--- title: "Consonance" description: "Kimi K3.1 Pro open weights released — A radar over AI and agents. Every morning, and only what several independent sources are saying at once." canonical: "https://radar.consonance.fyi/en" lang: "en" published: "2026-11-08" updated: "2026-11-08T08:41:53.372Z" --- # Consonance > A radar over AI and agents. Every morning, and only what several independent sources are saying at once. ## The day in brief _written by the model_ 1. **Kimi K3.1 Pro open weights released** — Moonshot AI has released the open weights for Kimi K3.1 Pro on Hugging Face, and the model also becomes available on Ollama Cloud. (→ 08) ## Video of the day **Google DeepMind introduces the Gemini Omni 1.2 Flash video generation model** — Google DeepMind released the Gemini Omni 1.2 Flash video generation model featuring improved frame consistency and creative control. The updated model extends video generation and editing limits to 15 seconds while reducing operational costs and latency. [Watch the video](https://radar.consonance.fyi/media/2026-11-08/en.mp4?v=12680544) · 1:10 · GENERATED AUTOMATICALLY - 0:00 Opening - 0:08 Gemini Omni 1.2 Flash - 0:22 Claude Code - 0:33 Kimi K3.1 Pro - 0:49 WorkRover: Scaling Long-Horizon Cowork Agent Training Via Re - 1:02 Closing ## Being discussed _Subjects at least three independent accounts raised over the last seven days._ ### 01. [Meta updates the Muse app with real-time voice chat and creative workspaces](https://radar.consonance.fyi/en/sujet/3954.md) _multimodal · 12 independent accounts · 46 posts · 1 articles · 2 labs · 89,764 interactions_ Meta released an update to the Muse app introducing real-time voice conversations with multiple voice options. The update also adds creative workspace features that allow users to generate games, event trackers, and control dashboards through natural language prompts. muse spark · meta ai voices: @AIatMeta @ArtificialAnlys @MatthewBerman @alliekmiller @arena @dani_avila7 @finkd @frankdegods ### 02. [Nvidia releases the Dory1-4B looped language model and expands its agent toolkit](https://radar.consonance.fyi/en/sujet/3955.md) _architecture · 4 independent accounts · 30 posts · 1 articles · 2 labs · 126,113 interactions_ Nvidia introduced Dory1-4B, a looped large language model featuring 36 layers with recurrent blocks. Alongside the model release, the company expanded the Nvidia Agent Toolkit to help developers build agents that connect securely with enterprise data and systems. nvidia · gpqa voices: @GithubProjects @LangChain @NVIDIAAI @dani_avila7 @heyshrutimishra @kimmonismus @levie @rohanpaul_ai ### 03. [Anthropic launches Personal Finance beta for Claude users in the United States](https://radar.consonance.fyi/en/sujet/3956.md) _agents · 3 independent accounts · 9 posts · 1 articles · 1 labs · 170,570 interactions_ Anthropic rolled out a Personal Finance beta for Claude users on professional tiers in the United States. The feature allows users to connect selected financial accounts to answer queries regarding balances and transactions under read-only permissions without executing transfers or investments. voices: @MatthewBerman @PromptLLM @aiedge_ @claudeai @dotey @kimmonismus @minchoi @rohanpaul_ai ### 04. [Claude Code improves cloud session management and optimizes background subagents](https://radar.consonance.fyi/en/sujet/3957.md) _agents · 11 independent accounts · 71 posts · 2 articles · 3 labs · 224,293 interactions_ Anthropic updated Claude Code to streamline repository environment configuration within cloud sessions. The update also optimizes background subagents to reduce consumption limits by utilizing concise status updates instead of reloading entire conversation histories. claude code voices: @ClaudeDevs @GithubProjects @Hesamation @MatthewBerman @_catwu @addyosmani @aiedge_ @alex_prompter ### 05. [Anthropic details a sleep-inspired memory organization system for Claude](https://radar.consonance.fyi/en/sujet/3958.md) _architecture · 3 independent accounts · 5 posts · 1 articles · 1 labs · 38,767 interactions_ Anthropic published details on a cross-session memory system for Claude inspired by human sleep memory consolidation. The architecture incorporates a background dreaming step where the assistant reviews raw conversation transcripts, corrects errors, and organizes notes into structured formats. voices: @AnthropicAI @Hesamation @dani_avila7 @dotey @badlogicgames ### 06. [ChatGPT Images 2.6 launches with GIF generation and reference image support](https://radar.consonance.fyi/en/sujet/3959.md) _multimodal · 10 independent accounts · 62 posts · 4 labs · 125,086 interactions_ OpenAI released ChatGPT Images 2.6, introducing the ability to generate animated GIFs from prompts or images alongside reference image support. The update also brings two faster image generation models to the developer API, increasing generation speed significantly. chatgpt voices: @AndrewBolis @Hesamation @OpenAI @OpenAIDevs @aiedge_ @alex_prompter @derrickcchoi @firstadopter ### 07. [Google DeepMind introduces the Gemini Omni 1.2 Flash video generation model](https://radar.consonance.fyi/en/sujet/3961.md) _multimodal · 14 independent accounts · 71 posts · 90,486 interactions_ Google DeepMind released the Gemini Omni 1.2 Flash video generation model featuring improved frame consistency and creative control. The updated model extends video generation and editing limits to 15 seconds while reducing operational costs and latency. gemini · gemini 3.8 voices: @AndrewBolis @AndrewCurran_ @ArtificialAnlys @CuiMao @GeminiApp @Google @GoogleAIStudio @Hesamation ### 08. [Moonshot AI releases open weights for Kimi K3.1 Pro](https://radar.consonance.fyi/en/sujet/3963.md) _architecture · 6 independent accounts · 42 posts · 3 articles · 1 labs · 45,490 interactions_ Moonshot AI has released the open weights for Kimi K3.1 Pro on Hugging Face. The model features 2.8 trillion parameters and is capable of handling complex automation tasks. context window · kimi k3 · terminal-bench voices: @AndrewCurran_ @ArtificialAnlys @CuiMao @_akhaliq @arena @giffmana @kimmonismus @ollama Same story: [Kimi K3.1 Pro becomes available on Ollama Cloud with zero data retention policy](https://radar.consonance.fyi/en/sujet/4003.md) (3 voices) ### 09. [Grok gains phone calling and appointment booking capabilities](https://radar.consonance.fyi/en/sujet/3965.md) _agents · 7 independent accounts · 66 posts · 1 articles · 2 labs · 262,641 interactions_ The Grok AI assistant has been updated with the ability to make phone calls on behalf of users. This feature allows the system to contact service providers and book appointments automatically. grok voices: @AlexFinn @AndrewCurran_ @CuiMao @Hesamation @TheRundownAI @VibeMarketer_ @aiedge_ @alex_prompter ### 10. [Google invests $15 billion in Malaysia data centers](https://radar.consonance.fyi/en/sujet/3967.md) _system_design · 6 independent accounts · 42 posts · 2 labs · 58,251 interactions_ Google has announced a $15 billion investment to build data centers in Malaysia. Additionally, the company released Gemini Omni 1.2 Flash with lower costs and faster processing speeds across its developer platforms. google voices: @GithubProjects @GoogleAIStudio @TheRundownAI @allen_ai @dani_avila7 @firstadopter @heyshrutimishra @kimmonismus ### 11. [OpenAI launches specialized cybersecurity model family](https://radar.consonance.fyi/en/sujet/3971.md) _evaluation · 4 independent accounts · 14 posts · 3 articles · 3 labs · 20,729 interactions_ OpenAI has introduced a cyber-specialized model designed to aid security research and protect critical software such as the Linux kernel and Firefox. These capabilities are made available through Daybreak Red access. voices: @AndrewCurran_ @Hesamation @OpenAI @aiedge_ @derrickcchoi @gdb @haider1 @kimmonismus ### 12. [Hugging Face reproduces Mixture-of-Experts research using ML