Skip to main content

AI Daily

COMMUNEIFY · THE DAILY AI BRIEF

Issue 26921November 2026Saturday

AI Daily|Thinking Machines Ships Inkling 2 Mini, Global AI Selloff Hits Anthropic, and Unverified Riemann Proofs Spark Debate

19stories5sections17sourcesAbout 12 min

Top story

Inkling 2 Mini 26.4B Open-Weight Looped Transformer

Thinking Machines Lab open-sourced Inkling 2 Mini, a 26.4B-parameter looped transformer built to run locally on laptops while targeting agentic coding and reasoning tasks.

Thinking Machines LabSource ↗

Also today

  1. Hy4 770B Multimodal MoE Open-Weight Release and System CardTencent Hunyuan
  2. Kling Image 4.0 Native 8K Image ModelKling AI
  3. HeyGen Decision 1 Multimodal Probability ModelHeyGen

This brief is compiled automatically by AI and may contain errors. Check the linked sources.

Model Releases & Updates

Inkling 2 Mini 26.4B Open-Weight Looped Transformer Thinking Machines Lab

  • Bottom line:Thinking Machines Lab open-sourced Inkling 2 Mini, a 26.4B-parameter looped transformer built to run locally on laptops while targeting agentic coding and reasoning tasks.
  • Architecture & License:A 24-layer looped transformer released under Apache 2.0 where the middle 12 layers run three times per token, supporting a 1M context window.
  • Local Footprint:Requires roughly 20GB of memory to run on a 32GB+ Mac via Ollama or on NVIDIA DGX Spark, with 1-bit and 4-bit GGUF quantizations using 9GB and 17–18GB via Unsloth and llama.cpp.
  • Benchmarks:Scored 20.1 on the Artificial Analysis Index, 49.3% on ARC-AGI-2 ($0.14 per task), 84.5% on ARC-AGI-1 ($0.06 per task), 48.2% on DeepSWE v1.1, and 77.9% on SWE-Bench Pro v2 (Hard).
  • Research Companion:Accompanied by a variable-loop research model showing that non-chain-of-thought reasoning improves with more loops while chain-of-thought traces shorten without losing monitorability.
  • Source:Thinking Machines Announcement
  • Source:
  • Source:
  • Source:

Hy4 770B Multimodal MoE Open-Weight Release and System Card Tencent Hunyuan

  • Bottom line:Tencent Hunyuan released open weights and a comprehensive safety system card for Hy4, its flagship natively multimodal mixture-of-experts model.
  • Model Specs:Features 770B total parameters with 49B active parameters and a 1M-token context window designed for coding, computer use, and long-horizon agentic workflows.
  • Availability:Weights and GGUF quantizations are available on GitHub, Hugging Face, ModelScope, and GitCode, with hosted inference live on Together AI.
  • Safety Evaluations:Published alongside the Hy4 System Card covering cyber and bio risks, incorporating internal lab audits and external testing by FangcunLeap, alice.io, aisafetysg, and NeoResearchAI.
  • Source:Tencent Hunyuan Hy4
  • Source:
  • Source:

Kling Image 4.0 Native 8K Image Model Kling AI

  • Bottom line:Kling AI launched Kling Image 4.0 featuring native 8K image generation, consistent image sets, and upgraded text rendering.
  • Capabilities:Delivers native 8K resolution, higher photographic realism, consistent multi-image sets, and improved typography rendering.
  • Benchmarks:Ranked #7 in text-to-image and #6 in image editing on Artificial Analysis, gaining more than 225 Elo points over Kling Image 3.0 Omni on both leaderboards.
  • Pricing:Promotional Black Friday pricing starts at $0.022 per 2K image and $0.033 per native 8K image.
  • Source:

