A Google-affiliated research team showed that an AI coding system can get better at using a model without retraining the model itself. That is a real result. It does not show that Gemini trained its own successor, or that Google used the technique to accelerate its November launches. World of AI's roundup puts those possibilities next to a verified design leaderboard, Anthropic's first trading day and a set of product updates you can actually try.
Watch World of AI's November 20 Roundup
Source: World of AI's November 20, 2026 episode and its transcript. Research, rankings, filings and rollout details were checked against the linked sources on November 20. The video includes a sponsored TestSprite segment. Follow World of AI on X.
What Google's RSI Paper Actually Shows
An agent harness is the software around a model: the instructions it receives, the tools it can call and the rules that decide when it should retry or stop. Two agents can use the same underlying model and still give very different results because their harnesses differ.
In the November 13 preprint Self-Improvement Memory Enables Continual Recursive Self-Improvement, researchers from Google Cloud AI, Google DeepMind and Penn State tested whether an optimizer could keep improving that harness when the task set or the model changed. The memory records why earlier changes were tried and how they performed, so a later run can reuse those lessons instead of inheriting only the edited harness.
Across ten matched coding comparisons, the version that carried forward the memory scored higher in nine, with an average gain of 2.36 benchmark points over continuing with the harness alone. The paper also includes Claude-to-Gemini transfers. It changes how a model is used, not its weights. A repeatable gain under those conditions is evidence for harness-level self-improvement, not proof of a system autonomously training a sequence of more capable frontier models.
Google's Gemini 4 Neon launch is official. The episode's suggested Helium positioning as a lightweight model and the reported Krypton checkpoint are a different kind of evidence. I found no Google release note confirming those names, specs or launch plans at the time of this roundup. A teaser or a fast release schedule tells you nothing conclusive about what happened inside training.
Gestalt Beat Fable on One Design Benchmark
On the Contra Labs and BenchFlow Frontier Design Benchmark, judges compare websites blind, using client-style design briefs. The leaderboard showed a 218-point Elo gap between the stealth entry Gestalt and Claude Fable 5.2 on November 20.
| Rank | Entry | Elo |
|---|---|---|
| 1 | Gestalt (stealth) | 1,931 |
| 2 | Human designers recruited from Contra | 1,739 |
| 3 | Claude Fable 5.2 | 1,713 |
The gap measures preference in those design comparisons. It isn't a 218-point improvement in coding ability, an accessibility score or evidence of higher conversions. Contra does not identify Gestalt's provider on the leaderboard, so attributing it to Google, Anthropic or a Chinese lab would be guesswork. For your own site, compare the designs blind, then test whether people can complete the intended task on a phone and with a keyboard.
Anthropic's IPO: Put Each Number in the Right Column
Anthropic began trading on Nasdaq as ANT on November 19. The final prospectus covered 980 million shares at $83, or $81.34 billion before expenses and before the underwriters bought additional shares. The November 20 closing notice puts the final public offering at $93.5 billion gross including that option. About $91.6 billion of the public offering went to Anthropic before expenses; the rest came from shares sold by existing holders. The concurrent $26 billion private placement brings the combined gross total to about $119.5 billion, which is neither the IPO alone nor Anthropic's net proceeds.
The shares first traded at $111.50 and closed their first session at $97.17, roughly 17% above the offer price but below the first trade. Those are three prices from different points in the transaction, not a current quote. First-day market recap.
The episode also discusses Dario Amodei's stated commitment to pace training and deployment around safety. The test of that commitment is whether it changes a training or release decision when the pressure to ship is real.
Genie 4, GPT-Live-1.1 and What Else Shipped
| Update | What changed |
|---|---|
| Genie 4 | Change the world by prompting while you explore it, with audio that adapts to what's happening. Google's release also adds a Gemini creation assistant and a gallery for remixing worlds. Genie 4 started rolling out in Project Genie to eligible Ultra subscribers. The updated sessions of up to three minutes described in the announcement apply to Genie 3. |
| Antigravity SSH | Open a project on a remote Linux host from the Antigravity desktop app. The agent and terminal run on that host, which makes its permissions just as important as your local workspace settings. |
| ChatGPT mobile and Health | The redesign adds a dedicated dot tab and lets eligible users pin features such as Projects. It's a limited rollout; OpenAI's launch notes exclude the EEA. Separately, the U.S. Health rollout expands to Android with Health Connect, desktop, sites and dots, subject to plan and feature eligibility. |
| GPT-Live-1.1 API | The voice model is available to developers. WebRTC standby can pause inference during quiet stretches while keeping the call connected. OpenAI lists standby for both GPT-Live-1 and GPT-Live-1.1. |
| Microsoft Canvas | MAI Image 2.7 Flash powers an editing workspace for turning sketches into images or diagrams and modifying selected regions without restarting the whole image. |
| OpenMotion | A third-party open-source project that turns references and brand assets into editable motion graphics. It works with DeepSeek Harness and other agents; this isn't a native DeepSeek video model launch. |
Still in beta: In a November 19 post on X (archived copy), Elon Musk said several Grok 4.8 variants were being tested and that a wide-release variant had not been chosen. That doesn't give users a model ID or a rollout date.
Confirmed leadership change, unconfirmed next move: NVIDIA announced that Jensen Huang resigned and CFO Colette Kress took over as interim CEO. The episode's speculation about Huang starting another semiconductor company is not a confirmed plan.
DeepMind Institute's Four Paths Beyond AGI
In What Comes After AGI?, published November 19 and based on an earlier technical report, Google DeepMind researcher Tim Genewein examines how broadly human-level AI might progress to systems that outperform large groups of experts. The essay explores four routes:
- Scale existing approaches: Apply more compute and other resources to methods already working.
- Find a new paradigm: Discover approaches that make possible work today's architectures struggle with.
- Coordinate many agents: Run enough capable agents in parallel that the collective can tackle work an individual agent cannot.
- Use AI to improve AI: Let stronger systems participate in the research and engineering that produce the next generation.
These routes can overlap. They're scenarios, not evidence that any lab has reached AGI or achieved sustained autonomous model improvement. For a claim about today's agents, ask what actually changed: the model, the harness, or the number of agents doing the work. Then look for a result that holds up on work the system didn't optimize against.
Video Chapters
| Time | Topic in the original video |
|---|---|
| 00:00 | Intro |
| 01:56 | Teasers for the next Gemini model |
| 02:55 | Gemini 4 Helium speculation |
| 05:00 | Did Google achieve RSI? |
| 06:35 | Research results and unverified model reports |
| 08:14 | Gestalt tops the design benchmark |
| 08:50 | Speculation about who built Gestalt |
| 10:00 | Anthropic goes public |
| 11:38 | IPO launch figures |
| 12:22 | Anthropic's first trading day |
| 14:37 | Dario Amodei on pacing and safety |
| 16:13 | Grok 4.8 beta variants |
| 17:34 | ChatGPT mobile redesign |
| 18:18 | ChatGPT Health expands |
| 19:24 | Genie 4 in Project Genie |
| 20:36 | Native SSH in Antigravity |
| 21:03 | GPT-Live-1.1 in the API |
| 22:28 | DeepMind Institute's four paths beyond AGI |
| 24:21 | Microsoft Canvas and MAI Image 2.7 Flash |
| 25:03 | DeepSeek and OpenMotion |
| 26:01 | Unverified report about Jensen Huang |
| 26:24 | Outro |