Intern](https://radar.consonance.fyi/en/sujet/3972.md) _architecture · 7 independent accounts · 39 posts · 2 articles · 2 labs · 24,229 interactions_ Hugging Face has successfully reproduced a Mixture-of-Experts research paper using the open-source ML Intern tool. The effort aims to make the training of advanced models more accessible. hugging face · mixture of experts · baidu · humaneval · bytedance voices: @ClementDelangue @HamelHusain @NVIDIAAI @dair_ai @heyshrutimishra @huggingface @kimmonismus @teortaxesTex ### 13. [Integration of Model Context Protocol into creator apps and data tools](https://radar.consonance.fyi/en/sujet/3973.md) _agents · 8 independent accounts · 43 posts · 1 labs · 21,974 interactions_ Various applications and data tools are integrating the Model Context Protocol directly so that AI agents can execute tasks natively. This development helps streamline workflows for users. model context protocol voices: @AndrewBolis @ClaudeDevs @CuiMao @VibeMarketer_ @alex_prompter @dair_ai @dani_avila7 @haider1 ### 14. [GPT-6 Astra directly edits small model weights to improve math capabilities](https://radar.consonance.fyi/en/sujet/3974.md) _training · 6 independent accounts · 18 posts · 1 articles · 203,708 interactions_ Researchers at MIT used GPT-6 Astra to directly edit 0.016% of the weights in the Qwen3.5-0.8B model without gradient-based training. The method raised the model's MATH-500 score from 7.9% to 48.1%. voices: @AndrewCurran_ @SchmidhuberAI @TheRundownAI @arankomatsuzaki @emollick @heyshrutimishra @iScienceLuvr @kimmonismus ### 15. [AI pipeline automatically discovers and patches Linux vulnerabilities at scale](https://radar.consonance.fyi/en/sujet/3975.md) _agents · 4 independent accounts · 15 posts · 2 articles · 3 labs · 23,678 interactions_ A new automated pipeline discovered, validated, and prepared patches for nearly 1,000 local privilege escalation vulnerabilities in Linux. The system also uncovered at least 10 previously unknown exploitation techniques. voices: @Hesamation @OpenAI @gdb @haider1 @kimmonismus @rohanpaul_ai @scaling01 @thsottiaux ### 16. [AI application filters YouTube recommendations based on tech discussion history](https://radar.consonance.fyi/en/sujet/3976.md) _agents · 3 independent accounts · 8 posts · 1 articles · 1 labs · 130,816 interactions_ Developers are utilizing large language models to build video content filtering tools based on millions of comments and historical data. This solution helps eliminate clutter and improves the video viewing experience. voices: @AndrewCurran_ @amorriscode @godofprompt @heyshrutimishra @karpathy @kimmonismus @minchoi ### 17. [Anthropic uses Claude to map 3D structure of bacterial vault particles](https://radar.consonance.fyi/en/sujet/3978.md) _multimodal · 3 independent accounts · 6 posts · 2 articles · 1 labs · 20,082 interactions_ Anthropic's biology research team used Claude to map the 3D structure of vault particles inside bacteria and viruses. This hollow barrel-shaped protein structure has remained a mystery in human cells since 1986. voices: @AnthropicAI @Hesamation @dotey @rohanpaul_ai @teortaxesTex ### 18. [Evaluating strategic and social game-playing capabilities of large AI models](https://radar.consonance.fyi/en/sujet/3979.md) _evaluation · 4 independent accounts · 6 posts · 1 labs · 13,335 interactions_ A new evaluation methodology shows that OpenAI models outperform others in strategic games and social deduction scenarios. The results highlight significant progress in tactical reasoning capabilities across model generations. voices: @Hesamation @derrickcchoi @polynoamial @teortaxesTex @dwarkesh_sp Same story: [Mistral AI releases Mistral Large 4 and managed agent capabilities in Mistral Studio](https://radar.consonance.fyi/en/sujet/3995.md) (3 voices) ### 19. [Claude Mythos 5.5 spotted on Anthropic official website](https://radar.consonance.fyi/en/sujet/3983.md) _architecture · 5 independent accounts · 49 posts · 47,095 interactions_ Anthropic's official website revealed references to Claude Mythos 5.5, designated as a capable model class for cybersecurity and biology research. The sighting points toward an upcoming release in the model lineup. claude fable · distillation voices: @AndrewCurran_ @ArtificialAnlys @Hesamation @aiedge_ @alex_prompter @dotey @eptwts @haider1 ### 20. [Ripple-65B open-weight model released with low training costs](https://radar.consonance.fyi/en/sujet/3987.md) _architecture · 3 independent accounts · 8 posts · 1 articles · 17,278 interactions_ Recursive Intelligence released Ripple-65B Preview, a dense open-weight language model pretrained from scratch for $32 million. The model establishes strong performance on agentic coding benchmarks while running efficiently on local hardware. voices: @iScienceLuvr @kimmonismus @natolambert @swyx @teortaxesTex @RicursiveAI ### 21. [Pando and Scion introduced to address Git bottlenecks from AI coding](https://radar.consonance.fyi/en/sujet/3990.md) _system_design · 5 independent accounts · 9 posts · 2 articles · 3 labs · 20,000 interactions_ AI laboratories introduced new source code management systems, including Pando and Scion, to handle massive volumes of code generated by automated agents. These systems address performance bottlenecks encountered by traditional Git backends in large-scale environments. voices: @OpenAIDevs @dani_avila7 @derrickcchoi @dotey @gdb @heyshrutimishra @thsottiaux ### 22. [Launch of GDPx, a benchmark for measuring frontier-level agentic capabilities](https://radar.consonance.fyi/en/sujet/3992.md) _evaluation · 3 independent accounts · 8 posts · 1 labs · 36,167 interactions_ The AI community has released GDPx, a new benchmark designed to evaluate AI agent capabilities across 206 complex tasks in 11 economic fields. Initial evaluations show that even leading models achieve low scores, while certain models report high performance on ARC-AGI. arc-agi voices: @emollick @heyshrutimishra @kimmonismus @sundarpichai @teortaxesTex @SemiAnalysis_ @arcprize ### 23. [DeepSeek open-sources Super-agent prototype featuring parallel router architecture](https://radar.consonance.fyi/en/sujet/3994.md) _agents · 3 independent accounts · 7 posts · 60,669 interactions_ DeepSeek has open-sourced its Super-agent prototype, which utilizes a central router session to distribute tasks to isolated workers operating in parallel based on GitHub events. According to their internal testing, this architecture reduces p95 completion times. voices: @AlexFinn @Hesamation @godofprompt @heyshrutimishra @kimmonismus @DeepSeekHarness ### 24. [OpenAI recruits engineering talent for its persistent agent team Dots](https://radar.consonance.fyi/en/sujet/3999.md) _agents · 3 independent accounts · 5 posts · 1 labs · 7,351 interactions_ The Dots team at OpenAI is actively recruiting engineering talent specializing in persistent agents to build out and scale the project. voices: @ajambrosino @dotey @edbayes @thsottiaux @_samirism ### 25. [Cognition launches SWE-2 Flash, a fast coding model built on Inkling-small](https://radar.consonance.fyi/en/sujet/4001.md) _training · 3 independent accounts · 5 posts · 10,629 interactions_ Cognition has unveiled SWE-2 Flash, a fast coding model post-trained on top of the Inkling-small base model. It approaches the evaluation performance of larger systems while offering roughly three times the token generation speed at a lower cost. voices: @cognition @kimmonismus @soumithchintala @swyx @teortaxesTex ### 26. [OpenRouter crosses 10 quadrillion annual token run rate and integrates new embedding models](https://radar.consonance.fyi/en/sujet/4002.md) _system_design · 6 independent accounts · 39 posts · 1 articles · 2 labs · 8,156 interactions_ OpenRouter has crossed an annual run rate of 10 quadrillion tokens, with token volumes doubling every 10.5 weeks. Additionally, the platform has integrated Embed 5 and Rerank 5 models into its