AI Daily｜OpenAI Achieves Automated Research Intern Milestone; Nvidia in Talks to Invest $2.5B in Thinking Machines Lab; Claude Fable 5.1 Tops Agent Arena

AI Daily｜OpenAI Achieves Automated Research Intern Milestone; Nvidia in Talks to Invest $2.5B in Thinking Machines Lab; Claude Fable 5.1 Tops Agent Arena

Model Releases & Updates

NeoMME (260M & 800M) Single-Tower Multimodal Encoders — H Company

• TL;DR: H Company unveiled NeoMME, a lightweight family of bidirectional multimodal encoders that completely eliminate separate vision towers and causal decoders in favor of unified masked diffusion pre-training.
• Key Highlights:
  • Employs a single unified Transformer tower to process multilingual text and raw $32 \times 32$ image patches interchangeably, dramatically cutting memory footprint and cross-attention latency.
  • Trained via discrete masked diffusion across vision and language modalities, outperforming conventional modular two-tower architectures on dense vision-language alignment tasks.
• Specs: 260M & 800M Parameter Sizes / Open Source / Single-Tower Bidirectional Encoder
• Links: MarkTechPost

Viggle-Animate (Local Character Replacement) — Viggle AI

• TL;DR: Viggle launched Viggle-Animate on Hugging Face, enabling users to swap characters into existing video footage in three steps on a local PC without complex pose, mask, or depth pipelines.
• Key Highlights:
  • Generates consistent character video animations from a single reference image without requiring multi-stage ComfyUI workflows (no pose extraction, face swapping, or depth conditioning needed).
  • Runs directly on consumer GPUs with significantly reduced inference overhead while preserving physical lighting and temporal motion coherence.
• Specs: Open Weights / Hugging Face Demo / Consumer GPU Compatible
• Links: Hugging Face Space |

  Thanks @cocktailpeanut, glad that Viggle-Animate held up better than expected.
  
  We're looking forward to more feedback from the community! Try here: https://t.co/nBkf9n9el1 https://t.co/AGfeS5m9qL

  — ViggleAI (@ViggleAI) September 6, 2026

Claude Fable 5.1 (Max) Captures #1 on Agent Arena — Anthropic & LMSYS

• TL;DR: Anthropic’s Claude Fable 5.1 (Max) secured the #1 overall rank on the LMSYS Agent Arena leaderboard, establishing a new Pareto frontier for real-world autonomous task execution.
• Key Highlights:
  • Achieved a +15.8% net performance leap across 6,700+ verified agentic sessions, leading all models in implicit user sentiment with a +42.5% Praise vs. Complaint margin.
  • Scored #1 in Confirmed Success (+22.4%) and #4 in Bash Recovery (+13.1%) with zero recorded tool hallucinations, operating at a median task cost of $4.14.
• Specs: Frontier Agentic Reasoner / #1 on Agent Arena / Available via Anthropic API & Claude Enterprise
• Links: Agent Arena Leaderboard |

  Claude Fable 5.1 (Max) by @AnthropicAI has landed in the Agent Arena at #1! At a $4.14 median cost per task and a +15.8% net improvement, it has reshaped the Pareto frontier!
  
  Claude Fable 5.1 is both the most performant and costly among all Agent Arena models today.
  
  Based on… https://t.co/PUoNbtild0 pic.twitter.com/x2jv6H6rJr

  — Arena.ai (@arena) September 6, 2026

Product Releases & Updates

Canvases for Agent Workflows — GitHub Copilot

• What’s New: GitHub introduced Copilot Canvases, providing developers with a persistent, interactive visual canvas to track, orchestrate, and review multi-agent executions. Instead of scrolling through infinite CLI or chat streams, Canvases aggregate file diffs, terminal execution states, and human-in-the-loop approvals on a unified surface.
• Who It’s For: Software engineers, DevOps teams, and developers managing long-running agentic coding tasks.
• Try It: GitHub Blog |

  Stop scrolling...
  
  ...through those agent chats.
  
  Canvases give your workflow a durable, shared surface to track what ran, changed, and needs review.
  
