Ilya's SSI to Release First Model Using Test-Time Training Architecture
新智元 · wechat · 2026-08-25
Safe Superintelligence (SSI), founded by Ilya Sutskever, is rumored to release its first major model this week. a16z partner Martin Casado hinted at encountering the "most important new model of the year."
The model is reportedly based on the Test-Time Training (TTT) architecture. Unlike traditional models that freeze weights after pre-training, TTT allows the model to update its weights in real-time during inference, internalizing new information rather than relying solely on context windows. This implies the model possesses continuous learning and evolutionary capabilities.
NVIDIA has partnered with SSI, scaling its compute power by 10x over the next 12 months. Jensen Huang stated the decision came after gaining "rare access" to SSI's research results.
Related event: SSI Rumored to Release First Model This Week(4 posts)→
More from AGI Musings
- 11 configurations of firm/worker/agent require different designs, shifting focus from automation to collaboration — random_walker · 2026-08-25
- Best-case tech adjustment: incumbents keep jobs, youth enter new roles — soumitrashukla9 · 2026-08-25
- Nicolas Cole: 100x AI output comes from explicit process documentation — cen6wkf · 2026-08-25
- Hinton: No Sharp Line Between Tool and Subject, Agency Emerges — MacrinePhD · 2026-08-25
- The AI Disillusionment Law: Frontier Models Feel Godlike for Days, Then Dumb — D3VAUX · 2026-08-25
- AI-flooded complaints strain UK public bodies, reports BBC — emax · 2026-08-25