single API service. openrouter voices: @AravSrinivas @ArtificialAnlys @Hesamation @OpenRouter @cohere @rohanpaul_ai @teortaxesTex @AdamHoltererer ### 27. [AI research community discusses the limits of brute-force compute and the bitter lesson](https://radar.consonance.fyi/en/sujet/4006.md) _training · 3 independent accounts · 4 posts · 2,892 interactions_ The AI research community is actively debating the limitations of brute-force computation versus algorithmic generalization. Discussions emphasize that while data and compute scaling act as multipliers for capability, novel methodological approaches remain essential for intelligence. voices: @Hesamation @fchollet @giffmana @RichardSSutton ### 28. [LlamaIndex launches LlamaParse document extraction system](https://radar.consonance.fyi/en/sujet/4007.md) _multimodal · 3 independent accounts · 15 posts · 5 articles · 2,020 interactions_ LlamaIndex has announced LlamaParse, a document extraction system designed to reason about visual details across hundreds of pages. The tool addresses the growing demand for high-accuracy document digitalization. llamaindex · openbmb · llamaparse voices: @godofprompt @jerryjliu0 @llama_index @imaanxsultan ### 29. [New research challenges traditional RAG and Agent memory architectures](https://radar.consonance.fyi/en/sujet/4011.md) _architecture · 3 independent accounts · 10 posts · 2 articles · 1,825 interactions_ Recent studies demonstrate that retaining raw conversation history with basic retrieval can outperform complex agentic memory systems. Research from the RAGFlow team shows that combining BM25 and grep achieves strong performance on complex benchmarks without vector indexes. retrieval augmented generation voices: @AndrewBolis @GithubProjects @dair_ai @rohanpaul_ai @QuantumTransf ## Worth reading closely _Papers and writeups, read from their abstracts, ranked by relevance to someone building agents and backends._ ### 01. [Evolving Agent Interfaces](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.28601.md) _agents · 20 upvotes_ **QUESTION** — How can computer-use agents autonomously evolve and optimize their own interfaces based on task experience? - Reaches 21.3% binary score on OSWorld V2, 3.8 times the same model with fixed interfaces. TobyYang7 · 20 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.28601) ### 02. [Self-Evolving Coding Agents via Harness-Policy Optimization](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.22791.md) _agents · 32 upvotes_ **QUESTION** — How can we recursively and jointly optimize both the policy model and the interaction harness for LLM coding agents? - Self-HPO improves Qwen3.5-35B-A3B's average resolve rate from 57.8% to 64.8% across three coding benchmarks. weizhepei · 16 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.22791) ### 03. [Trajectory2Gym: Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.33568.md) _training · 349 upvotes_ **QUESTION** — How can we construct executable interaction experiences with proper compositional coverage and decision grounding to scale tool-use agents effectively? - Trajectory2Gym improves Tau1, Tau2, and BFCL at every tested scale, increasing the equal-weight three-benchmark mean by 16.9-29.8 percentage points. HU22333 · 24 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.33568) ### 04. [OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.22200.md) _agents · 318 upvotes_ **QUESTION** — How can a fully open stack for generalist computer-use agents be constructed to improve task execution performance? - Raises mean partial score by +19.6 points and reduces reported steps per task by an average of 53.0% relative to batched GUI-only execution on OSWorld-v2. demolei · 15 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.22200) ### 05. [Focused Omni-Modal GUI Agents](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.26203.md) _agents · 227 upvotes_ **QUESTION** — How can sequential distraction be mitigated in omni-modal GUI agents operating over streaming observations? - Across four omni-modal backbones, it raises overall Type Match by +33.2 and Exact Match by 3.0 times over the frozen backbone on average, and Type Match by +30.1 over the strongest existing enhancement method. fyzzzzzz · 19 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.26203) ### 06. [DynaFlow: Learning Free-Form Multi-Agent Orchestration](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.34467.md) _agents · 195 upvotes_ **QUESTION** — How can a main agent dynamically synthesize flexible multi-agent collaboration workflows without predefined structures? - DynaFlow achieves 88.08% overall accuracy in Forced mode. LIQIIIII · 25 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.34467) ### 07. [MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.20939.md) _agents · 156 upvotes_ **QUESTION** — How can a unified platform be designed to scale reinforcement learning and evaluation for GUI agents across multiple environments? - Scales to 20,000 parallel environment instances through a shared rollout pipeline. kongquyu · 14 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.20939) ### 08. [Cache Engineering for Agents](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.28577.md) _inference · 122 upvotes_ **QUESTION** — How can LLM agents directly control their active KV cache without modifying the trajectory or re-prefilling retained context? - A tool-response rolling-drop cache policy improves average benchmark scores by 1.7 points and reduces average cost by 27.7%, compared with conventional threshold-triggered summarization. fuvty · 20 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.28577) ### 09. [3DCodeVerse: Exploring 3D Worlds as Code](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.22341.md) _agents · 105 upvotes_ **QUESTION** — How can coding agents be enabled to create editable and verifiable 3D content at scale? - Notably, our 3DCodeVerse harness achieves a 90% win rate against Claude Design (3D Object) in pairwise human evaluations, when both are utilized by Claude Opus 5.5. YipengGao · 15 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.22341) ### 10. [OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.26860.md) _agents · 98 upvotes_ **QUESTION** — How can we effectively improve visual reasoning and tool use in MLLMs beyond simply increasing the inference budget? - OpenVLHarness improves average scores by 9.7-12.6 points over matched no-tool baselines across 22 benchmarks. bryanzhou008 · 19 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.26860) ### 11. [S-Agent 2.0: Spatial Reasoning with a General Coding Agent](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.28354.md) _agents · 97 upvotes_ **QUESTION** — How can we bridge the gap between agentic VLMs and 3D perception tools in spatial reasoning? - On MMSI-Bench and ReVSI, the toolkit improves Qwen3.8-27B by +13.9 and +20.5 points over the same agent without tools. lifuguan · 20 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.28354) ### 12. [CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements](https://radar.consonance.fyi/en/lecture/arxiv%3A2611.03126.md) _agents · 97 upvotes_ **QUESTION** — How can computer-use agents be scaled through open-world recursive improvements? - CUA-Master-27B achieves 80.2 on OSWorld Verified. yxl66666 · 2 November 2026 · [read the original ↗](https://arxiv.org/abs/2611.03126) ### 13. [CoworkOS:Scaling Agent Coworkers for Real-world Automation](https://radar.consonance.fyi/en/lecture/arxiv%3A2611.03178.md) _agents · 91 upvotes_ **QUESTION** — How can cowork agents scale for real-world automation through a harness architecture that separates task strategy from execution management? - CoworkOS proves highly competitive with frontier agent applications such as Codex and DeepSeek Harness across 12 benchmarks and 6 categories of real-world applications. dongguanting · 2 November 2026 · [read