Monday, November 9, 2026

Miracle Makes Video Editable Live

Miracle Makes Video Editable Live

Today’s Overview

Good morning, Higgsfield is teasing a video editor that lets changes land while a clip plays, Nvidia's board is investigating intelligence claims around Jensen Huang, and researchers are probing whether agent upgrades quietly break things they used to handle. There's plenty of useful tooling too, from Gemini video updates to local web translation. Let's dive in.

Top Stories

Nvidia Board Opens Huang Intelligence Probe

Nvidia's board has opened an independent investigation after reporting that CEO Jensen Huang was targeted by a foreign intelligence effort. A committee of independent directors, supported by outside counsel, will oversee the review. The board says it has no reason to believe national security or Nvidia secrets were compromised and says Huang is cooperating.

  • The review is being handled by independent directors with outside independent counsel assisting the process.
  • The board framed the issue around U.S. national and economic security as central to Nvidia and its shareholders.
  • Nvidia said it still has confidence in Huang and his leadership team while the inquiry proceeds.

Higgsfield Previews Miracle Video Editing

Higgsfield previewed Miracle, a video editing feature that separates clips into editable layers. Users can recolor objects or swap on-screen text without regenerating the whole clip. The feature is currently limited to beta testers.

  • The product page emphasizes real-time edits that appear while the clip is playing rather than waiting in a render queue.
  • Higgsfield says users can edit one element while preserving face, light, and camera movement so the rest of the shot stays stable.
  • The page also highlights live text and layer control including pulling out an element and putting it back without visible seams.

OpenAI Raises $11 Billion Tranche

OpenAI closed an $11 billion first tranche of a funding round at a $1.25 trillion valuation. The round may reach about $20 billion in the coming weeks. It was previously shopped at a higher target of around $30 billion or more at a $1.4 trillion valuation, but was cut back as AI stocks sold off.

  • The report says the $11 billion is part of a larger round that has not yet closed.
  • The latest figure falls short of late September fundraising targets that had OpenAI seeking about $30 billion.
  • No investors were named in the input, and OpenAI did not comment on the reported financing.

Research & Analysis

Latent Sparse Attention Targets Ultra-Long Context

Latent Sparse Attention groups attention keys into clusters in a learned latent space, then lets queries search promising clusters rather than every key. The approach is designed for ultra-long contexts where full attention becomes prohibitively expensive. The researchers describe distilled and native versions, with large speedups reported on GLM and Qwen in 10-million-token tests.

  • The paper targets contexts above 1 million tokens where full attention is described as too expensive for practical use.
  • D-LSA can be added to pre-trained models through distillation instead of requiring native training from scratch.
  • N-LSA is trained on next-token loss alone and matched full attention accuracy while speeding training at 128,000 tokens.

Anthropic Explains Claude Agent Memory

Anthropic described how Claude agents use downtime to improve memory by checking and reorganizing saved notes against logs of prior work. The system treats memory as context Claude can read across sessions rather than a change to model weights. In tests, newer models handled a Pokémon task with far fewer memory files than an older model.

  • Anthropic distinguishes agent memory from in-context learning and from compaction summaries used to continue a single long conversation.
  • The system can use transcripts to correct mistakes create notes, build memory representations, or decide what to load into future sessions.
  • Anthropic frames saved memory as a prediction about what a future session will need to know.

Agent Evolution Can Hide Regressions

This study finds that agent harness evolution can improve benchmark performance without changing model weights, but the gains are not reliably clean. Across evolved agents, new task coverage came with losses on tasks that were previously handled. The authors argue for preservation-aware evaluation and revise the EvoAudit process for selecting, analyzing, mutating, and gating agents.

  • The study evaluates four methods across harnesses, general tasks, and coding tasks.
  • The authors use repeated evaluations, controlled expansion, component interventions, and admission replays to expose where gains break down.
  • The proposed EvoAudit revision changes select, analyze, mutate, and gate as a constructive follow-up to the diagnostics.

Tiny Weight Edit Boosts Qwen Math

Researchers reported that a direct 1,023-parameter weight edit improved Qwen3.5 by 29.45 points on MATH-500. The result comes from work on Weight-Space Programming, where agents directly inspect and modify model parameters. The broader finding is promising but uneven, with reliability and transfer still unresolved.

  • The study examined 16 editing trajectories across seven benchmarks and two Qwen3.5 model scales.
  • Reported held-out gains reached 36.72 points on IFEval and 17.11 points on SciBench in addition to the MATH-500 result.
  • Six of the 16 trajectories kept the original weights when candidate improvements could not be reliably validated.

Trending AI Tools

  • Gemini Omni 1.2 Flash Google's video model update extends clips to 15 seconds, cuts 720p pricing to $0.07 per second, and adds conversational editing.

  • OpenCode Browse Lets coding agents control a user's existing Chromium browser sessions while keeping interactions local.

  • Sakutto An open-source Chrome extension that translates pages on-device with the 0.8B PLaMo 3 Translate model.

Quick Hits

  • NVIDIA Cosmos 3 adds four open-weight 16B-parameter models for interactive world simulation, driving scenes, multi-sensor data, and photorealistic video from simulator controls.

  • Mistral Studio launched alongside the company's Le Chonk release.

  • Thinking Machines Lab funding is reportedly led by Accel, with a $5 billion round at a $40 billion valuation.

  • Cat Wu joins OpenAI after running product for Claude Code until October 30.

  • Inkling 2 Small is a 594B-parameter Mixture-of-Experts model with image input, interactive creation capabilities, and planned Apache 2.0 weights after safety testing.

  • Egoist AI Passport offers a way to securely share personal data with AI tools.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.