首页 分类 维基 关于
SITEMAP 订阅 EN
AI Twitter Hots · 2026-11-03 · 第 053 期 · 10 条 2026 年 11 月 3 日 周二 ·

Weight-Space Programming goes mega-viral, DeepSeek Harness ships Creator Mode, UNCUT minimal harness clears 500★ in a day

Today's highlights

  1. Models editing models dominate the timeline: MIT CSAIL-linked discussion pushes Weight-Space Programming — GPT-6 Astra reading/writing small-model weights; Neel Nanda and others frame it as scaled-up classic mech-interp, not spooky magic.
  2. Harnesses keep productizing and going minimal: DeepSeek Harness ships Creator Mode for user-written plugins; UNCUT hits 500+ stars in a day on "no silent cuts"; Chinese timelines debate Claude Code self-orchestration vs Grok Bot outsourcing.
  3. Everyday agent friction: Antigravity's monthly Mission Report, T3 Code / Magpie recirculating, an unconfirmed Fable 5.5 rumor, and a 13.8GB agent disk cleanup.

1. Weight-Space Programming: Astra edits Qwen weights directly

Key points: @rikiyatakehi posts with MIT CSAIL collaborators that GPT-6 Astra edited Qwen3.5-0.8B weights to improve math: only ~0.016% of parameters, no gradient training, fixing reasoning loops, with math up to ~6× and instruction following ~3×. They call it Weight-Space Programming — frontier LLMs natively reading/writing neural net weights. @kimmonismus relays MATH-500 7.9%→48.1%; @NeelNanda5 argues Astra looks surprisingly good at old-school mechanistic interpretability rather than anything spooky, and stays cautious on unlimited scaling.

Why it matters: Highest-engagement model-research thread of the day (likes/bookmarks in the tens of thousands). Treat the model as a weight editor — and pair the hype with the blog plus skeptical reads before treating the demo as a training replacement.

Weight-Space Programming

2. DeepSeek Harness: Creator Mode — turn ideas into plugins

Key points: @DeepSeekHarness ships Creator Mode — "Your Harness. You Define It." — turning ideas into reusable plugins, from a design workspace and sandbox game to a 3D LLM visualizer. The video post is extremely high engagement (~17k likes, ~9k bookmarks), extending yesterday's dsh-superagent prototype heat and expanding the harness story from "run tasks" to "user-defined capability surface."

Why it matters: Open harness teams are racing for the extensible plugin layer. If Creator Mode really lowers the bar, community remixes accelerate — and the comparison to Claude Code Skills / MCP gets sharper.

DeepSeek Harness Creator Mode

3. Claude Code self-orchestration vs Grok Bot: onenewbite's challenge

Key points: @onenewbite asks why so many people drive Claude Code via Grok Bot instead of letting Claude Code orchestrate Claude Code. He says he remote-controls a master session (chief of staff) from phone Claude Code, spinning child sessions in parallel, with multi-account login on mobile. Follow-ups note Claude Code Desktop's native multi-session plus a custom coordination skill, and argue harness plugins are basically customized AI–user interfaces.

Why it matters: High-signal Chinese-timeline friction on whether an outer orchestration layer is worth it — Grok Bot's cloud-computer coworker story versus Claude Code's native session tree. The right answer depends on whether you need cross-app agency or full-context local orchestration.

Claude Code self-orchestration

4. Antigravity: November Mission Report

Key points: @antigravity posts this month's Mission Report (video) covering product updates and content highlights. Community follow-ups mention changelog themes like Marketplace, built-in browser, and freer attachments, plus case strings (Blender scenes, multilingual calibration, Fortran→Rust). Separate comments "scraping" training rumors from the video are second-hand — keep them distinct from the official report.

Why it matters: Google's coding agent keeps mindshare with a monthly narrative; shipped case studies matter more than a one-line changelog for whether it finishes real engineering work.

Antigravity Mission Report

5. pi: a spare agent builds pico + an eval harness

Key points: @badlogicgames keeps showing pi durable / codemode: a spare agent writes a tiny language (pico, QuickJS/JS vs Armin's djs/TS) and an eval harness — tasks different models should complete. Self-described "exceptionally cursed," but the point is clear: harness authors using agents to produce evaluation assets.

Why it matters: Not a product launch — a veteran harness practitioner closing the "agent builds agent infra" loop. Read alongside DeepSeek / UNCUT the same day: evals and minimal interfaces are becoming shared ground.

pi pico eval

Key points: Multiple posts say T3 Code topped GitHub trending (one cites ~+3,100 stars in a day): one desktop/web/phone app driving local Claude Code / Codex / Cursor on your own subscriptions. The repo dates to ~2026-02 with tens of thousands of stars already — today's heat is a trending recirculation, not a fresh open-source drop.

Why it matters: Demand remains strong for a multi-surface shell that drives local agents on existing subs. Digests should separate new explosions from old-project comebacks and avoid echoing "just open-sourced" framing.

T3 Code

7. UNCUT: a "lossless" minimal harness clears 500★ in a day

Key points: GitHub Balanov/uncut (created_at ~2026-11-01, MIT) gets dense X write-ups: ~837 lines of Go, single ~7MB static binary; no TUI/plugins/background agents/auto-retry/context compression; only four tools (read/write/replace/run_command); No Silent Cuts — truncation fails loudly instead of silent summarization. @viest_w and @fishjar compare Claude Code / Codex compression, sandboxes, and extension surfaces, framing UNCUT as a control group, not a replacement.

Why it matters: In a week of thicker plugin stacks, UNCUT stress-tests transport honesty. Good for CI/containers and people who want to read a whole harness; also a reminder that no compression means long jobs hit context walls.

UNCUT

Key points: @0xGky and others keep recommending yetone/magpie (MIT; repo ~2026-09, not newly open-sourced today): menu-bar one-click model swaps for Claude Code / Codex / Gemini CLI / OpenCode / Cursor CLI; a local Gateway for protocol translation so Codex can hit DeepSeek and Claude Code can hit Kimi, plus multi-account quota rotation. Chinese posts stress fewer config edits.

Why it matters: Model routing / subscription reuse is still a pain point. Note the repo has existed for months — today's posts are recirculation. Check subscription terms and default credential-reading behavior before installing.

Magpie

9. Fable 5.5: unconfirmed Claude checkpoint rumor

Key points: @minchoi roundups a circulating unreleased Claude checkpoint nicknamed "Fable 5.5": full animations and 3D scenes from code alone (no separate image/video model), with same-prompt contrasts vs Opus 5.5, Gemini, and Astra (giraffe gait, open-world RPG, Spider-Verse-style short). The thread notes Anthropic has not confirmed.

Why it matters: If the visual coding claims hold, demo standards jump overnight — but this is still a second-hand demo chain. Watch-list material, not a ship note.

Fable 5.5

10. Agent disk spring-cleaning: 13.8GB of hidden junk

Key points: @illyism posts an "AI spring-cleaning" on a Mac that freed ~13.8GB — tens of thousands of old Grok CLI sessions, thousands of Cursor transcripts, Codex SQLite history, dead Keychain tokens, and editor caches. Includes a pasteable audit + 30-day auto-prune prompt for Cursor / Claude Code.

Why it matters: Local agent state is becoming disk and privacy debt. If products keep transcripts forever by default, the ops cost flips from delightful to janitorial — practical reminder for multi-agent users.

Agent junk cleanup