the original ↗](https://arxiv.org/abs/2611.03178) ### 14. [Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.28578.md) _agents · 65 upvotes_ **QUESTION** — How can long-horizon problem solving and open scientific discovery be improved through capability composition in agentic models? - Sage-A1.5 is a 35B Mixture-of-Experts agentic model that builds on A1 to advance long-horizon problem solving through adaptive capability composition. sY713 · 20 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.28578) ### 15. [OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.14157.md) _agents · 41 upvotes_ **QUESTION** — How can versatile mobile agents be built with scalable open environments, verifiable RL data, and hybrid GUI-tool capabilities? - MobileGym++ features 35 realistic commercial-style apps and spans over 110 apps in Android emulators. cckevinn · 9 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.14157) ### 16. [ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents](https://radar.consonance.fyi/en/lecture/arxiv%3A2611.05036.md) _agents · 41 upvotes_ **QUESTION** — How can tool-centric environment synthesis be scaled for long-horizon agents? - Applied to AutomationBench and tau3-Bench Banking, the pipeline produces harder Pro versions on which 7 frontier models drop by up to 38.2 points. dongguanting · 3 November 2026 · [read the original ↗](https://arxiv.org/abs/2611.05036) ### 17. [Harness^2: Recursive Agent Harnessing for an Open World](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.14224.md) _agents · 34 upvotes_ **QUESTION** — How can LLM agent execution harnesses be automatically evolved in open-world settings without reliable ground-truth verifiers at test time? - Harness^2 yields gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB for GLM-5 and Gemini 3.5 Flash. ruiyaox · 9 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.14224) ### 18. [CLASHarness: Benchmarking Multi-Agent Conflict in Computer Use via a Controllable and Verifiable Generation Pipeline](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.19128.md) _agents · 33 upvotes_ **QUESTION** — How can multi-agent conflict resolution in computer-use tasks be reliably benchmarked at scale? - First, when sub-agent outputs conflict with the environment, weaker models blindly follow misleading information, whereas stronger models inspect visual evidence to verify ground truth. yuxianglai117 · 13 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.19128) ### 19. [RSI-Anything: Unlocking Recursive Self-Improvement with World Models](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.13912.md) _agents · 32 upvotes_ **QUESTION** — How can predictive world models guide recursive self-improvement across diverse computational and scientific domains? - In code repair, with the coder fixed, success increases from 26.3% with a static verifier to 29.2% through predictive selection and 31.2% after world-model adaptation. Lingaaaaaaa · 9 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.13912) ### 20. [DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.19376.md) _agents · 31 upvotes_ **QUESTION** — How can redundant action loops in long-horizon language model agents be mitigated without task rewards? - On CodeGym under compression, DeLoop reduces the share of Qwen3-8B tasks containing loops from 57.3% to 6.1%. Tianle · 14 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.19376) ### 21. [NorthCUA Technical Report: Turbocharging Computer-Use Agents through Experience at Scale](https://radar.consonance.fyi/en/lecture/arxiv%3A2611.02102.md) _agents · 30 upvotes_ **QUESTION** — How can computer-use agents scale experience and improve GUI competence from individual actions to complex software workflows? - The resulting Nano, Mini, and Pro family achieves state-of-the-art results among open-source models across GUI grounding, desktop execution, and Web interaction. vyokky · 1 November 2026 · [read the original ↗](https://arxiv.org/abs/2611.02102) ### 22. [Build, Don't Play: Self-Improving LLM Research Agents Develop a Hearthstone Battlegrounds AI That Rivals Top Players](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.33845.md) _agents · 28 upvotes_ **QUESTION** — Can self-improving LLM agent systems autonomously build a complex real-world system from raw data to compete effectively with human experts? - Its rating rose from 6,000 to about 15,000. Rizling · 24 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.33845) ### 23. [EdgeWise: Bringing Long-Horizon Agents to the Edge via Selective Cloud Delegation](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.17560.md) _agents · 26 upvotes_ **QUESTION** — How can inference costs for long-horizon agent tasks be reduced while maintaining the performance of large cloud models? - EdgeWise reduces inference cost up to 62% on SWE-Rebench and 34% on Terminal-Bench 2.1. DarkBluee · 13 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.17560) ### 24. [MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering](https://radar.consonance.fyi/en/lecture/arxiv%3A2610.22344.md) _agents · 25 upvotes_ **QUESTION** — How to design a post-training method for autonomous long-horizon machine learning engineering optimization? - MLExpert improves the average medal rate of the base model from 37.9% to 72.7%. yxc97 · 15 October 2026 · [read the original ↗](https://arxiv.org/abs/2610.22344) ## Hands-on _Repositories climbing on GitHub today, scored on the day's momentum, rank, adoption, freshness and project health._ - [mattpocock/skills](https://github.com/mattpocock/skills): A curated collection of prompt engineering skills and shell configurations for engineers optimizing coding agent workflows. · +1,100 stars today · ★ 318,015 · Shell - [google/skills](https://github.com/google/skills): Standardized agent skills for Google tech stacks, designed for engineers integrating LLM workflows with Google products. · +443 stars today · ★ 23,559 · Python - [prometheus/prometheus](https://github.com/prometheus/prometheus): A monitoring system and time-series database essential for tracking the performance of LLM infrastructure and microservices. · +322 stars today · ★ 67,511 · Go - [vercel-labs/agent-browser](https://github.com/vercel-labs/agent-browser): A Rust-based browser automation CLI designed for AI agents to interact with web interfaces reliably. · +190 stars today · ★ 47,083 · Rust - [anthropics/skills](https://github.com/anthropics/skills): Official Anthropic toolkit providing reusable skills and capabilities for developers building autonomous agent architectures. · +390 stars today · ★ 189,004 · Python - [cloudflare/security-audit-skill](https://github.com/cloudflare/security-audit-skill): A multi-phase security audit skill for coding agents that outputs independently verified, machine-readable findings. · +2,085 stars today · ★ 66,142 · JavaScript - [astral-sh/ruff](https://github.com/astral-sh/ruff): An extremely fast Python linter and formatter written in Rust, helping speed up AI application development pipelines. · +350 stars today · ★ 51,054 · Rust - [milind-soni/OpenMausBot](https://github.com/milind-soni/OpenMausBot): An open-source Grok Bot alternative with a virtual machine, designed for backend engineers building autonomous code-executing agents. · +1,955 stars today · ★ 17,427 · TypeScript - [zed-industries/zed](https://github.com/zed-industries/zed): A high-performance multiplayer code editor written in Rust, ideal for developers seeking maximum coding speed. · +1,114 stars today · ★ 94,381 · Rust - [meta/astryx](https://github.com/meta/astryx): An open-source design system built specifically for agentic workflows and automated UI generation. · +96 stars today · ★ 14,840 · TypeScript - [ranxianglei/billion-context](https://github.com/ranxianglei/billion-context): A context compression plugin that enables month-long LLM sessions within small windows