Interview
David Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience
- The outlook posits that superintelligence will emerge from autonomous systems capable of lifelong learning from experience, explicitly rejecting the path of LLM-based intermediate products or human priors as sufficient for a "true super learner."
- Ineffable is founded as a dedicated, research-led elite lab with a small team of 20 people to pursue this approach without the constraints of larger organizations, utilizing a "research-led culture" to push back on product pressure.
- The company has secured access to the largest Ruby cluster on GCP by year-end, consisting of GPUs rather than CPUs, enabling experiments with 100,000 times more compute than early reinforcement learning examples like OpenAI's Hide and Seek.
- The technical strategy involves training agents in simulated "crucibles" with high emergent complexity and non-stationary data distributions to discover novel fields of science, technology, and mathematics without human data weighting, relying on the human brain as an existence proof for feasibility.
- Timeline expectations include achieving extremely strong evidence of progress within one year and reaching true superintelligence within three years, driven by favorable timing in technology, scaling laws, and compute availability.
- Future risks and governance include the limitations of current language-based alignment, the potential for monitoring difficulties as agents gain power, and a commitment to gating all real-world deployment on sufficient evidence from the science of superintelligence developed in simulation.
- The anticipated outcome of superintelligence is described as an encounter with "magic" that brings wonderment, discovers unknown fields, and creates a future balance of actors where systems support and protect humanity.
- The founders, including those from AlphaStar and AlphaGo projects, believe the "short-term sacrifice" of removing human data is necessary to avoid historical biases and access capabilities far beyond what is possible with human data, predicting the process will "just have to work."