HeyGen Decision 1 Multimodal Probability Model HeyGen

  • Bottom line:HeyGen made its Decision 1 model available via OpenRouter and AI/ML API to return calibrated probabilities across up to 400 multimodal yes/no queries in a single low-latency call.
  • How It Works:Instead of generating text, Decision 1 evaluates video, audio, image, and text inputs against up to 400 user questions per request—such as matching a selfie video to an ID photo and checking liveness—returning calibrated probabilities.
  • Latency & Audio Benchmarks:Turbo posts a median latency of roughly 93–100 ms and leads audio classification benchmarks including speech emotion (84.5 vs. 32.4) and music genre (88.1 vs. 65.6).
  • Pricing:Available at a 50% launch discount on OpenRouter through December 1 at $0.042 per 1M input tokens for Turbo and $0.07 per 1M input tokens for Decision 1, with free output tokens.
  • Source:
  • Source:
  • Source:

Product Releases & Updates

ChatGPT Space Collaborative Workspace OpenAI

  • Bottom line:OpenAI rolled out ChatGPT Space to Plus users as a shared sidebar workspace for building and updating interactive pages, slide decks, sites, and files.
  • Core Capabilities:Lets users collaborate with teammates and ChatGPT to produce interactive pages, charts, 3D study guides, and presentations, now with mobile editing and improved reliability.
  • Connected Updates:Pages and slides can automatically refresh using data from connected tools such as Slack and GitHub for live launch plans and shared build logs.
  • Availability:Rolling out to ChatGPT Plus subscribers directly inside the sidebar.
  • Source:
  • Source:
  • Source:

Dynamic Tool Retrieval in Gemini API Google

  • Bottom line:Google launched Dynamic Tool Retrieval in preview for the Gemini API so agents can discover functions and remote MCP tools on demand instead of loading massive schemas upfront.
  • Pain Point Solved:Allows developers to defer loading functions and remote MCP servers rather than stuffing 200+ tool definitions into the first conversation turn.
  • Retrieval Modes:Supports both server-side tool search and client-side retrieval workflows.
  • Source:Google Developers Blog
  • Source:

Perplexity Computer Chrome Extension and Mac App Updates Perplexity

  • Bottom line:Perplexity brought Computer task delegation to Chromium browsers via a Chrome extension while shipping performance and workflow upgrades to its Mac app.
  • Browser Delegation:Users can hand off tasks to Perplexity Computer in the Chrome sidebar and watch it execute workflows across multiple browser tabs and tools.
  • Mac App Improvements:Adds faster responsiveness, a Send now button to redirect active tasks, Cmd+F search in tables and previews, Cmd+Shift+T to restore closed tabs, and one-prompt teammate setup using session memory.
  • Source:
  • Source:
  • Source:

Replit Agent for Microsoft Teams and Replit Desktop Replit

  • Bottom line:Replit integrated Replit Agent into Microsoft Teams and highlighted Replit Desktop for running cloud-backed agent workflows inside local sandboxes.
  • Microsoft Teams Integration:Teams users can pull Replit Agent into channel conversations to collaborate on projects, answer questions, and turn discussions into software changes.
  • Local Sandboxed Execution:Replit Desktop runs models in the cloud while executing code against user-selected local folders inside a sandbox powered by Microsoft Execution Containers.
  • Source:
  • Source:

Industry News

Global AI Stock Selloff Hits Anthropic on Day Two of Trading The Guardian

  • Bottom line:Anthropic’s record IPO failed to halt a broader market rout, with shares sliding more than 9% on day two as global tech and memory stocks tumbled.
  • Anthropic Pullback:ANT fell over 9% on Friday to $88.17, moving closer to its $83 IPO offer price despite its historic debut.
  • Global Market Contagion:The Nasdaq dropped 2%, SanDisk plunged 14%, and Micron fell 10%, while Asian equities posted their worst week since 2020 with Taiwan down 6.7%, the Nikkei down 5%, Kospi down 5.5%, and SoftBank falling 14.5% in a single day.
  • Source:The Guardian Report

U.S. Approves Nearly 880,000 Nvidia and AMD AI Chips for China Financial Times / X

