Gemini 4 RSI Leak Destroys Astra and Fable in Benchmarks

Value of the Information & Strength of the Argument

The video compiles scattered rumors and leaks into a coherent narrative, offering a valuable roundup for those tracking AI developments. However, its argumentation is largely based on inference: the anomalous behaviors of an anonymous Arena model are presented as proof of Gemini 4’s existence, but without direct confirmation. The presenter does acknowledge this, framing everything as speculation. The strength lies in the aggregation of information, though the reasoning occasionally leaps from correlation to causation. No rigorous methodology is applied; instead, it relies on community observations and unverified charts. Consequently, the video is informative but not analytically rigorous.

Scientific Rigor, Source Quality, Title Accuracy

The video cites several official sources in its description, including an arXiv paper, Google’s blog on Q2 earnings, OpenAI’s GPT-6 Astra page, and Anthropic’s Claude Fable page. These lend some credibility to the contextual claims. However, the primary leak evidence comes from unofficial channels like NokiaPowerUser and Qiniu News, which are not independently verified. The title is somewhat clickbait (‘Destroys’), though the content does cover the benchmark claims. Overall, the rigor is moderate: the presenter distinguishes between confirmed facts and rumors, but the lack of verifiable primary sources for the core claims keeps the scientific quality low.

The title accurately reflects the video's focus on leaked benchmarks that suggest Gemini 4 outperforms competitors, though the word 'Destroys' is sensationalist. The content matches the premise, albeit with appropriate caveats.

4/10

The video relies heavily on unverified leaks and speculation. While it cites official sources for context (e.g., Google's earnings call, an arXiv paper on Dream-RSI), the core claims about Gemini 4's identity and performance remain unconfirmed. The presenter clearly labels rumors as such, but the overall reliability is limited by the absence of official confirmation.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction: The leak of an anonymous model under the name 'Gemini 3.8 Flash' on Arena is revealed.
• 01:00Example of the model producing a highly detailed PS5 SVG render in 10 minutes, unlike the real flash's fast but rough output.
• 04:00Showcase of other visual outputs (voxel pagodas, BMW M4, 3D helicopter) and the model's ability to handle complex multi-part prompts.
• 06:00Leaked benchmark numbers: SWE-V1.1 (88%), GDP Eval-A-V2 (2064 Elo), Terminal Bench 2.1 (95.3%), OSWorld 2.0 (86.8%) – all ahead of competitors.
• 08:00Back-end specs: 10M input context, 256K output, native memory, internet access, sandboxing, and robot motor control support.
• 11:00Discussion of Google's struggles, RSI research (Dream-RSI paper), and potential reveal timeline in October.

• arXiv: Dream-RSI paper — Referenced as research on recursive self-improvement, used to support the RSI angle.
• Google CEO blog on Q2 2026 earnings — Cited for Pichai's statement that Gemini 4 is needed to stay competitive.
• NokiaPowerUser leak article — Primary source for leaked specs and benchmark claims.
• Qiniu News benchmark article — Secondary source for leaked benchmark comparisons.
• OpenAI GPT-6 Astra page — Official page for one of the competitors mentioned.
• Anthropic Claude Fable page — Official page for another competitor mentioned.
• VentureBeat article on Dream-RSI — Covers the Dream-RSI research and its impact on agent efficiency.