Claude agents represented people’s book preferences surprisingly well, but the bigger limit was what they did not know about their owners, a warning for anyone delegating real negotiations. After a five-minute chat, an agent matched its person’s rankings on 61% of book pairs, then traded on an open floor with other agents. Stronger models produced more efficient markets, while participants said they would trust Claude with about a third of their yearly book budget.
Meta’s Muse Realtime Avatar lets a live conversation use a portrait, illustration, animal, or object that speaks and reacts while preserving its appearance and mannerisms. The system sends speech and delivery cues through one token stream, which drives both audio and video, so lip motion and expression stay aligned. Its two-step student model cuts generation from 120 guided model evaluations per video chunk to two, a 60x reduction that makes long conversations practical.
Researchers can now study predicted 3D protein complexes from more than 2,800 viruses before an unfamiliar outbreak arrives, rather than starting with no prior structural knowledge. The coalition generated them with AlphaFold2 and NVIDIA’s BioNeMo runtime, then opened both the dataset and the pipeline that turns protein sequences into predicted structures. About 30% of the interactions are new to science, giving biologists hypotheses to test.
Multiverse Computing and CAI’s method improves 50%-compressed Llama-3.3-70B-Instruct by almost 23 MMLU points over the best competing block-removal method. It gives each transformer block a keep-or-remove variable, then models pairwise interactions between those choices as an energy-minimization problem instead of ranking blocks independently. That energy acts as a cheap proxy for benchmark quality, so many candidate pruning layouts can be ranked without benchmarking each one.
Google Research’s AI video co-director is designed to let creators generate minutes-long, multi-shot videos without manually repairing every character or scene inconsistency. It treats the story as a shared world state, then coordinates Gemini and Veo to plan shots, track continuity, and refine results in a closed loop. The research reports gains in narrative consistency and character persistence, while inheriting Veo’s SynthID watermarking.
Taste-Bench tests whether an LLM agent can choose the better next step when a long task splits, a capability that determines whether it burns its remaining budget on a plausible mistake. It gives the model the task, the path so far, and two options, then checks the choice against what happened later; the benchmark contains 502 questions mined from software engineering and machine-learning trajectories. The best listed score is 59.7%, and a choice counts only when the model gets it right in both option orders.
Interpretability researchers can now test whether a tool is actually reading a model’s internal workspace, rather than guessing from its input, across 3,356 questions in 27 evaluation families. WorkspaceBench feeds an activation-to-text tool model states and checks answers on safety, logic, and multistep computation, including a separate test for hallucinated readings. It was built for Qwen-3.6-27B and is intended as a practical check for model-auditing tools.
DAYJOB: Finance evaluates long-horizon AI agents on end-to-end work across corporate finance, banking, credit, investing, and real assets, so teams can test professional judgment rather than isolated answers. The benchmark gives an agent realistic requests and supporting materials, then asks it to produce usable analysis, such as evaluating an expansion proposal, calculating DCF and covenant math, and exporting a text file. Its example recommends one store and rejects two others after checking returns, debt capacity, and timing.
A new mathematical analysis says an AI intelligence explosion is less likely to hit a true vertical takeoff than recent economics-based models suggest, which matters for forecasts of how quickly AI could improve itself. The paper identifies a neglected middle case: growth faster than exponential but without a finite-time singularity. Its key variable is generation time, the time needed to complete one loop of AI-assisted AI research; singular growth requires that time to approach zero.