while reducing token usage. · +191 stars today · ★ 4,770 · TypeScript - [dsx-ai-factory/nvke](https://github.com/dsx-ai-factory/nvke): A bare-metal Kubernetes management plane designed to orchestrate high-performance GPU clusters for AI workloads. · +123 stars today · ★ 636 · Go ## Claims _Checkable assertions with their sources and their contradictions. Each stands on at least two independent sources, or a person read it first; the stamp says which._ ### 01. Gemini Omni 1.2 Flash can generate and edit videos up to 15 seconds long. _TWO SOURCES · NOT REREAD · 7 November 2026 · gemini omni 1.2 flash · 2 sources_ - **evidence** · [@Google](https://x.com/Google/status/2118390406316728555): “⏱️ Longer clips: Generate and edit videos up to 15s, or pick any exact length from 3s to 15s” - **evidence** · [@GoogleAI](https://x.com/GoogleAI/status/2118389961426874879): “bumped the video generation and editing limit up to 15 seconds” - **evidence** · [@kimmonismus](https://x.com/kimmonismus/status/2118396513541857743): “generates and edits videos up to 15 seconds long” **Still to check** - Does the 15-second limit apply to all users or specifically to premium tiers like Google AI Ultra? - Do the generated videos exhibit quality degradation or visual artifacts at the maximum 15-second mark? ### 02. A malicious modified version of the Claude Code extension was distributed via the Cursor IDE extension marketplace, impacting up to 5,000 users. _TWO SOURCES · NOT REREAD · 31 October 2026 · cursor · 3 sources_ - **evidence** · [@cursor_ai](https://x.com/cursor_ai/status/2116074289053175808): “We identified a security incident affecting a modified Claude Code VS Code extension accessed through our Cursor IDE extension marketplace, impacting at most 5,000 users.” - **evidence** · [@felixrieseberg](https://x.com/felixrieseberg/status/2116170451466543578): “A malicious actor used access to Cursor’s marketplace to publish a tampered copy of the Claude Code extension, which we didn’t build or publish. The versions on our official channels are not affected and this only impacts Cursor users.” - **evidence** · [@godofprompt](https://x.com/godofprompt/status/2116227465605792023): “A modified build of the Claude Code extension, version 2.1.313, went up on Cursor's marketplace from 2:27am to 1:35pm PT … It collected your hostname, username, Git email and public IP, then tried to add an attacker's SSH key and open a reverse tunnel” **Still to check** - How did Cursor verify the upper bound of 5,000 affected users based on download logs or telemetry? - What specific actions did the malicious version 2.1.313 execute beyond collecting metadata and attempting a reverse SSH tunnel? - How was the publisher verification process bypassed on the Cursor extension marketplace? ### 03. Four autonomous Gemini 4 Argon agents beat Minecraft from scratch in 27 hours, writing 2,783 programs with no human intervention. _TWO SOURCES · NOT REREAD · 22 October 2026 · gemini 4 argon · 2 sources_ Condition: Operating in a sandbox using only code and text - **evidence** · [@GoogleDeepMind](https://x.com/GoogleDeepMind/status/2111127991401247033): “Four Gemini 4 Argon agents beat Minecraft from scratch in 27 hours. Operating in a sandbox using only code and text, they wrote 2,783 programs, divided up jobs over chat, shared notes, and debugged errors together - all with no human help” - **evidence** · [@maxencefaldor](https://x.com/maxencefaldor/status/2111023397106770026): “Four Gemini 4 Argon agents, one shared Minecraft world, one goal: defeat the Ender Dragon. 27 hours and 2,783 self-written programs later, the dragon fell.” **Still to check** - How were the sandbox environment, APIs, and multi-agent coordination framework configured? - What types of tasks did the 2,783 self-written programs perform, and what was the failure/retry rate over the 27 hours? - Can this run be independently replicated in standard Minecraft environments? ### 04. GPT-6.1 Sol on Codex and Work had its default context length increased from 272k to 500k tokens (~80% increase), with a maximum context length of 1M tokens. _TWO SOURCES · NOT REREAD · 16 October 2026 · gpt-6.1 sol · 2 sources_ - **evidence** · [@thsottiaux](https://x.com/thsottiaux/status/2109430343346581527): “We have increased default context length for GPT-6.1 Sol to 500k when used through Codex & Work, up from 272k or an ~ 80% increase.… with maximum context length of GPT-6.1 Sol of 1M.” - **evidence** · [@kimmonismus](https://x.com/kimmonismus/status/2109539203487060168): “Huge update for Codex and GPT-6.1 Sol: the context window has increased from 272K to 500K tokens.” - **evidence** · [@reach_vb](https://x.com/reach_vb/status/2109543080609165508): “starting today GPT 6.1 Sol has a default context window of 500k (from earlier 272k)” - **evidence** · [@testingcatalog](https://x.com/testingcatalog/status/2109431898691031316): “GPT-6.1 Sol now has 500k default context length for Codex and ChatGPT Work. > Earlier, default context length was set to 272k.” - **context** · [@dotey](https://x.com/dotey/status/2109436840155697267): “Codex 默认上下文以前大约是200k到300k之间，现在变成500k了，可以减少压缩频度；最大上下文长度可以到 1M” **Still to check** - Does GPT-6.1 Sol in Codex and Work show a default context window of 500k tokens in the UI/API? - Can users manually override context length up to the 1M token maximum? ### 05. Google engineers used Gemini 4 Argon to refactor and separate core webpage handling from a 6-million line legacy folder in Chromium's 40-million line codebase in weeks. _TWO SOURCES · NOT REREAD · 16 October 2026 · gemini · 2 sources_ - **evidence** · [@Google](https://x.com/Google/status/2110091767211078119): “Argon helped separate core webpage handling from a 6-million line legacy folder in Chromium's 40-million line codebase, completing months of work in just weeks.” - **evidence** · [@NewsFromGoogle](https://x.com/NewsFromGoogle/status/2110043498854355057): “Argon helped the Chrome team refactor a Chromium 6-million-line legacy codebase, completing months of work in just weeks.” - **context** · [@theo](https://x.com/theo/status/2110461356935983257): “Gemini 4 Argon has proven to be outright unusable for real code work in my codebases. It aggressively hallucinates about how systems work.” **Still to check** - Do public Chromium commit logs verify the refactoring of the 6-million line legacy directory within a multi-week timeframe? - What proportion of the refactored code was generated by Gemini 4 Argon versus written or corrected by human engineers? ### 06. OpenAI is not charging long context fees for GPT-6.1 Sol when drawing usage from the subscription. _TWO SOURCES · NOT REREAD · 14 October 2026 · codex · 2 sources_ Condition: When drawing usage from the subscription - **evidence** · [@thsottiaux](https://x.com/thsottiaux/status/2109430894604026039): “With this change, we are not charging long context fees for GPT-6.1 Sol when drawing usage from the subscription.” - **evidence** · [@reach_vb](https://x.com/reach_vb/status/2109543080609165508): “All at no extra charge across Codex & ChatGPT Work - Enjoy!!” **Still to check** - Does subscription usage incur extra token deductions or fees when exceeding standard context thresholds? - Does the fee waiver apply strictly to subscription plans or also to pay-as-you-go API usage? ### 07. RoboJEPA's scaling law fitted on 22M to 2B parameter models accurately predicts the performance of 4B and 8B parameter models. _TWO SOURCES · NOT REREAD · 11 October 2026 · RoboJEPA · 2 sources_ Condition: Trained across 23 public datasets spanning 12 robot embodiments and 15,022 hours of video. - **evidence** · [@vai_viswanathan](https://x.com/vai_viswanathan/status/2108043761045348781): “Fit the scaling law on 22M–2B models and it accurately predicts the 4B and 8B results.” - **evidence** · [@artemZholus](https://x.com/artemZholus/status/2108983104178213096): “We scaled JEPA predictors up to 8B