  How it works 👇https://t.co/pMwe9YhsuS

  — GitHub (@github) September 6, 2026

Gorgon Halo AI Processor (192GB Unified Memory) — AMD

• What’s New: At IFA Berlin, AMD unveiled its Gorgon Halo APU architecture, expanding unified system memory to 192GB. The platform is designed to run 300-billion-parameter LLMs entirely on local client workstations and high-end laptops, aiming to shift high-volume local reasoning away from costly cloud API billing.
• Who It’s For: Local AI developers, privacy-sensitive enterprises, and workstation power users.
• Try It: AMD Official

YoLive: Real-Time Multiplayer Video Generation — Yoroll & MiniMax

• What’s New: Yoroll launched YoLive, a live, continuous multiplayer storytelling platform powered by the open-source MiniMax H3 Superfast model (capable of synthesizing 10 seconds of high-fidelity video in under 4 seconds). Viewers submit story prompts and vote in real time, dynamically steering the generated film’s visual narrative.
• Who It’s For: Content creators, game designers, and digital communities exploring interactive generative media.
• Try It: Yo.Live

Industry News

OpenAI Reports “Automated Research Intern” Milestone; Chief Scientist Jakub Pachocki Warns of Alignment Limits

• What Happened: OpenAI published internal operational data revealing that its research agents now operate at a 3.1:1 ratio compared to human researcher workdays, accelerating per-researcher experiment velocity to 1.6x the 2025 baseline and setting a target for fully autonomous AI researchers by March 2028. Concurrently, Chief Scientist Jakub Pachocki published a reflective essay titled An Alien Mind, warning that Chain-of-Thought (CoT) monitoring effectiveness degrades as models grow more advanced. Pachocki advocated for potential voluntary scaling pauses, third-party safety verifications, and international coordination around recursive self-improvement (RSI).
• Why It Matters: Signals that frontier AI development is rapidly crossing into closed-loop automated R&D inside major labs, while prompting internal safety leadership to publicly acknowledge the widening gap between raw model capabilities and robust alignment verification.
• Source: OpenAI Research Acceleration | OpenAI: An Alien Mind

Nvidia in Advanced Talks to Invest $2.5B in Mira Murati’s Thinking Machines Lab

• What Happened: The Information reported that Nvidia is negotiating an anchor investment of approximately $2.5 billion in Thinking Machines Lab, the AI venture founded by former OpenAI CTO Mira Murati. The startup is raising a $5B–$6B funding round led by Accel at a pre-money valuation exceeding $40 billion.
• Why It Matters: Highlights relentless capital concentration around elite frontier lab spinouts, ensuring strategic silicon allocations for emerging foundation model developers while cementing enterprise customization and open-weights distribution strategies.
• Source: The Information

Seattle Times and Newsday File Federal Copyright Lawsuit Against OpenAI and Microsoft

• What Happened: Major regional news organizations The Seattle Times and Newsday jointly filed a copyright infringement lawsuit against OpenAI and Microsoft in US federal court. The complaint alleges unauthorized ingestion of copyrighted investigative reporting to train GPT models and generate verbatim excerpts within Copilot and ChatGPT search features.
• Why It Matters: Adds significant legal pressure to foundation model providers by expanding the litigation coalition beyond national outlets (such as The New York Times) to encompass hundreds of regional and local news publishers seeking mandatory licensing structures.
• Source: The Verge

Authors and Publishers Clash Over Anthropic’s $1.5B Copyright Settlement Payouts

• What Happened: Following the court approval of Anthropic’s $1.5 billion copyright settlement—which offers roughly $3,000 per infringed title across nearly 500,000 literary works—authors publicly reported disputes with major publishing houses and literary agencies attempting to claim full payouts on reverted or shared-rights contracts.
• Why It Matters: Exposes administrative and legal friction in corporate AI restitution mechanisms, highlighting how legacy intellectual property contract ambiguities complicate collective copyright settlement distribution.
• Source: TechCrunch

Jensen Huang and Greg Brockman Spark Industry Debate on 100K-GPU Clusters and AGI Benchmarks

• What Happened: Nvidia CEO Jensen Huang and OpenAI President Greg Brockman confirmed that GPT-6 Astra was trained across a cluster of over 100,000 NVIDIA Grace Blackwell NVLink72 systems, with 400,000 GPUs scheduled to come online next. Huang’s statement that “AGI has arrived” drew substantial pushback from researchers, including Gary Marcus and François Chollet, who emphasized that saturating benchmarks or scaling compute does not satisfy comprehensive general intelligence definitions (such as ARC-AGI).
• Why It Matters: Quantifies the unprecedented scale of next-generation training infrastructure while reigniting foundational debates over evaluation rigor, economic sustainability, and marketing definitions across frontier AI.
• Source:

  we're now moving into the AGI era (whether you view it as this model, the last one, or the next one), and could not do it without close partners https://t.co/DBPF7QRqyJ

  — Greg Brockman (@gdb) September 6, 2026

  |

  Sad to see Jensen claim that AGI has arrived, with no evidence and no definitions.
  
