Tech • AI • Robotics • Game

ENFR

Google Achieved RSI?! New Stealth Model Beats Fable 5.2, Anthropic Post-IPO, Genie 4 & More! AI News

AIWorldofAINovember 20, 2026 at 08:15 AM27:05
Audio player

⏱ ≈ 3 min read · 27-min video — 24 min saved

TL;DR

Google is signaling another Gemini release as speculation grows that rapid model iteration, new self-improvement research, and intensified competition are accelerating the race from frontier AI toward possible superintelligence.

KEY POINTS

Google hints at another Gemini model

Posts from Gemini team members and a new periodic-table style teaser have fueled expectations that Google is preparing another launch within days. The next model appears to be Helium, following earlier element-themed names such as Carbon, Argon, Hydrogen, and the newly released Gemini 4 Neon.

Gemini 4 Neon raises questions about Google’s pace

Gemini 4 Neon, released a day earlier, is described as a smaller model that nevertheless outperforms larger siblings such as Carbon and Argon in some tasks. Reports tie Neon to a newer training and post-training pipeline, suggesting Google may be extracting substantial gains from improved data, tooling and optimization rather than just model size.

Self-improvement research adds to RSI speculation

A newly published Google research paper explores continuous AI self-improvement at the harness level rather than direct model-weight rewriting. The system stores a structured memory of what changes were made, why they were made, and whether they improved results; across 10 coding runs, the memory-based method performed better in 9, with an average gain of 2.36 points.

Recursive self-improvement remains unconfirmed

The paper does not prove autonomous recursive self-improvement of core model weights, but it shows how agents can refine surrounding tools, prompts and workflows over time and transfer those gains across models. That nuance matters: the evidence points to self-improving systems architecture, not definitive proof that Gemini is independently redesigning itself.

A stealth model called Gestalt jumps up the rankings

A previously unknown model named Gestalt has appeared on a frontier-design benchmark and is reportedly ahead of Fable 5.2 by nearly 200 Elo points. Reports also cite a 90% win rate in blind arena-style tests, though its developer remains unclear, with speculation ranging from Anthropic and Google to Chinese labs.

Anthropic launches a huge Nasdaq debut

Anthropic has begun trading on NASDAQ under the ticker ANT after pricing its IPO at $83 a share, later opening at $111.50 and closing at $97.17. Reported figures tied to the offering and private placement imply access to well over $100 billion in capital, a scale that could reshape spending on compute, data centers and next-generation model development.

Anthropic stresses safety while rivals accelerate

Chief executive Dario Amodei said AI could compress 100 years of scientific and medical progress into 10 and argued model deployment should proceed cautiously even under business or political pressure. The message sets up a contrast with Google’s visible release tempo, though investors will watch whether safety-first rhetoric holds as competition intensifies.

xAI, OpenAI and Google expand products

Elon Musk said Grok 4.8 is still in reinforcement-learning testing, with multiple variants quietly evaluated through live usage. OpenAI is rolling out a redesigned mobile ChatGPT experience, health integrations and GPT Live 1.1 in the API, while Google has introduced Genie 4, which can generate editable interactive worlds with native audio and conversational controls.

DeepMind researchers map paths beyond AGI

Researchers at Google DeepMind are openly discussing four routes from AGI to ASI: scaling models and data, discovering new paradigms, coordinating millions of agents, and recursive self-improvement. One striking estimate in that discussion is that effective AI compute has been rising by roughly 10x per year, implying a dramatic collapse in the cost of running powerful systems over the next decade.

Broader ecosystem momentum continues

Microsoft has introduced Canvas for iterative AI image editing, while DeepSeek has upgraded video generation through an open-motion plugin with hundreds of templates and more than 152 design styles. Separate market chatter also points to a possible $50 billion semiconductor venture linked to Jensen Huang, though that remains unconfirmed.

CONCLUSION

The immediate story is Google’s next likely Gemini launch, but the larger trend is a rapidly accelerating AI industry in which model releases, infrastructure spending and self-improvement techniques are converging. The central question is no longer only when AGI arrives, but how quickly major labs can turn iterative gains into a durable strategic lead.

Ask a question

More from AI