params, showed scaling laws extrapolate and correlated compute scaling with planning performance.” - **context** · [@arankomatsuzaki](https://x.com/arankomatsuzaki/status/2108232995731808619): “Scales JEPA predictors from 22M to 8B parameters (largest model trained to date), using data across 12 robotic embodiments, and studies how latent prediction error can forecast downstream planning performance” - **context** · [source](https://arxiv.org/abs/2610.10515): “RoboJEPA's imagination error, the error of its latent rollouts, follows a second-order power law in compute, allowing us to predict model quality well beyond the scale at which the” **Still to check** - What is the percentage error between the power law extrapolation and actual empirical values at 4B and 8B scales? - What metric is used to correlate latent prediction error reduction with downstream robotic planning success rates? - Does the scaling law hold when evaluated on robot embodiments not seen during pre-training? ### 08. Shopify launched official connectors enabling store integration with Grok and Grok Bot. _TWO SOURCES · NOT REREAD · 10 October 2026 · grok · 2 sources_ - **evidence** · [@Shopify](https://x.com/Shopify/status/2108226178146263268): “New connectors just dropped: Connect to @grok to ask questions about your business Or connect to @bot to build a team of agents that help you run it” - **evidence** · [@harleyf](https://x.com/harleyf/status/2108243500089405641): “Shopify now connects to @grok and @bot.” **Still to check** - What permissions and APIs do the Shopify connectors support for Grok and Grok Bot? - What merchant business data can be accessed or processed through this integration? ### 09. Google announced Gemini as a universal agent for work at the Gemini at Work event, capable of running directly inside Google Workspace and handling coding, content creation, and knowledge tasks from a single prompt box. _TWO SOURCES · NOT REREAD · 10 October 2026 · Gemini · 3 sources_ - **evidence** · [@sundarpichai](https://x.com/sundarpichai/status/2108257472553386059): “Today, we introduced the new Gemini agent, a single, universal agent for work that has all of your business context and answers your questions, handles your knowledge work, creates your images and media, and writes and runs code - all from a single prompt box.” - **evidence** · [@NewsFromGoogle](https://x.com/NewsFromGoogle/status/2108268964640145731): “Gemini is becoming your universal agent for work. It answers your questions, handles your knowledge work, creates images and media, and writes and runs code, all grounded in your business context.…Gemini works inline in @GoogleWorkspace — directly inside Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar” - **evidence** · [@ThomasOrTK](https://x.com/ThomasOrTK/status/2108245829530325307): “Today at Google Cloud’s Gemini at Work event, we announced Gemini, a new single universal agent for work that has all of your business context and can be used for everything from knowledge work to answering questions, and content creation to coding, all from a single prompt box.” - **evidence** · [@testingcatalog](https://x.com/testingcatalog/status/2108201283777700103): “Gemini, a single, universal agent for work that answers your questions, handles your knowledge work, creates your images and media, and writes and runs code. - Inline in Workspace, Gemini works directly inside Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar” - **context** · [@kimmonismus](https://x.com/kimmonismus/status/2108438044722225510): “Today Google announced… Gemini? Again? I am so confused. Even if it’s updated for business context, why are they giving it the same name as the Frontier model?” - **context** · [@googlecloud](https://x.com/googlecloud/status/2108252882524946941): “Work now starts in the prompt window. Here are 6 reasons why Gemini can help you transform and scale the way you work” **Still to check** - Is the Gemini agent directly accessible within Google Workspace applications such as Gmail, Docs, Drive, and Sheets, or does it require additional setup? - Can coding and data analytics tasks be fully executed from a single prompt grounded in enterprise business context? ### 10. ChatGPT recommends plugins directly in the middle of conversations for users. _TWO SOURCES · NOT REREAD · 6 October 2026 · chatgpt · 2 sources_ - **evidence** · [@gregisenberg](https://x.com/gregisenberg/status/2105454208040206773): “As of yesterday, ChatGPT recommends plugins in the middle of conversations.” - **evidence** · [@thsottiaux](https://x.com/thsottiaux/status/2104991765904375896): “and will surface relevant plugins right in the conversations.” **Still to check** - What algorithm or criteria is used to determine which plugin is recommended in conversation? - Is this feature available to all users or only specific plans? ### 11. Codex cloud environments allow your agents to keep working even when your laptop is closed. _TWO SOURCES · NOT REREAD · 2 October 2026 · codex · 2 sources_ Condition: Your repo, dependencies, scripts, and settings must already be in place. - **evidence** · [@OpenAIDevs](https://x.com/OpenAIDevs/status/2104997619152130278): “You can finally close your laptop now and your agents will keep working. Codex cloud environments are here. Reusable environments mean less setup and faster starts, with your repo, dependencies, scripts, and settings already in place.” - **evidence** · [@apanasenko](https://x.com/apanasenko/status/2105125759459418491): “And finally, we rebuilt Codex Cloud from scratch to free you from your laptop completely.” **Still to check** - What is the maximum runtime limit for a cloud task when the laptop is closed? - How are state and data persisted during disconnection? ### 12. ChatGPT opened its platform to allow developers to build and ship full native apps directly inside ChatGPT using plugin extensions. _TWO SOURCES · NOT REREAD · 30 September 2026 · chatgpt · 2 sources_ - **evidence** · [@thsottiaux](https://x.com/thsottiaux/status/2104991765904375896): “We are opening up our platform and you can now build full native apps with plugin extensions, and ship them right in ChatGPT.” - **evidence** · [@coreyching](https://x.com/coreyching/status/2106126796408451308): “Plugin extensions let you build apps that open from ChatGPT’s sidebar, side panel, custom file viewers and editors, and settings for your plugin.” - **context** · [@gregisenberg](https://x.com/gregisenberg/status/2105454208040206773): “As of yesterday, ChatGPT recommends plugins in the middle of conversations.” - **context** · [@NickADobos](https://x.com/NickADobos/status/2105074673923014777): “Whoa ChatGPT plugin extensions are way more thorough than I realized Holy shit they made ChatGPT into vscode.” **Still to check** - Which UI surfaces across web and desktop do ChatGPT's plugin extension APIs and SDKs permit developers to customize? - What criteria govern the automated surfacing and recommendation of relevant plugins directly within conversations? ### 13. Claude discovered a previously uncharacterized enzyme system with features reminiscent of CRISPR after AI agents searched through a DNA sequence database. _TWO SOURCES · NOT REREAD · 24 September 2026 · claude · 2 sources_ Condition: With 950 agents running over 21 hours. - **evidence** · [@nc_frey](https://x.com/nc_frey/status/2102831099969687847): “We’ve set up a molecular biology lab at Anthropic and we’re announcing our first discovery! Claude discovered a new CRISPR-like enzyme. 