Ineffable's two founders: no human data and no pre-training, first build an agent that learns for itself in simulated worlds

A 'superlearner': start from nothing and keep learning past human knowledge

Silver splits it into depth and breadth. Depth is starting from knowing nothing and discovering things far beyond what humans know, as humanity collectively went from 'grey mush in our heads' to language, science and mathematics, with no blueprint and no trillion-item dataset. Breadth is being droppable into any stream of experience, simulated, digital or physical. He says no one has cracked this in AI and that it is the essence of what superintelligence needs.

“a system that can just keep learning and discovering knowledge forever.”David Silver

Giving up human data: a short-term sacrifice

He says the bet is to give up the human data and priors everyone relies on: a short-term sacrifice of some performance for a potentially far higher ceiling. He cites his own history: AlphaGo used human data with the team in the loop and took about two years to beat Lee Sedol; AlphaZero gave up the human data and in a few hours surpassed that work, beating the world champion in chess in four hours and, he recalls, taking 'a couple of days' in Go. He says he is not religious about it: weak priors that apply across environments are fine; what he avoids is strong priors encoding everything known about one problem.

Crucibles and simple objectives: not writing down what counts as science

Day to day there are two jobs: build a reinforcement-learning stack that runs at massive scale, and build simulated environments they call 'crucibles'. Crucibles are not hand-built puzzles; complexity is meant to emerge from physics engines with millions of interacting objects and millions of interacting agents, and Lasse adds that they want a range, with simple fast ones to iterate on. On objectives, they do not leave a breadcrumb trail toward a chosen outcome; they give agents simple objectives and hope to be surprised, with milestones being the agent rediscovering the steps humans took. He says they have tried many reward functions and has no single answer.

Build it in simulation first, then take the algorithm to the real world

He argues that if superintelligence can be built in simulation it will happen ten to a hundred times faster, because simulation parallelises without limit, is reproducible and is the safest place to do it. He agrees with Rich Sutton that what is learned in simulation should not be expected to transfer to the real world, which is far more complex; so the thing to discover in simulation is the superlearner algorithm itself, which would then be applied in the real world, with simulations made ever more faithful to check it. He grants that some things can only be expressed in reality, quantum computation for one. Lasse notes that the weather model MetNet beat physics-based simulation, so simulation has a ceiling.

Compute and team: a cluster, GPUs, twenty people

Lasse says they are first customers of the largest cluster on Google Cloud, starting 'by the end of the year' and then 'actually one month from now'; it is Nvidia GPUs, and he notes that on Vera Rubin one rack can act as one machine, which matters for reinforcement learning's fast iteration; simulation, inference and training all run on GPUs rather than large CPU fleets. Silver says many of today's large-scale RL ideas trace back to Lasse's early IMPALA and SEED RL; Lasse says Ineffable uses derivatives of a few such algorithms. The team is 20 people, kept small on purpose, so each has a lot of compute. Silver says that compared with OpenAI's hide-and-seek experiment they 'could probably run experiments that use 100,000 times more compute'.

Safety: study the science of superintelligence in simulation, release into the real world on evidence

Silver says alignment that rests on language and human-data priors will reach its limits as agents train further, because they overwhelm the priors and learn to say one thing and do another, and monitoring chains of thought gets harder. His alternative is to study 'the science of superintelligence' in simulation, learning under what conditions a society of superintelligences interacts peacefully, and to share it publicly; monitoring can avoid language, for example 'shadow actions' that show what an agent would do without letting it, or another superintelligence regularising the size of its effects. The company's commitment is to gate anything it does in the real world on enough evidence from that science. He supports pacing the real-world frontier while moving as fast as possible in simulation.