Thursday, November 5, 2026

Astra Learns to Rewrite AI

Astra Learns to Rewrite AI

Today’s Overview

Good morning, AI is moving in some strange new directions: voters are pushing back on the data centers powering the boom, OpenAI is embedding deeper into Abu Dhabi's government and industry, and MIT researchers say Astra can edit model weights directly. Meanwhile, video editing and robotics papers are pushing longer context and tighter perception into action. Let's dive in.

Top Stories

Voters Push Back on Data Centers

Voters backed limits on data centers in nearly every local measure that appeared on U.S. midterm ballots. The results point to growing public resistance around AI infrastructure and signal tougher scrutiny ahead for the industry.

  • The tracker maps each ballot outcome into two camps, Restrict and Allow, so differently worded local measures can be compared directly.
  • Its scope includes bans, megawatt caps, zoning referendums, developer agreements, voter-approval requirements, and advisory questions about whether to allow data centers.
  • The data is compiled from Ballotpedia tracking, MultiState's September roundup, county election offices, and local reporting as of mid-October 2026.

Astra Rewrites AI Models From the Inside

MIT researchers published work showing OpenAI's GPT-6 Astra can improve some AI models by directly rewriting their weights instead of retraining them. The result hints at a path toward self-improving models and more automated AI research.

  • The authors call the approach Weight-Space Programming, using language models to reason about and modify neural network weights directly.
  • In one test, Astra read 17,253 weights with no tools and identified the game a small network had learned to play.
  • For a stuck game policy, Astra edited 3.5% of the weights in one shot and moved success from 0% to 100%.

OpenAI Deepens Its Abu Dhabi Push

OpenAI signed four agreements in Abu Dhabi spanning government, energy, investment, and enterprise deployments. The deals extend OpenAI's role in public-sector AI and major enterprise workflows across the emirate.

  • The meeting centered on using advanced AI to accelerate government transformation and innovation across Abu Dhabi's strategic sectors.
  • Abu Dhabi framed the effort around becoming the world's first fully AI-native government with AI integrated across services and economic sectors.
  • The agreements were signed across government, energy and investment sectors after the meeting with Sam Altman.

Research & Analysis

MoviEdit Tackles Long-Form Video Editing

MoviEdit is an end-to-end framework for long-form, multi-shot video editing that tackles data, modeling, and evaluation together. It introduces automated training data generation, a cascaded bidirectional-autoregressive architecture, and a benchmark for systematic evaluation.

  • The paper was published on Nov. 2 and listed as Hugging Face's number three paper of the day.
  • The author list includes researchers affiliated with Adobe on the Hugging Face paper page.
  • The project points readers to a dedicated site for complete results and benchmark access.

AI Agents Run a Quantum Spin Experiment

A Tokyo lab of AI agents is said to have run a full quantum physics experiment and discovered a new law. The source frames it as a notable result for agentic experimentation, though validation would determine its broader significance.

  • The arXiv paper is titled Collaborative AI agents for autonomous experimental research with solid-state quantum spins.
  • The agents could generate and reject research questions, require theoretical checks, and redirect investigations when evidence failed to support a claim.
  • When an initial experiment showed signal drift, the agents treated it as inconclusive and revised the measurement protocol without relaxing acceptance criteria.

TurnWise Improves Streaming Video RL

TurnWise is a turn-level GRPO framework for streaming video understanding. Instead of assigning one advantage to an entire trajectory, it gives credit to individual observe-or-respond decisions to reduce redundant replies and improve follow-up behavior.

  • The paper was published on Nov. 4 and ranked as Hugging Face's number one paper of the day.
  • Its core setting requires models to repeatedly choose whether to respond or keep observing as a video stream unfolds.
  • The authors say the method improves learning without a learned critic or additional rollouts.

NEO-VLA Rethinks Robot Vision

NEO-VLA replaces a fixed visual encoder with a native vision-language-action design where image and language features evolve jointly through the backbone. The paper argues that embodied control benefits from access to visual representations across layers, not just a final encoded output.

  • The paper was published on Nov. 4 and listed as Hugging Face's number two paper of the day.
  • The motivation is that manipulation depends on visual cues such as local geometry, boundaries, depth, and subtle state changes.
  • The model uses lightweight patch convolutional layers before a flow-matching action expert draws from representations at different depths.

Trending AI Tools

  • Kimi K3.1 Pro Moonshot AI released public weights for a 2.8 trillion-parameter open-weight model with a 1 million token context window.

  • Grok Bot phone calls Paid users can have the agent sit on hold, book appointments, and negotiate bills on U.S. numbers.

  • ChatGPT Images 2.6 OpenAI added GIF output, reference images, faster generation, markup, SVG export, and new API models.

Quick Hits

  • OpenAI and Cathedral are partnering on an internal security sprint using OpenAI models to challenge defenses.

  • OpenAI to acquire Pierre bringing the startup behind Codex code review tools and its Code Storage service into the platform.

  • ChatGPT search growth has reportedly reached about 20% of Google's search volume, with voice mode at 250 million weekly active users.

  • Conductor joins Meta as the AI coding startup brings its multi-agent management platform into the company.

  • Claude legal tools add ready-made workflows for contract review, cite checks, and regulatory scans.

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.