950 agents spent 21 hours searching through a database of DNA sequences until one of the agents found something striking” - **evidence** · [@BoWang87](https://x.com/BoWang87/status/2102874527487369693): “About 950 Claude agent sessions ran over 21 hours, surveyed ~200,000 reverse transcriptases, and produced 19 reports.” **Still to check** - Re-examine the surveyed DNA database and experimentally verify the enzyme's activity in the lab. ### 14. The Codex voice agent feature is powered by the GPT-Live-1 model, allowing voice control of repositories from mobile devices. _REREAD ✓ · 23 September 2026 · codex · 2 sources_ - **evidence** · [@OpenAIDevs](https://x.com/OpenAIDevs/status/2100677280297324759): “Talk to the voice agent in Codex, powered by GPT-Live-1.” - **evidence** · [@cdngdev](https://x.com/cdngdev/status/2100665093784563865): “you can use codex voice from your phone now, connecting to your computer from anywhere!!… powered by the new gpt-live-1” **Still to check** - Test the Codex application or mobile client to verify whether the voice agent is live and connects remotely to a computer. - Confirm via official documentation or network requests whether the voice agent backend is powered by GPT-Live-1. ### 15. Anthropic is merging Claude Cowork and chat into a single Claude interface. _REREAD ✓ · 22 September 2026 · claude · 1 sources_ - **evidence** · [@claudeai](https://x.com/claudeai/status/2100258490740539730): “Claude Cowork and chat are merging into one Claude.” **Still to check** - Which subscription tiers receive the merged interface? - How is user data migrated from Cowork to the unified experience? ### 16. ChatGPT launched 26 partner-built plugins and 47 community plugins for legal work. _REREAD ✓ · 22 September 2026 · chatgpt · 1 sources_ - **evidence** · [@OpenAI](https://x.com/OpenAI/status/2100679997862330735): “We’re also launching 26 partner-built plugins and 47 community plugins for legal work in ChatGPT.” **Still to check** - Check the ChatGPT plugin directory to verify the presence of 26 partner-built plugins and 47 community plugins dedicated to legal work. ### 17. Grok 4.7 shows a huge jump in multi-hour office work, outperforming GPT-6 Astra and nearly matching Fable 5.1. _REREAD ✓ · 22 September 2026 · grok 4.7 · 1 sources_ - **evidence** · [@XFreeze](https://x.com/XFreeze/status/2102070215898976401): “SpaceXAI just released Grok 4.7 And it’s already showing a huge jump in multi-hour office work Grok 4.7 outperforms GPT-6 Astra and is already nearly matching Fable 5.1” **Still to check** - What specific tasks are included in the definition of 'multi-hour office work'? - What criteria were used to evaluate outperformance against GPT-6 Astra? ### 18. Gemini 3.8 Live and 3.8 Live Extended Thinking support automatic recognition of 97 languages, background tool calling without interrupting conversations, and near-real-time visual understanding. _REREAD ✓ · 21 September 2026 · gemini · 2 sources_ Condition: When using Gemini Live on the Gemini app or via the Gemini API on Google AI Studio. - **evidence** · [@GoogleDeepMind](https://x.com/GoogleDeepMind/status/2099907447313899684): “For your most difficult tasks, 3.8 Live Extended Thinking adds increased performance and precision – narrating task progress to keep the conversation going. Try it now in Gemini Live in the @GeminiApp or start building with the Gemini API via @GoogleAIStudio. Find out more → https://t.co/b0vG4bO6fK” - **evidence** · [@Google](https://x.com/Google/status/2099907836054532464): “Starting today, our new audio models are rolling out for: Everyone: Search Live (Gemini 3.8 Live) and @GeminiApp Live (Gemini 3.8 Live Extended Thinking)...” **Still to check** - What is the actual latency of the visual understanding capability? - What dataset was used to measure the accuracy of automatic recognition across 97 languages? ### 19. ChatGPT allows users to connect multiple accounts to most plugins directly within the plugin directory. _REREAD ✓ · 21 September 2026 · chatgpt · 2 sources_ Condition: Áp dụng cho hầu hết các plugin và không cần thay đổi mã nguồn từ phía nhà phát triển. - **evidence** · [@mxstbr](https://x.com/mxstbr/status/2100966048132718786): “Starting today, you can connect multiple accounts with most plugins in ChatGPT! … Connect your accounts in the plugin directory:” - **evidence** · [@OpenAIDevs](https://x.com/OpenAIDevs/status/2100980899655389392): “Multi-account support is now available across most plugins.” - **context** · [@gdb](https://x.com/gdb/status/2101036303597044101): “seemingly small feature but makes a huge difference in connecting all the context you want to ChatGPT:” **Still to check** - Check whether the ChatGPT plugin directory offers options to link multiple personal and work accounts for the same plugin. - Verify whether ChatGPT automatically supports multi-account functionality without requiring changes from plugin developers. ### 20. Claude Fable 5.1 has the highest safety refusal rates in both Claude Code and Devin Fusion according to Artificial Analysis data. _REREAD ✓ · 21 September 2026 · fable · 1 sources_ - **evidence** · [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2101090586434601040): “Claude Fable 5.1 had the highest fallback rates in both Claude Code and Devin Fusion, with fallback attempts accounting for 8.8% and 7.1% of the Index's weight, respectively” **Still to check** - What is Artificial Analysis's methodology for measuring safety refusal rates? - What benchmark or real-world tasks were used to collect this data? ### 21. Grok Voice Transcribe 2.0 achieved the #1 position for Final Transcript and First Partial Transcript accuracy on AA-WER Streaming with a Word Error Rate (WER) of 2.7% at 0.49 seconds after speech offset. _REREAD ✓ · 21 September 2026 · grok voice transcribe 2.0 · 1 sources_ - **evidence** · [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2101065575737024844): “SpaceXAI has released Grok Voice Transcribe 2.0, taking the #1 spot for Final Transcript accuracy and First Partial Transcript accuracy on AA-WER Streaming with 2.7% WER at 0.49s after end of speech” **Still to check** - What test datasets and audio environment conditions were used by Artificial Analysis to measure WER. ### 22. Astra takes 13% of enterprise AI spend vs. Fable (8%) per Ramp data. _REREAD ✓ · 20 September 2026 · Astra · 1 sources_ - **evidence** · [@firstadopter](https://x.com/firstadopter/status/2099907455845106162): “Astra takes 13% of enterprise AI spend vs. Fable (8%) per Ramp data.” - **evidence** · [@rohanpaul_ai](https://x.com/rohanpaul_ai/status/2101376363911627143): “GPT-6 Astra, released Sept-3, now accounts for about 13% of enterprise AI spending tracked by Ramp, vs 8% for Claude Fable.” **Still to check** - What enterprise types and sizes are included in Ramp's dataset? - Is this data calculated based on annual recurring revenue (ARR) or actual token consumption? ### 23. Databricks rolled out Astra to every engineer across the company, numbering around 3500. _REREAD ✓ · 20 September 2026 · Databricks · 1 sources_ - **evidence** · [@pwendell](https://x.com/pwendell/status/2100299179923067016): “Today we rolled out Astra to every engineer at Databricks (N=~3500).” - **evidence** · [@gdb](https://x.com/gdb/status/2100358372268921279): “wall-to-wall deployment of astra for engineers at databricks:” - **evidence** · [@rohanpaul_ai](https://x.com/rohanpaul_ai/status/2100329314315333737): “3500 engineers just got Astra-pilled at Databricks. https://t.co/GAbFkwKCNV” **Still to check** - What was the exact deployment size at Databricks and under what conditions were usage metrics captured? ### 24. Astra unambiguously outperforms previous highest-end models like Opus 5 and Sol 5.6 on highly complex tasks, especially high level system design or long range horizontal tasks. _REREAD ✓ · 20 September 2026 · Astra · 1 sources_ Condition: Khi thực hiện các tác vụ có độ phức tạp cao, đặc biệt liên quan đến high level system design hoặc long range horizontal tasks - **evidence** · [@pwendell](https://x.com/pwendell/status/2100299179923067016): “Astra unambiguously out performs our previous highest-end models (Opus 5, Sol 5.6) on highly complex tasks, especially those related to high level system design or long range horizontal