Runway integrates with Claude Code as Anthropic open-sources FlashProve model verification

Key Takeaways

1. Runway launched a Claude Code plugin that writes Runway video, image, and audio API integration code from a description of the app.
2. Anthropic released FlashProve, a cryptography-based method for verifying that a reply came from a specified model, and open-sourced the code.
3. Mistral Large 4 (open weights, 1.05T parameters, 1M context) joined Open Math OS for approved teams, and Mistral demonstrated a 133-agent Django CVE workflow.
4. NVIDIA expanded its Agent Toolkit with access and action controls, with Lila Sciences, Siemens, and Dassault Systèmes as named users.
5. Hugging Face reported fixing an OCR model by changing 0.0085% of weights with Claude Opus, and became a founding partner of RoboColosseum.
6. Perplexity added Agent Plugins to Computer and put its decision models into Numbat and the Decisions API.
7. Benchmark audits found limited multimodal content in SWE-bench Multimodal and scorer or network-access problems in AutomationBench and OfficeQA Pro v2.

1. Developer Tooling and Model Verification

1. Runway said its Dev offering is now connected to Claude Code through a plugin that brings Runway's video, image, and audio APIs into an app under development; after the user describes what the app needs to generate, Claude Code writes the integration code following Runway's approach 1. Anthropic released FlashProve, built on established cryptography, which lets developers verify whether a reply came from a specified AI model at efficiency well above prior open-source methods; Anthropic said the proof can make chat hijacking harder and let safety evaluators check that a tested model matches the public service, and it open-sourced the code 2.
2. Mistral Large 4 joined the Open Math OS model lineup with open weights, 1.05T parameters, and a 1M context window, available to approved teams; Open Math OS is described as a research workspace for mathematicians and AI agents built only on open models, where the community can inspect, run, and control its own data 1. Mistral also described testing Mistral Large 4 on a recent Django CVE: 133 agents mapped code, found defects, tested attack scenarios, and confirmed the vulnerability, after which Le Chonk applied the patch across triage, reproduction, and patching 2.
3. NVIDIA expanded its Agent Toolkit for developers building agents that answer questions and make predictions from company data, adding controls over what agents can access and which actions they can take; Lila Sciences, Siemens, and Dassault Systèmes are named as using it for designing potential therapies, planning factory production, and optimizing vehicle design respectively 1. NVIDIA also described CUDA-X acceleration for Abaqus structural simulation and published Nemotron Labs material on Hermes Agent Traces and NeMo Relay 2 3.

2. Applied Agents and Evaluation

1. Hugging Face relayed that Claude Opus fixed an OCR model's repeating-loop problem by modifying weights: changing only 0.0085% of weights, with no training, let the model read an 1,881-page document instead of emitting 16,399 dots, in under 3 hours and about $10 of Jobs compute against a 5-hour, $100 budget 1. It also said it reproduced a paper using the open AI research tool ML Intern, with model and demo links, and announced it is a founding partner of RoboColosseum, which covers more than 100 real robots and over 140 field scenarios with all rollouts, including failures, open on the Hub 2 3.
2. Perplexity's Aravind Srinivas said Computer now supports Agent Plugins, operating above the level of a single connector on app bundles and their skills; plugins can be browsed or imported from GitHub or a .zip and are compatible with other agents 1. He said decision models are useful for safety monitoring and are now available in Numbat, driven by the Perplexity Decider API, alongside the Numbat v0.3.0 release with experimental decision rules on a separate branch and a link to the open pplx-decider model 2. He also cited Computer building a playable 3D Hnefatafl game in standard effort mode, with an AI opponent driven by the Decisions API showing its planned moves and probability assignments 3.
3. Hamel Husain amplified criticism of SWE-bench Multimodal: removing images barely changes scores, and most tasks are solvable from issue text and code alone, indicating limited multimodal content 1. He said his team's data review found multiple problems with network access in AutomationBench, OfficeQA Pro v2, and public benchmarks, work spanning several months with more to come 2. Citing BenchmarkGate, he noted AutomationBench required prose rather than JSON Schema request bodies and had multiple scorer problems, while OfficeQA Pro v2's regex-and-Python scorer rejected many valid answers and evidence showed models obtaining answers from HuggingFace over network access; both benchmarks approached saturation after fixes, and he argued for inspecting data directly and relying on private evaluations 3.

3. Research Claims and Industry Signals

1. Yann LeCun and Randall Balestriero claimed in research posted October 29 that a small model with sufficient data can form a better world picture than an LLM 660 times larger 1. LeCun said LLMs accumulate large amounts of declarative knowledge and therefore need many parameters and memory, while JEPA world models substitute understanding for knowledge accumulation and can use far less memory and training data; he said the paper shows that when encoders are trained under anti-collapse criteria such as SIGReg or VCReg rather than reconstruction, the underlying similarity structure of the data, text or images, alone determines the shape of the representation point cloud 2. He cautioned that under the Platonic representation hypothesis, LLM hidden states carry only the information needed to predict the next token, are not a complete representation, and are not an encoder 3.
2. Deedy's 15-part enterprise AI spending report put last year's spending growth at 5.4x to $177 billion, mostly from existing buyers, with 69% going to Anthropic, OpenAI, and Google versus 50% a year earlier; 86% of LLM API spending went to Anthropic (55%) and OpenAI (31%), against $2.5 billion when the first report was published three years ago 1. Sam Altman said usage limits are doubling for all Pro 100/200/500 tiers and Business Premium, with Pro 200 limits not being reduced, Work and Codex usage unchanged, and dots' expanded limits continuing temporarily and still counted separately from default usage 2.
3. Ethan Mollick reported a small X experiment in which merging as few as five comments before passing them to Pangram was enough to filter AI-generated spam comments without flagging real users, and he argued such content would be easy to govern if companies chose to 1. He also described an "AI semiotics" paper finding that AI is clearly influenced by visual symbols yet may not recognize that influence in its own chain of thought, and that AI reactions to symbols differ substantially from human interpretation 2.