31 stories
In the second of three posts on Opus 4.7, Mowshowitz reviews official benchmark scores, community reactions, and usage recommendations, characterising the model as a substantial improvement over Opus 4.6 for coding while noting issues with adaptive thinking and instruction following.
Stanford CRFM's IRT-based adaptive testing dynamically selects the most informative questions per model, achieving AUC-ROC of 0.85 on training sets across 22 HELM datasets while matching average-score rankings.
Stanford CRFM evaluated 10 models from five organizations on five long-context tasks capped at 128K tokens; GPT-4.1 led overall while the highest MRCR score reached only 0.256.
Stanford's CRFM and Arabic.AI released HELM Arabic, covering seven established Arabic-language evaluation tasks with full transparency into model requests and reproducible results.
Developer Simon Willison noted that ChatGPT Images 2.0 inserted editorial text into a generated image without any instruction from the user, confirmed by inspecting the original prompt.
Databricks tested GPT-5.5 on OfficeQA, its own internally developed benchmark built from 89,000 pages of US Treasury Bulletins, and reported a 46% reduction in errors on a full agent-harness workflow compared to GPT-5.4. No independent validation of the results has been published.
Databricks customers can now access GPT-5.5 on AWS, Azure, and GCP, with all model inference and Codex coding workflows governed through Unity AI Gateway's single control plane.
Adept has introduced Fuyu-Heavy, a multimodal model designed for digital agents that the company says scores higher than Gemini Pro on MMMU and outperforms it across AI2D, VQAv2, and ChartQA, while matching Claude 2.0 on MT-Bench in a chat-tuned variant.
Publication held: Ars Technica reported (citing Bloomberg) Google's potential up-to-$40B Anthropic investment, but this article cannot be published until direct Bloomberg verification or a Tier 0/1/2 source confirms the deal.
OpenAI published a system card for GPT-5.5 but the document URL returned 403 at write time and could not be independently verified.
Researchers at Apple ML Research show that a model's top-k logits can reveal task-irrelevant image information that model owners expect to be inaccessible, including cases where logit exposure matches what is accessible through direct residual stream projections.
The Verge frames OpenAI's GPT-5.5 release alongside the Anthropic rivalry and the imminent Musk trial, noting safeguard and token-efficiency claims. This article draws on a single tier-3 source; see also OpenAI's primary announcement.
OpenAI's GPT-5.5 matches GPT-5.4 per-token latency while scoring higher on coding and agentic benchmarks, and uses fewer tokens to complete the same Codex tasks.
Apple ML researchers model scene dynamics by working directly in a compressed motion embedding space derived from large-scale trajectory data, enabling efficient motion generation conditioned on text or spatial inputs.
Google is investing $10 billion now at a $350 billion Anthropic valuation, with $30 billion more contingent on performance targets, while expanding compute supply to 5 gigawatts on top of an existing TPU deal.
OpenAI released GPT-5.5 with improved agentic coding and computer use, while simultaneously upgrading Codex with browser control, document editing, and auto-review — a combination AINews describes as a superapp strategy.
Google's Simula framework decomposes synthetic dataset creation into independently controllable axes — diversity, complexity, and quality — and has already been deployed in Gemma safety models, Android scam detection, and spam filtering.
NVIDIA's developer blog describes a first-place finish in the March 2026 Kaggle Playground Series S6E3 churn prediction competition, achieved using GPT-5.4 Pro, Gemini 3.1 Pro, and Claude Opus 4.6 agents that produced 600,000 lines of code and ran 850 experiments in a human-in-the-loop workflow.
Apple presents five research highlights at ICLR 2026 in Rio de Janeiro, covering parallel RNN training, SSM length generalization, unified image understanding and generation, real-time 3D scene synthesis, and a new protein folding approach.
Google Cloud's ICLR paper introduces a structured memory framework that distills both failed and successful agent trajectories into transferable reasoning strategies, beating memory-free baselines by up to 8.3% on web benchmarks.
Accenture, Bain, BCG, Deloitte, and McKinsey gain early access to Gemini models and direct DeepMind technical engagement in a push to close a gap between enterprise AI ambition and production deployment, citing a Deloitte figure that only 25 percent of organizations have moved AI into production at scale.
AutoAdapt automates selection between RAG and fine-tuning approaches, plans adaptation strategies against real deployment constraints, and replaces manual hyperparameter search with a budget-aware refinement loop adding approximately 30 minutes and $4 in overhead.
The new open-weights models — a 284B-parameter Flash and a 1.6T-parameter Pro — introduce hybrid attention compression and FP4/FP8 mixed precision, reducing memory requirements compared to V3.2. The Register notes benchmark claims should be treated with caution.
Apple's ParaRNN framework uses Newton's method to parallelize nonlinear RNN training, achieving a 665x speedup over sequential approaches and enabling billion-scale RNN language models for the first time.
A paper from Apple ML Research introduces a benchmark covering four context understanding tasks across nine datasets, finding that pretrained dense models underperform fine-tuned alternatives and that 3-bit quantization causes uneven performance reduction across task types.
Google's Auto frame feature uses 3D scene reconstruction and generative inpainting to re-render photos from a different viewpoint, fixing parallax and perspective distortion without a reshoot.
Google DeepMind's new text-to-speech model scores 1,211 on the Artificial Analysis TTS Elo leaderboard, introduces natural-language audio tags for vocal direction, and watermarks all output with SynthID.
Google Research has released Vantage on Google Labs, a research experiment that uses an Executive LLM to steer AI avatar conversations toward targeted skill assessment, with AI scoring accuracy comparable to human expert agreement in two validation studies.
A new distributed training architecture from Google DeepMind uses asynchronous compute islands to train large models across distant locations, achieving over 20x faster synchronization than conventional methods and maintaining 88% goodput in simulated high-failure environments.
Microsoft Research and collaborators introduce ADeLe, a framework that characterizes both benchmarks and LLMs using shared capability scores, achieving approximately 88% prediction accuracy on unseen tasks for models including GPT-4o and LLaMA-3.1-405B.
A review of LinkedIn profiles by TechCrunch shows Thinking Machines Lab has hired more researchers from Meta than from any other single employer, even as Business Insider reported Meta has poached seven of TML's founding members.