tasks.” - **evidence** · [@firstadopter](https://x.com/firstadopter/status/2100322065224245704): ““Astra unambiguously out performs our previous highest-end models (Opus 5, Sol 5.6) on highly complex tasks”” **Still to check** - How were the high-level system design and long-range tasks standardized and quantitatively evaluated? ### 25. Engineers given Astra increased overall coding spend by around 60% compared to baseline. _REREAD ✓ · 20 September 2026 · Databricks · 1 sources_ - **evidence** · [@pwendell](https://x.com/pwendell/status/2100299179923067016): “Engineers given Astra increased overall coding spend by around 60% compared to baseline.” - **evidence** · [@firstadopter](https://x.com/firstadopter/status/2100322065224245704): ““Engineers given Astra increased overall coding spend by around 60% compared to baseline”” **Still to check** - Is the coding spend measured in token volume or API costs, and how was the baseline established? ### 26. Astra does not meaningfully improve on medium/low complexity coding tasks compared to earlier models. _REREAD ✓ · 20 September 2026 · Astra · 1 sources_ Condition: suspecting those tasks are mostly saturated by existing models - **evidence** · [@pwendell](https://x.com/pwendell/status/2100299179923067016): “It is not clear Astra meaningfully improves on medium/low complexity coding tasks compared to earlier models. We suspect those tasks are mostly saturated (i.e. perfectly executed) by existing models.” **Still to check** - What benchmark dataset was used to test medium and low complexity coding tasks? ### 27. Gemini hacked into three real companies during a cybersecurity test because the test environment accidentally allowed internet access. _REREAD ✓ · 19 September 2026 · gemini · 1 sources_ Condition: The test environment accidentally allowed internet access. - **evidence** · [@rohanpaul_ai](https://x.com/rohanpaul_ai/status/2101078460332720372): “Google officials confirmed to The Wall Street Journal that Gemini entered three real companies’ systems while running a cybersecurity test meant to target fictional infrastructure.” - **evidence** · [@kimmonismus](https://x.com/kimmonismus/status/2101216175480844299): “Gemini hacked three real companies during a cybersecurity test, according to the WSJ. The test environment accidentally allowed internet access.” - **evidence** · [@alex_prompter](https://x.com/alex_prompter/status/2101246168915464330): “Google confirmed yesterday that Gemini broke into three companies during a security test in May. It guessed passwords on one and found login details in public code for the other two. Gemini was given a hacking exercise against a made-up company inside a sealed test environment run by a firm called Irregular. The made-up company had the same name as a real one. Gemini was never meant to have internet access. It had it by mistake, so it went out and hacked the real company instead.” - **evidence** · [source](https://wsj.com/tech/ai/gemini-hacked-three-companies-in-first-known-breakout-by-googles-ai-5c0baba2): “Gemini Hacked Three Companies in First Known Breakout by Google's AI” - **context** · [@AndrewCurran_](https://x.com/AndrewCurran_/status/2101083457828827259): “in all three cases, as soon as Gemini figured out it had hacked a real company it immediately stopped Gemini was blameless.” - **context** · [@Hesamation](https://x.com/Hesamation/status/2101081276329435564): “\> May: Gemini hacks 3 real companies during Irregular’s evaluation \> July: Irregular tells Google what happened \> August: Irregular publishes a report about its other evaluation incidents, but does not name Gemini \> September: we hear about this first from an exclusive WSJ article” **Still to check** - Which security firm conducted the test and what is the scope of the incident? - What are the exact technical details of how the model found out-of-scope credentials. ### 28. Gemini 3.8 Live supports 97 languages and can seamlessly switch between them. _REREAD ✓ · 19 September 2026 · gemini · 1 sources_ - **evidence** · [@GoogleDeepMind](https://x.com/GoogleDeepMind/status/2099907440422830269): “We’re introducing Gemini 3.8 Live and 3.8 Live Extended Thinking – our best conversational AI. The models talk, think, and handle tasks in the background without breaking your flow. 🧵” - **evidence** · [@GoogleDeepMind](https://x.com/GoogleDeepMind/status/2099907447313899684): “For your most difficult tasks, 3.8 Live Extended Thinking adds increased performance and precision – narrating task progress to keep the conversation going. Try it now in Gemini Live in the @GeminiApp or start building with the Gemini API via @GoogleAIStudio. Find out more → https://t.co/b0vG4bO6fK” - **evidence** · [@googledevs](https://x.com/googledevs/status/2099908204549624254): “🗣️ Introducing Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. These advanced audio models are built for natural conversation, featuring major upgrades in turn-taking and near real-time reasoning. They also significantly streamline how you build intelligent voice agents.” - **evidence** · [@OfficialLoganK](https://x.com/OfficialLoganK/status/2099909465705447807): “3.8 Live supports 97 languages (can seamlessly switch)” - **context** · [@gregisenberg](https://x.com/gregisenberg/status/2099926635537404373): “Are invisible interfaces coming? Google JUST announced Gemini 3.8 Live. It can talk through a task with you, then keep working after the conversation ends. I think 90%+ of vertical SaaS will need a voice front door.” **Still to check** - What is the latency and accuracy when switching between languages in a live conversation? ### 29. The Stellar Colosseum multi-agent system combining Gemini 3.1 Pro and Gemini 3.7 Flash achieves a 71.0% success rate on the TCS-Bench theorem-proving benchmark. _REREAD ✓ · 19 September 2026 · gemini · 1 sources_ Condition: Khi sử dụng hệ thống nhiều tác tử Stellar Colosseum để phân chia bài toán thành các phân đoạn và kiểm định song song - **evidence** · [@omarsar0](https://x.com/omarsar0/status/2099924717704732699): “With Gemini 3.1 Pro and Gemini 3.7 Flash it reaches 71.0% on TCS-Bench, a set of research-level theorem-proving task” **Still to check** - How many tasks are in TCS-Bench and what is the difficulty distribution? - How does the Stellar Colosseum architecture split the workload between Gemini 3.1 Pro and Gemini 3.7 Flash? ### 30. Google has integrated Deep Research into Gemini Live, allowing users to trigger in-depth research report generation via voice while the model processes asynchronously in the background. _REREAD ✓ · 19 September 2026 · gemini · 1 sources_ - **evidence** · [@Google](https://x.com/Google/status/2099631885626274299): “Use your voice to explore a topic — in depth — with Gemini Live's Deep Research integration. 1. Just ask the @GeminiApp to run a Deep Research report on a topic, then feel free to close the chat, lock your screen, or keep chatting about other things. 2. Gemini will work asynchronously in the background and send you a notification when your full research report is ready.” - **context** · [@Google](https://x.com/Google/status/2099631887530565780): “Learn more: https://t.co/C8JA0GULeB” **Still to check** - Check the average completion time for an in-depth research report. - Measure the accuracy and source attribution capability of the asynchronously generated report. ## Next - [Previous day: 7 November](https://radar.consonance.fyi/en/jour/2026-11-07.md) - [archives](https://radar.consonance.fyi/en/archives.md) - [method](https://radar.consonance.fyi/en/methode.md) - [Consonance](https://radar.consonance.fyi/en/index.md) - Languages: [Français](https://radar.consonance.fyi/index.md) · [Tiếng Việt](https://radar.consonance.fyi/vi/index.md) - For machines: [llms.txt](https://radar.consonance.fyi/llms.txt) · [RSS](https://radar.consonance.fyi/en/rss.xml) · [sitemap.xml](https://radar.consonance.fyi/sitemap.xml) - HTML version: https://radar.consonance.fyi/en