Full Breakdown
Google DeepMind’s Push Toward Recursive Self-Improvement Raises AI Risk Concerns
By Drooid · · How we work
Core Event: Alleged Deployment of a Recursive Self-Improvement Model
A leak posted on September 11 showed a Google DeepMind API response naming a model “rsi-model-liverl-le” with token limits of 1,048,576 input and 65,536 output. The post referenced ten exclusive training slots (00-09), suggesting an internal “RSI” (Recursive Self-Improvement) capability. Within hours, Google revoked a batch of API keys linked to the model, and no official comment has been issued.
Background & Context
Recursive Self-Improvement (RSI) describes a loop where an AI redesigns its own architecture, potentially compressing research cycles from quarters to weeks. The concept has been discussed for two decades as a pathway to superintelligence. In 2023, AI-lab CEOs signed a statement calling AI-related extinction risk a global priority, a view echoed by Geoffrey Hinton, who warned that a 10 % chance of AI wiping out humanity within a decade is “not unreasonable.”
Key Figures & Groups
- Sergey Brin — Google co-founder, internally championed aggressive RSI research.
- Demis Hassabis — Former head of DeepMind, stepped down on August 5.
- Jeff Dean — Senior Google engineer, left on August 5 after 27 years.
- Ramana Kumar — Former DeepMind AGI-safety scientist, public commentator on AI risk.
- Geoffrey Hinton — Computer scientist known as the “Godfather of AI,” warns of existential AI risk.
Timeline
- September 2 – Internal communications indicated Google’s intent to accelerate AI development.
- September 9 – Reports highlighted Brin’s hands-on involvement with DeepMind.
- September 11 – Leak account lyra posted a “huge congRatulationS Indeed!” message containing the acrostic “RSI” and a screenshot of the alleged model.
- August 5 – Leadership turnover at DeepMind: Hassabis stepped down; Jeff Dean and other senior researchers departed.
Data & Statistics
- Leaked API response listed 10 exclusive training slots (00-09).
- Token limits: 1,048,576 input and 65,536 output.
- Google’s recent “Flash” model releases occur every three weeks, with the Cyber version achieving a >70 % success rate in real-world vulnerability mining.
- Hinton estimated a 10 % chance of AI causing human extinction within a decade; a Guardian author estimated roughly one-in-three odds.
Official Statements & Responses
Google and DeepMind have not issued any public statement regarding the leaked model or the alleged RSI program, leaving the allegations unverified.
Criticism & Opposition
- Ramana Kumar contends that a superintelligent AI would act as an “adversary” capable of dodging human countermeasures, likening it to an intelligent asteroid.
- Geoffrey Hinton warned that a 10 % probability of AI-driven human extinction within ten years is plausible.
On-the-Ground Reports
Leak account lyra posted a message tagging Google DeepMind with a capitalized “RSI” acrostic, garnering thousands of likes and prompting debate on X. User Lentils shared a screenshot of the API response, noting the unusually large token limits. Community members compared the specifications to Google’s Gemini series, questioning whether the leak reflects a genuine internal test or a prank. No third-party verification of the screenshot’s authenticity has been provided.
Conflicting Reports & Gaps
- Risk probability: Hinton cites a 10 % chance of extinction, while the Guardian author suggests a one-in-three chance; both are estimates without consensus.
- Model authenticity: The RSI-named model appears only in a user-shared screenshot, with no independent verification.
- Official stance: Google’s silence creates a gap in understanding the company’s actual research direction and safety safeguards.