  I would urge him to read https://t.co/j6QDjPa2zg by @hendrycks, @Yoshua_Bengio & many others (including myself), and to consider how Astra is doing on the kinds of examples I laid out in my bet… https://t.co/NU0v2w7GzT

  — Gary Marcus (@GaryMarcus) September 6, 2026

Research Papers

AI Research Preference Models (RPMs): Ranking ML Experiments Before Spending GPU Compute — Meta FAIR, Oxford & UCL

• Motivation: Frontier machine learning research wastes substantial GPU compute running suboptimal, redundant, or failing hyperparameter configurations and architectures.
• Key Innovation: Meta FAIR, University of Oxford, and UCL introduced AI Research Preference Models (RPMs). RPMs act as predictive meta-judges that evaluate and rank unexecuted ML experimental proposals and code modifications, pruning unpromising branches before allocating cluster time.
• Results: Significantly increases the success rate of automated ML research pipelines by eliminating low-yield runs, achieving comparable discovery breakthroughs with a fraction of total GPU cluster hours.
• Paper: MarkTechPost

Amortizing Test-Time Reasoning into Distilled Skills for Lightweight Agents — Microsoft & Cornell University

• Motivation: Test-time reasoning and long-horizon search (e.g., $o$-series tree searches) are computationally expensive and introduce high latency during real-time agent deployments.
• Key Innovation: Researchers developed an automated framework that harvests 35–50 execution trajectories from complex reasoning runs, extracts recurring failure modes, and compiles them into structured Markdown skill libraries injected directly into the system prompts of lightweight, non-reasoning models.
• Results: Enables compact models (like GPT-5.4-mini) to recover 55% to over 100% of full test-time reasoning benchmark gains while drastically reducing inference token costs and latency.
• Paper:

  What if you could pay the reasoning cost once, then reuse what the model learned across future tasks?
  
  New Microsoft paper finds that some expensive test-time reasoning can be replaced with a small set of rules learned from previous agent runs.
  
  The paper tests a cheaper… pic.twitter.com/dMYAlCDPLs

  — Rohan Paul (@rohanpaul_ai) September 6, 2026

STAIR: Structured Generative Retrieval Grounded in Document Tables of Contents — IBM Research

• Motivation: Standard chunk-based RAG methods discard document structural hierarchies, while generative retrieval models often suffer from hallucinated internal index pointers.
• Key Innovation: IBM introduced STAIR, a generative retrieval framework that utilizes document tables of contents (ToC) as an explicit hierarchical addressing structure, forcing the model to generate indices grounded in real document topology.
• Results: Achieved an 82.6% Recall@1 on SearchTome (outperforming BM25 at 59.5% and fine-tuned Differentiable Search Indices at 76.9%) while keeping hallucination rates strictly below 0.05%.
• Paper:

  Great RAG paper from IBM.
  
  There are some really good ideas on how to solve common RAG issues.
  
  It's well known that retrievers chunk long documents by length, which discards the hierarchy the document already has.
  
  So they propose using a table of contents.
  
  A table of contents… pic.twitter.com/AewQkTWEA6

  — elvis (@omarsar0) September 6, 2026

Other Highlights

The Collapsing Software Patch Window vs. Autonomous AI Code Repair

• Overview: Venture analysis from a16z revealed that 87% of exploited software vulnerabilities are now attacked on or before public disclosure day (up from 23% in 2020). Meanwhile, an independent empirical study by 1Password evaluating over 6,000 LLM-generated security patches found that autonomous code repairs successfully resolved bugs only 26% of the time, while introducing new security vulnerabilities in 4.5% of cases. Researchers emphasized that human review remains strictly non-negotiable for AI-generated security fixes.
• Link:

  The window to patch software bugs is collapsing
  
  Of the bugs hackers actually exploit, ~87% are now being attacked on or before the day the bug is public knowledge
  
  That share was 23% in 2020
  
  Charts of the Week: https://t.co/RfYfzHpbLI pic.twitter.com/eoXU2dXoRF

  — a16z (@a16z) September 6, 2026

  |

  “The expected value of a fully LLM-generated, non-human-reviewed patch is a net-negative by a considerable margin.” https://t.co/mY4mgUgUJN

  — Gary Marcus (@GaryMarcus) September 6, 2026

claude-transplant: Open-Source Multi-Account Session Manager for Claude Desktop

• Overview: A lightweight open-source macOS menu bar and CLI utility that allows developers using multiple Claude Pro/Team accounts to migrate and preserve Claude Code Desktop session histories, transcripts, and context sidecars without losing project state during account switching.
• Link: GitHub Repository