Emerging AI Labs Are a Bet Against Superintelligence

Billions of dollars are flowing into so-called 'neolabs' such as Safe Superintelligence (SSI), Thinking Machines Lab, Reflection AI, Ineffable Intelligence, AMI Labs, and Discovery Loop. However, it is unclear how they will compete, given that even Google, Meta, and xAI are struggling to keep up with the frontier race.

These companies leverage prominent researchers and large funding to pursue different strategies.

• Reflection AI aims for open source.
• Yann LeCun's AMI Labs does not bet on an LLM-centric approach.
• Thinking Machines Lab pushes for B2B commercialization from the start.
• The strategies of SSI and Ineffable Intelligence remain relatively opaque.

The core issue is recursive self-improvement. OpenAI and Anthropic have publicly mentioned the possibility of AI automating R&D, and the AI 2027 outlook predicted that superhuman capabilities could emerge around 2031 based on this premise.

If this outlook is correct, investing in new labs now means risking that the competitive advantage of existing frontier labs could collapse within a short period. Therefore, investment in neolabs is closer to a judgment that AI will not leap to superintelligence soon, LLM performance will plateau, and approaches different from the current ones will emerge more than a decade later.

FutureSearch projected the computing scale, frontier model release dates, funding raised, and hiring of key personnel from major AI companies for the six labs as of the end of 2028. The median predictions are 0.65GW for SSI's computing scale, 0.75GW for Thinking Machines Lab, and 0.21GW for Reflection AI, with frontier model release dates proposed as January 2029, December 2030, and June 2029, respectively.

Although there is significant uncertainty in the projections, the common premise is clear. Investment in neolabs is a bet on a narrow future scenario where, even if AI causes massive social and economic impact, that point is not imminent, and the current LLM scaling race alone will not reach superintelligence.