interactively on a single RTX PRO 6000 GPU.
— via Hugging Face
A new deep learning model called E(3)-Pose tracks a fetus's head as it moves during an MRI, so the slices don't end up scattered at odd angles and leave gaps that can hide brain abnormalities. It gets there by building the head's left-right symmetry into the network itself instead of making the model learn that physics from data.
— via MIT News: AI, Alex Ouyang | Abdul Latif Jameel Clinic for Machine Learning in Health
Meta's Muse Spark Open packs about 816 billion parameters but activates only around 42 billion per token, so it needs the memory of a giant model and runs at the speed of a much smaller one. The open-weight model takes text, images, audio and video, handles up to one million tokens of context, and was trained for agentic coding.
— via Hugging Face
