Oil painting of a small robot building a spiral staircase step by step from an open notebook, representing self-improving AI

Quick answer

No, not in the science-fiction sense. On 13 November 2026 Google researchers published “Self-Improvement Memory Enables Continual Recursive Self-Improvement”, which shows an AI system getting better at improving the tools and instructions around a model by remembering what it tried before and whether it worked. The model’s own weights don’t change. Talk that Google has “achieved RSI” comes from rumours and remarks, not from anything Google has confirmed. The same week brought a mystery design model called Gestalt, Anthropic’s stock market debut and Genie 4.

Key takeaways

  • Google’s SIM paper gives an optimiser a memory of past edits, reasons and results. In coding tests it beat the version without memory in 9 of 10 runs, by 2.36 points on average.
  • Improvements carried across models. Changes learned with one model (Claude Opus) still helped when the system switched to Gemini.
  • “Google has achieved RSI” is speculation. It is based on leaks, a fast release pace and a comment from Sergey Brin, not an announcement.
  • Gestalt, an unnamed stealth model, tops Contra Labs’ design benchmark with 1,931 Elo, ahead of Claude Fable 5.2 (1,713). Its maker is unknown.
  • Anthropic began trading on Nasdaq as ANT on 19 November, closing its first day at $97.17, up about 17% on its $83 IPO price.

🛠️ Practical guides

Ready to put this into practice? These step-by-step guides answer two common follow-up questions:

“Recursive self-improvement”, or RSI, is the idea of an AI that makes itself smarter, which then makes it better at making itself smarter. It is one of the most talked-about ideas in AI, and this week it was everywhere. Here is what Google’s new research actually shows, what is still rumour, and the rest of the week’s news.

This round-up is based on the 20 November news video from WorldofAI, with the main claims checked against the original sources where we could find them.

Watch: the source video

What this video covers: WorldofAI’s AI news round-up for 20 November: Google’s periodic-table teaser and Gemini 4 Helium hints, speculation that Google has reached recursive self-improvement, the Self-Improvement Memory paper, the Gestalt stealth model on Contra Labs’ design benchmark, Anthropic’s Nasdaq debut, Grok 4.8 testing, ChatGPT’s mobile redesign and Health expansion, Genie 4 in Project Genie, Antigravity SSH support, GPT Live 1.1, DeepMind’s essay on paths to superintelligence, Microsoft’s Canvas, DeepSeek’s OpenMotion plugin and rumours about Jensen Huang’s next move.

Video by WorldofAI (@intheworldofai). The video was sponsored by TestSprite, an AI testing tool. We have no link with the sponsor.

What is Google’s Self-Improvement Memory paper?

It is a method that gives an AI “optimiser” a long-term memory of every change it has made to an AI system, why it made it, what it expected and what actually happened. The paper comes from researchers at Google Cloud AI and Google DeepMind.

Here is the setup in plain English. Many AI products wrap a model in a “harness”: the prompts, tools, rules and checking steps around it. A second AI can keep tweaking that harness to improve scores. The problem is that, without memory, it forgets which tweaks worked and repeats mistakes. SIM keeps a record so each round builds on the last.

Result reported in the paperFigure
Final score with SIM vs without (headline test)84.75 vs 73.42
Coding runs where SIM did better9 out of 10
Average coding gain over harness-only continuation+2.36 points
Did improvements transfer to a different model?Yes, from Claude Opus to Gemini

The transfer result is the interesting part for everyday users. The lessons were stored as notes about the harness, not baked into one model, so they still helped after a model switch. You can copy that idea with something as simple as a change log, which we cover in How to Keep a Prompt Change Log So Your Improvements Survive a Model Switch.

Has Google actually achieved recursive self-improvement?

There is no public evidence that it has. WorldofAI pointed to several hints: a leak about a project reportedly called Lyra, reports that Gemini 4 Neon was built on Gemini 4 Hydrogen with newer training, Google releasing four models in about two weeks, a periodic-table teaser that seems to point to a “Gemini 4 Helium” Flash model, and Sergey Brin reportedly saying Google is “on the RSI treadmill”.

Those are signs of fast progress, but they are not the same thing. The SIM paper improves the harness, not the model’s own weights, and the authors describe it in those terms. Full RSI would mean a model meaningfully improving its own training and design with little human help. Google hasn’t claimed that.

🔎 Unconfirmed / rumour: Rumour: Gemini 4 Helium, the Lyra project and an RSI breakthrough at Google are unconfirmed. Treat them as speculation until Google announces something.

Anthropic has been more open about this topic; see Anthropic Recursive Self-Improvement: What the IPO Filing Change Means, Plus Fable 5.5 and This Week’s AI News. Google’s newest released model, Gemini 4 Neon, is covered in our Gemini 4 Neon Review: What the Tests Show on Speed, Cost and Quality.

What is the Gestalt stealth model?

Gestalt is an unreleased “stealth” model that currently ranks first on Contra Labs’ Frontier Design Benchmark, which judges AI on real design briefs. It scored 1,931 Elo, above Contra’s human “golden” designs (1,739) and Claude Fable 5.2 (1,713). Contra reports that it won 78% of blind head-to-head votes against Fable 5.2.