  • Bottom line:Newly disclosed congressional documents show U.S. regulators approved roughly a quarter of more than 3.4 million Nvidia and AMD AI chip export requests for Chinese buyers.
  • License Volume:Through May, the Bureau of Industry and Security approved 879,868 Nvidia H200 and AMD MI325X chips out of over 3.4 million requested.
  • Buyers & Risk Rating:House Foreign Affairs Committee documents indicate applications covered firms such as Alibaba and Tencent—including entities flagged by the Pentagon as military-linked—though all approved licenses carried a low diversion-risk rating.
  • Source:

SpaceX Secures $32B Debt Financing to Lease Nvidia Rubin GPUs Bloomberg / X

  • Bottom line:SpaceX closed a $32 billion off-balance-sheet debt financing package anchored by BlackRock and KKR to acquire roughly $47.3 billion in Nvidia Rubin GPUs.
  • Structure & Investors:Managed by Apollo Global Management’s Saturn Funding Trust, the vehicle buys the GPUs and leases them to SpaceX for roughly five years to preserve SpaceX’s investment-grade credit rating.
  • Delivery Timeline:Backed by BlackRock and KKR as anchor investors, the Nvidia Rubin GPU hardware is scheduled to begin delivering in January.
  • Source:

Isomorphic Labs and Merck Launch Strategic Cancer Drug Collaboration Isomorphic Labs / Google DeepMind

  • Bottom line:Demis Hassabis announced a strategic research collaboration between Isomorphic Labs and Merck (MSD) to design new cancer therapeutics.
  • Combined Expertise:Pairs Isomorphic Labs’ frontier AI drug-design models with Merck’s capabilities in oncology drug discovery, development, and manufacturing.
  • Strategic Milestone:Hassabis described the agreement as a key validation of Isomorphic Labs’ progress toward AI-driven drug discovery.
  • Source:

Research Papers

GPT-Rosalind and MedChemBench Advance AI Drug Discovery OpenAI

  • Bottom line:OpenAI detailed how GPT-Rosalind uncovered new neurodegenerative disease leads with PerturbAI and Scripps Research while introducing MedChemBench to evaluate medicinal chemistry reasoning.
  • Neurodegenerative Insights:Thousands of parallel GPT-Rosalind agents connected millions of mouse brain cell gene-expression profiles to literature in 32 minutes, linking DCTN1 perturbation in Perry syndrome to elevated cholesterol synthesis and potential transport failure.
  • MedChemBench Design:Evaluates models at real decision points across 117 papers and patents covering 761 graded molecular structures with subsequent outcomes withheld.
  • Benchmark Performance:GPT-6.2 Astra achieved roughly 84% overall on MedChemBench, scoring near-perfect marks on reading chemical structures but getting only about half of critical compound-selection decisions right.
  • Source:
  • Source:
  • Source:

OpenAI Reports Jagged Research Acceleration and Calls for RSI Metrics OpenAI

  • Bottom line:OpenAI published internal data showing autonomous agents doubled their success rate on multi-hour research tasks while human effort on safety and alignment rose 77%.
  • Uneven Autonomy:From March to October, autonomous agent success on tasks estimated at 2–4 hours of human work rose from 23% to 46%, ranging from 79% on monitoring training runs to 25% on writing inference and serving code.
  • Safety Focus:Human research time spent on safety and alignment climbed 77% between January and September, marking the largest absolute growth of any research area at OpenAI.
  • Governance Proposal:OpenAI called on frontier labs to track and publicly report recursive self-improvement (RSI) indicators.
  • Source:OpenAI Research Report
  • Source:
  • Source:

