Google's Genie 4 simulates interactive worlds with sound that you can redirect by typing. For builders, new tests show agents still slip on long work: coding success falls from 62.7% to 28.5% as features evolve.
Google announced Genie 4, a world model: set an event in motion and it simulates what happens next in real time, including how the world sounds and how people in it move and speak. You can change the weather, move the camera or add a person by typing, and Gemini adapts the prompt to the scene. It is a research preview for trusted testers, coming to Google AI Ultra subscribers in Project Genie.
Anthropic published a paper by Claude claiming a deterministic algorithm that solves 3-SAT, a classic logic puzzle, faster than the Exponential Time Hypothesis allows, which would refute it. The result was discovered by an internal research Claude model, and the Lean 4 formalization, a machine-checked proof, was also written by AI models. Anthropic says the result is purely asymptotic and says nothing about solving formulas in practice. Lean checks the proofs, but not that the formal statement matches the paper.
The team behind Devin released FrontierCode 2.0, a coding benchmark graded on whether open-source maintainers would merge the agent's change. Its reference solutions modify 67% more files and 89% more lines than FrontierCode 1.1, in codebases 25% larger, and add security patches and performance work. Claude Opus 5.5 still leads but drops from 54.6% to 45.5%, while Gemini 4 Argon drops from 47.5% to 26.3%.
Fluently streamed 22.2 hours of unscripted answers from 2,042 English learners to 21 real-time speech-to-text systems. Scribe v2 Realtime at 5.4% word error rate and MAI-Transcribe-2 at 5.6% are statistically tied for first. For the median system, errors range from 6.2% for German speakers to 12.8% for Vietnamese speakers. Streaming adds a median 1.4 points over the same model run offline, and more for beginners.
FeatureArc tests coding agents on features that grow in stages, each building on the agent's own earlier code. Across 10 agents and 73 tasks of 2 to 10 stages, success fell from 62.7% at the first stage to 28.5% when the whole feature was tested on the final code. Inherited defects appeared in 72.9% of failed later stages, and agents left the earlier error unnoticed about 95% of the time.
AutoVend gives one LLM agent 20 vending machines across five site types to run through a web interface for 180 simulated days, with a shared warehouse, procurement lead times, expiring stock and operating costs. Four of five agents tested fell short of a rule-based inventory-control baseline. Ranking models by returns per dollar of inference cost ordered them differently from returns alone.
LongChang builds simulated worlds where a physical law is changed, such as heavier objects falling more slowly, and lets an LLM explore with no hints about what to investigate. Separate LLM agents generate scenes and check reports against the altered laws. Among 12 LLMs, GPT-6-Astra achieved a discovery rate of 16.7% and the other 11 scored 0%. Exploration records show anomalies left uninvestigated.
Artificial Analysis launched Intelligence Index v5, its composite score of model ability, with 60% now based on private test sets or held-out answers so models can't train on the test. Claude Opus 5.5 leads at 58, ahead of GPT-6 Astra at 57 and GPT-6.1 Sol at 55. The gap between open-weights and closed models widened from 3 points to 12, and cost per task varies by more than 13x among the leading models.
Fine-tuning on harmless-looking instruction data can quietly erode a model's safety. SimpleLoRIF compares each example's gradient, the direction it pushes the model, with the direction of unsafe behavior, and scales to datasets of around a million examples. Removing the 20% of Tulu 3 most aligned with unsafe responses produced SafeTulu, which cut GCG attack success rate by up to 22% across Qwen, Llama and Olmo while largely preserving capability.
Audit-based certification checks a random fraction of an AI agent's actions to bound how often it breaks the rules. This paper shows the bound only holds if the agent cannot tell, before it acts, whether it is being audited. On the authors' monitor, a silent audit changed no output but added latency that tool timestamps expose. A fresh random coin per audited call closes that timing channel; padding only narrows it.
Recent studies claim LLMs have functional valenced states, internal states that play the role of pleasure or pain without needing to be felt. This paper reviews that evidence and argues that key aspects of valence remain untested in LLMs. It says current experiments do not distinguish valence from preferences, value judgments or role-play, and sketches new experiments aimed at valence itself.
OpenBMB released MiniCPM-OCR 1.0, a 1.3B model that turns document photos, scans and PDFs into text, tables, formulas, charts and seals. It scores 93.68 on Real5-OmniDocBench, which tests real-world captures such as curved, skewed and recaptured pages. With DSpark decoding acceleration, its throughput is about 1.49× PaddleOCR-VL-1.6 and 3.62× MinerU2.5 Pro. It runs on vLLM, SGLang, llama.cpp and Ollama.
SuperPlane 1.0 is an open-source, model-agnostic software factory: it reads a team's backlog, scores each task for clarity, complexity and verifiability, runs coding agents, and loops them back through failing CI and review until a pull request is ready for a person. Every change returns with evidence such as passing checks and screenshots of the real UI. Over the last month, 52% of the pull requests its team merged were automated through it.
You’re caught up.
That was all 13 stories in the 28 October issue.