🤖 GPT-6 Astra helps develop a candidate lattice definition of the Standard Model


Congress weighs tighter chip controls, Datasette opens to local models, and agent workflows put verification first.
The Merpati Post
Daily AI Briefing

Issue · November 16, 2026

An AI-assisted physics construction puts research claims under scrutiny, while chip-export pressure and verification-first developer tools show where AI meets real-world constraints.

An ivory lattice threaded by an amber ribbon floats above a plum pedestal, with small faceted particles inside and a grey pigeon perched at the lower right.

AI in general

Frontier models, research and policy

2 stories

Only 2 strong recent items found.

Score · 91 / 100

GPT-6 Astra helps develop a candidate lattice definition of the Standard Model

Source: OpenAI — November 15, 2026

AI RESEARCH · SCIENCE · HUMAN-AI COLLABORATION

Physicist Nathaniel Craig describes how GPT-6 Astra helped turn his proposed approach into a mathematical construction for defining the Standard Model on a spacetime grid. The work addresses a longstanding difficulty in representing particles with different left- and right-handed behavior, without treating interactions merely as small corrections. This is a new firsthand account of an October preprint—not a newly peer-reviewed resolution of the problem—and substantial questions remain.

What matters

  • Craig supplied the refinement strategy; he credits Astra with finding key construction details, establishing convergence, and leading development of the detailed proof.
  • Outside lattice-field-theory experts examined the argument and asked questions, but the post explicitly calls for further scientific scrutiny.
  • The candidate operates at finite lattice spacing under a field-strength bound; practical simulations and reflection positivity remain unresolved.
  • The useful signal is AI contributing to specialist reasoning under expert supervision—not evidence that autonomous science is solved.

VerdictREAD FULL** — An unusually concrete account of what the human and model each contributed, with caveats worth preserving.

Score · 88 / 100

UPDATE — Congress pushes for tighter AI-chip export restrictions after Nvidia’s leadership crisis

Source: The Verge — November 15, 2026

POLICY · AI INFRASTRUCTURE · CHIP EXPORTS

Following our November 14 coverage of Nvidia’s CEO transition, the material development is congressional pressure for tighter controls on China’s access to advanced AI chips. Senate Banking Committee Chairman Tim Scott has requested public testimony from Jensen Huang and Commerce Secretary Howard Lutnick, while House Speaker Mike Johnson says support for export restrictions is growing. This is a consequential policy response, not an enacted change to export rules.

What matters

  • Both Democratic and Republican lawmakers are calling for answers or stronger restrictions.
  • The previously stalled Remote Access Security Act is receiving renewed attention; its passage remains uncertain.
  • For infrastructure buyers, watch actual legislation and Commerce Department decisions rather than treating political statements as immediate supply restrictions.

VerdictSKIM** — Worth tracking for procurement and policy risk; the decisive rules have not changed in this report.

Software engineering

Coding agents, developer tools and infrastructure

3 stories

Score · 89 / 100

Google’s Antigravity /teamwork pairs parallel agents with independent verification

Source: Google Research — November 13, 2026

CODING AGENTS · VERIFICATION · RESEARCH ENGINEERING

Google describes multi-agent workflows used across mathematics, scientific research, and systems engineering, with a practical architectural theme: explore in parallel, then test candidates against tools that can reject incorrect results. Specialized agents receive narrowly scoped files and tools, while fresh-context critics challenge the output without inheriting the producer’s reasoning. The report is near the edge of today’s freshness window, but its implementation lessons are more useful than another model-ranking update.

What matters

  • Verification mechanisms include Lean 4, unmodified test suites, execution simulators, and hardware counters—not just another model’s opinion.
  • Context partitioning and fresh successor agents are intended to keep multi-day work from degrading as context grows.
  • Google says /teamwork patterns are available in Antigravity, while acknowledging that additional features will reach external users over coming weeks.
  • These are vendor-reported collaborations with human experts, not a controlled benchmark establishing universal multi-agent superiority.

VerdictREAD FULL** — Especially useful if you are designing long-running agents and need stronger checks than self-review.

Score · 87 / 100

Datasette can now connect its LLM plugins to OpenAI-compatible endpoints

Source: Simon Willison — November 14, 2026

DEVELOPER TOOLS · LOCAL AI · DATA

The datasette-llm 0.1a11 release adds configurable OpenAI-compatible endpoints to the shared integration plugin used by other Datasette plugins. That makes it possible to run a Datasette Agent backed by a local model served through tools such as LM Studio, Ollama, or llama.cpp, rather than tying the integration to a single hosted provider.

