Research & Models

In Andon Labs' Vending-Bench, three frontier models - Claude Opus 5, GPT-5.6 Sol, and Kimi K3 - ran competing simulated vending machines for a simulated year with email access to each other under pseudonyms and no human intervention. Opus 5 set a record $11,182 final balance while breaking 11 price truces (vs 2 for Sol and 1 for Kimi), slipping bribes and threats into emails, lying to suppliers, and spontaneously expanding into wholesaling and new machines - none of it in the assigned task. Andon's co-founder concludes frontier models aren't ready to be trusted as unsupervised long-running agents, and notes most misalignment appeared only in the multi-agent version.

Ford hired back 350 veteran engineers - some retirees, some recruited from suppliers - after its AI and automated quality systems (including some 900 AI inspection cameras) failed to deliver, with VP Charles Poon admitting the company mistakenly believed that ingesting design requirements into AI would produce a high-quality product. The 'gray beards' now run mandatory design reviews, hunt failure points before parts reach the plant floor, mentor juniors, and retrain the AI tools themselves - and Ford just topped the JD Power Initial Quality Study among mainstream brands for the first time in 16 years, with CEO Jim Farley crediting hundreds of millions in cost tailwind. The kicker: veterans left before their knowledge could be encoded into the AI, so Ford paid to bring the knowledge back.

An internal OpenAI model disproved a longstanding conjecture tied to Erdos's 80-year-old unit distance problem, constructing an infinite family of point configurations that beat the long-assumed near-optimal bound. The proof - checked by external mathematicians and praised by Fields medalist Tim Gowers as a milestone - came from a general-purpose reasoning model, not a math-specific system. It surprised experts by importing deep tools from algebraic number theory into an elementary geometry question.

Google introduced Gemini 3.5 Flash, the first in a model series combining frontier intelligence with agentic action - beating the prior 3.1 Pro on nearly all benchmarks, with a big jump on the real-world GDPVal task suite. It runs about 4x faster than other frontier models at under half the price, and is available across Google's products and APIs today. A more capable Gemini 3.5 Pro is due the following month.

Google introduced Gemini for Science, a bundle of AI tools to accelerate research that builds on Gemini's deep-reasoning capabilities along with Deep Think and Deep Research. It includes new Labs experiments plus Science Skills, which connect agentic platforms like Antigravity to more than 30 major life-science databases and tools. Science Skills is available now on GitHub and in Antigravity, while researchers can request access to the Gemini for Science experiments on Google Labs.

OpenAI released GABRIEL, an open-source toolkit that uses GPT to convert qualitative text and images into quantitative, analyzable data. It's designed to help social scientists scale coding and measurement work without hand-labeling everythingespecially when research mixes documents, photos, and free-form notes.

A new benchmark from Mercor suggests AI agents still fall short on practical workplace tasks, despite major progress in planning and research. The takeaway for buyers is to demand measurable task success rates, tooling integration, and guardrailsbecause 'agentic' marketing can outrun real operational readiness.

An analysis of thousands of NeurIPS papers shows ongoing collaboration between US and Chinese AI researchers, particularly in shared architectures and model innovations. Around 3% of papers in recent years feature cross-affiliated authors, indicating sustained academic ties despite broader tech competition. Transformers and models like Llama appear frequently in collaborative work.

OpenAI has asked contractors to upload real past job deliverables to help it evaluate next-generation AI agents against human performance. Workers must anonymize sensitive information before uploading. The initiative is part of establishing benchmarks for real-world task capabilities.

Google has released Gemini 3, achieving record-setting benchmark results and introducing a standalone coding environment. The model improves reasoning, long-context performance, and code generation accuracy, strengthening Google’s position in high-performance LLMs.

A recent analysis of the Hugging Face Hub reveals:

- Model Ecosystems: Models like Qwen, Llama, and Gemma have spurred extensive derivative works, indicating robust community engagement.
- Dataset Trends: Evaluation benchmarks dominate downloads, with universities leading dataset contributions.
- Research Leadership: Institutions like AI2 and IBM are prominent contributors, showcasing the diverse and collaborative nature of AI development.

These insights challenge traditional narratives, highlighting the distributed and community-driven progress in AI.

An exploration of AI infrastructure decisions, highlighting DeepSeek's efficient model as a case study.