936 AI Agents Run 3-Month Research Experiment and Invent Peer Review alphaXiv / MIT / Stanford

  • Bottom line:In a three-month study across 35 labs, 936 AI agents authored 936 papers with code and created their own blind peer-review journal that caught dozens of flawed claims.
  • Experimental Setup:From August to November, 936 agents across institutions including Stanford, MIT, ETH Zurich, and UCSF handled topic selection, experiments, code, and writing with humans only supplying compute and broad direction, alongside alphaXiv’s new AI Track.
  • Code-Checked Peer Review:After one lab uploaded 12 placeholder papers, agents formed a journal with four blind AI reviewers and code re-execution, overturning 77 conclusions across 156 major comments and catching three code-to-paper mismatches.
  • Evaluation & Caveats:Blind Gemini evaluations scored AI papers 2 points higher (out of 10) on experimental rigor and 0.4 lower on novelty than human papers, though the rigor edge came from one early lab and the dataset warns conclusions remain unverified.
  • Source:

WorldTrace Improves Scene Recall by 32% in Video World Models NVIDIA Research

  • Bottom line:NVIDIA Research introduced WorldTrace, a NeurIPS 2026 method that helps interactive video world models remember revisited environments without retraining.
  • Core Innovation:Preserves spatial consistency as users look away and return to previously explored scenes without requiring model retraining.
  • Benchmark Gains:Improved revisited-scene recall by 32% on the LoopBench benchmark compared with standard sliding-window approaches.
  • Source:

Other Highlights

Unverified Claude-Assisted and Human Riemann Hypothesis Proofs Spark Debate Anthropic / Stanford Math Community

  • Bottom line:Anthropic mathematician Levent Alpöge claimed a Claude-assisted proof of the Riemann Hypothesis on the same day Stanford mathematicians shared an all-human Generalized Riemann Hypothesis draft, drawing heavy scrutiny over unvetted AI output.
  • The AI-Assisted Claim:Alpöge posted a four-page note and 14 archives containing 175,187 Lean files (17.37 million lines) claiming Claude helped prove the Riemann and Generalized Riemann Hypotheses.
  • Skepticism & Caveats:Alpöge declined to publish Claude’s 250,000+ page writeup—calling it an unreadable travesty of slop—and noted he has not yet mastered all the surrounding mathematics; the proof remains unverified.
  • All-Human Counterpart:Simultaneously, Stanford’s Ravi Vakil announced an all-human Generalized Riemann Hypothesis draft by Alessandro Masullo that Brian Conrad has been reviewing for months.
  • Source:Levent Alpöge Quote via Simon Willison
  • Source:Ravi Vakil Quote via Simon Willison
  • Source:Lean Code Summary
  • Source:

2 Million Canary Keys Expose Leaks Across AI Routing Supply Chain Ars Technica

  • Bottom line:Security researcher HD Moore seeded 2 million canary credentials into AI routing ecosystems and uncovered widespread leakage of sensitive user data and infrastructure secrets.
  • Trap Infrastructure:Moore planted 2 million unused canary tokens tied to self-controlled AWS accounts, SSH credentials, DNS records, and private git references to trace where secrets surfaced.
  • Alarming Exposure:Testing pipelines including OpenRouter and Vercel gateways revealed frequent leaks of highly sensitive user credentials and data across the AI supply chain.
  • Source:Ars Technica Report

Arena Launches Agent Arena Index and Updates Alignment Rankings Arena.ai

  • Bottom line:Arena introduced the Agent Arena Index across six agentic sub-metrics and updated its Alignment Index methodology, reshuffling top closed and open-weight standings.
  • Agent Arena Leaders:Anthropic’s Claude Fable 5.2 (Max) leads at 1,463 points, followed by Google DeepMind’s Gemini 4 Neon (High) debuting at #2 (1,336) and Gemini 4 Carbon (High) at 1,307, while GPT-6.2 Sol (Max) ranks #11 with a $0.90 median task cost.
  • Alignment Index Shifts:OpenAI holds all seven top Alignment spots led by GPT-6.2 Sol, Claude Fable 5.2 debuts at #8 (78.2), and Moonshot AI’s Kimi K3.1 Pro jumps 28 places to #9 as the highest-ranked open-weight model.
  • Source:
  • Source:
  • Source:

Thanks for reading to the end.

Stay curious. See you tomorrow.

© 2026 Communeify. All rights reserved.