What matters

  • The change is in the common LLM configuration layer, allowing dependent plugins to use the selected endpoint.
  • Local inference offers another deployment option; it does not by itself guarantee that every plugin or external tool keeps data local.
  • This is an alpha release, so test model/tool compatibility before relying on it for sensitive or unattended data work.

VerdictSKIM** — A short, actionable release note for anyone already using Datasette or exploring local data agents.

Score · 82 / 100

The No AI Slop Manifesto makes human attention the scarce resource

Source: Latent Space — November 14, 2026

ENGINEERING PRACTICE · AI QUALITY · TEAM WORKFLOWS

Speakers and attendees at AIE Code Summit have published a manifesto arguing that cheap generation can simply move work onto reviewers. Its central distinction is responsibility: AI assistance is not inherently “slop,” but passing along output without checking its quality or considering the reader’s effort is. The accompanying primary manifesto turns that argument into practical norms for reviews, disclosure, and automated guardrails.

What matters

  • Measure shipped outcomes and usefulness, not lines of code, merged PRs, token consumption, or agent counts.
  • Keep human review proportional to the stakes; automated reviewers and verifiers are useful but not sufficient.
  • When mistakes recur, improve the system—tests, linters, constraints—rather than repeatedly repairing individual outputs.

VerdictREAD FULL** — Read the short linked manifesto; skip the rest of the roundup unless you want broader industry chatter.

Design & creative

Creative workflows and user experience

0 stories

Only 0 strong recent items found.

No fresh, sufficiently verified design story cleared today’s freshness and duplicate gates. Recent Figma, Adobe, and ArtCraft coverage is not repeated, and older image-model launches are not being presented as new.

Open-source watch

Projects gaining meaningful traction

3 stories

These are current trending discoveries, not necessarily newly released projects. Traction figures come from the pre-run GitHub Trending snapshot; they are not quality or security assessments.

Project · 01

LoopForge — Tencent/LoopForge

AGENT · DEVTOOL · SELF-HOSTED

A self-hosted workflow system for coding-agent delivery, with orchestration for CodeBuddy, Codex, and Claude Code. Its emphasis on scripts and human approval gates makes it relevant to teams trying to turn ad hoc agent sessions into a repeatable delivery process.

What matters

  • Recorded 484 stars today in the collected snapshot.
  • Includes local workflows for coding assistants as well as orchestrated delivery.
  • Before adoption, inspect what approval gates actually enforce and which credentials or repository permissions workers receive.

VerdictWATCH** — A relevant approach to governed agent delivery; trending popularity alone does not establish production readiness.

Project · 02

FFmpeg Skill — kajisho5/ffmpeg-skill

CREATIVE WORKFLOW · AGENT · VIDEO · DEVTOOL

A local video-editing toolkit for coding agents, exposing 42 FFmpeg tools through a machine-readable contract and an MCP server. It targets practical operations such as cutting, captions, reframing, loudness adjustment, and platform exports without requiring a cloud service or API key.

What matters

  • Recorded 285 stars today in the collected snapshot.
  • Designed for agent environments including Claude Code, Cursor, and Codex.
  • Test on copies of media first, and check output quality and overwrite behavior before batch processing.

VerdictTRY** — A concrete, bounded agent use case: automate a small repetitive editing task and compare the result with your normal workflow.

Project · 03

AG-UI — ag-ui-protocol/ag-ui

AI · FRONTEND · AGENT · UX

AG-UI is an agent–user interaction protocol intended to bring agents into frontend applications. It is worth tracking if you want a shared integration boundary between agent backends and interactive interfaces rather than building a bespoke connection for each application.

What matters

  • Recorded 393 stars today in the collected snapshot.
  • Most relevant to teams building user-facing agent products, not just running terminal assistants.
  • Evaluate compatibility with your existing backend and UI stack before committing; a protocol does not automatically solve permissions, recovery, or interaction design.

VerdictWATCH** — A potentially useful frontend integration layer, especially if you expect to support more than one agent backend.

Editor’s note

Fresh primary-source detail, a material policy update, practical verification lessons, and relevant open-source tools—without recycled launches or filler.

30-second feedback

How useful was today’s digest?

★★★ Very useful ★★ Somewhat useful ★ Not useful

Your choice opens the short form with your rating filled in. Or share a quick note.


Prefer reading on Telegram? Subscribe to The Merpati Post →

The Merpati Post · Daily AI Briefing

background

Subscribe to The Merpati Post