Contra says it is working with Gestalt’s team but hasn’t named the company. People online have guessed at links to a university research project and to Chinese labs, but none of that is confirmed. If a stealth model appears in a tool you use, read How to Use a Free Stealth AI Model Without Risking Your Data before sharing anything private with it.

How did Anthropic’s first day on the stock market go?

Anthropic listed on Nasdaq under the ticker ANT on 19 November 2026. Shares were priced at $83, opened well above that and closed the first day at $97.17, about 17% higher. They fell the next day during a wider sell-off in AI stocks. WorldofAI also noted Dario Amodei’s comment that Anthropic would pace how it trains and releases models.

For Claude users, nothing changes overnight: the apps, plans and prices are the same. Our full coverage of the listing and what to watch next is in the Anthropic IPO round-up.

⚠️ Warning: This is news, not investment advice. The creator shared personal views on the shares in the video; we don’t recommend buying or selling any stock, and newly listed shares can swing sharply.

What can Genie 4 do?

Genie 4 is Google DeepMind’s new world model: it creates interactive 3D worlds you can move around in, now with matching sound and the ability to change the world while you’re in it. Google says it is available in Project Genie for eligible Google AI Ultra subscribers.

  • Native audio that matches what happens on screen.
  • Real-time editing with text prompts, such as adding an avalanche while you ski.
  • A Genie agent, powered by Gemini, that helps you build worlds through conversation and can suggest changes.
  • A community gallery for exploring and remixing other people’s worlds.
  • Genie 3 improvements, including sessions of up to three minutes.

It is still a premium research product, but it shows where interactive AI video and games are heading.

What changed at OpenAI this week?

  • GPT Live 1.1 in the API. Developers can now use OpenAI’s real-time voice model, which adds a standby option to cut costs while a voice assistant sits idle.
  • ChatGPT mobile redesign. WorldofAI showed a refreshed app layout with a new “Dots” tab; it is rolling out gradually.
  • ChatGPT Health expansion. More health features reached US users. As always, treat AI health answers as a starting point for a conversation with a professional, not a diagnosis.

What else happened in AI this week?

  • Grok 4.8 testing. xAI appears to be testing several Grok 4.8 versions inside the Grok Bot on X; there is no release date.
  • Antigravity SSH support. Google’s Antigravity desktop app can now connect to remote servers over SSH, so developers can work on remote projects without leaving the app.
  • DeepMind on “what comes after AGI”. An essay from DeepMind researchers sets out four possible routes to superintelligence: scaling, new AI designs, millions of cooperating agents and recursive self-improvement. It argues effective AI compute has grown around 10 times a year. These are projections, not predictions.
  • Microsoft Canvas. A new workspace powered by MAI Image 2.7 Flash that lets you select part of an image and change only that part. Our guide How to Edit One Part of an AI Image Without Starting Again shows how to do this in any tool.
  • DeepSeek OpenMotion. An open-source plugin for making motion graphics with DeepSeek’s harness, with over 300 templates. WorldofAI says results compare well with pricier models; we haven’t tested it.
  • Jensen Huang rumour. Unconfirmed reports suggest Jensen Huang could raise around $50 billion for a new chip company. See Jensen Huang Resignation: What It Could Mean for Your AI Bill for background.

What is confirmed and what is only claimed?

StoryStatus
Google SIM paper and its resultsConfirmed (published paper)
Google has achieved recursive self-improvementSpeculation
Gemini 4 HeliumUnconfirmed (teaser and leaks only)
Gestalt tops Contra’s design benchmarkConfirmed by Contra; maker unknown
Anthropic listed as ANT at $83, closed day one at $97.17Confirmed
Genie 4 in Project Genie for AI UltraConfirmed by Google
GPT Live 1.1 in the APIConfirmed
Grok 4.8 variantsSeen in testing; not released
Jensen Huang’s $50bn chip ventureRumour

What should you do about this news?

  1. Start a change log for your AI prompts and skills. It is the low-tech version of the SIM idea and makes switching models far less painful.
  2. Don’t rush into stealth or leaked models with real customer data, however good their benchmark scores look.
  3. Ignore stock hype. Anthropic’s listing doesn’t change your Claude plan, and none of this is a reason to buy or sell shares.
  4. Try targeted image edits instead of regenerating images from scratch; it saves time and credits.

Frequently asked questions

What is recursive self-improvement (RSI)?

The idea of an AI that improves itself, then uses that improvement to improve itself further, in a repeating loop.

Did Google announce that it has achieved RSI?

No. The claim comes from speculation and leaks. Google’s SIM paper improves the setup around a model, not the model itself.

What is the SIM paper?

A Google research paper showing that giving an AI optimiser a memory of past changes and results helps it keep improving an AI system over time, even after switching models.

Who made the Gestalt model?

Nobody knows publicly yet. Contra Labs says it is working with the team but hasn’t named them.

Can I use Genie 4?

Google says it is available in Project Genie for eligible Google AI Ultra subscribers.

Was the video sponsored?

Yes, by TestSprite, an AI testing tool.

The bottom line

Google hasn’t built an AI that rewrites itself, but its SIM paper shows something practical: AI systems improve more steadily when they remember what was tried and what worked. You can use the same principle today with a simple change log. The rest of the week, from Gestalt to Genie 4 to Anthropic’s market debut, shows how fast things are moving. Stick to what is confirmed, and test new tools on your own work.

author avatar
Garry Knight
I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

Leave a Reply