iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
David Silver’s Ineffable Intelligence is pursuing reinforcement learning (RL) as another route to more capable AI: agents that learn by acting in environments and observing what happens, rather than depending on human-generated data as their foundation. That is a research thesis and long-term ambition—not evidence that the startup has built a general-purpose “superlearner” or can replace today’s large language models (LLMs).
What is David Silver trying to build?
Silver, a former Google DeepMind researcher, founded Ineffable Intelligence to pursue an ambitious RL-centered research program. In a company note dated January 15, 2026, he wrote: “The world needs a place where the full ambition of the reinforcement learning paradigm can flourish.” He described wanting to focus on discovering knowledge through experience in an environment, while acknowledging the risk that the effort might fail.
Ineffable describes its intended system as a “superlearner”: an agent that would gain skills and knowledge through its own experience, from motor abilities to intellectual discoveries. The company says it expects such a system to rediscover and go beyond human inventions including language, science, mathematics, and technology. Those are the company’s aims, not results demonstrated in the public evidence described here.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteIn a WIRED interview published April 27, 2026, Silver called the mission “making first contact with superintelligence.” That phrase captures the scale of the ambition; it should not be read as a claim that the company has achieved superintelligence.
#1 Best Overall
Why does Silver think learning from experience could matter?
Most current LLMs learn patterns from large collections of human-produced material. Silver argues that this gives them a powerful shortcut, but also makes their knowledge dependent on what people have recorded. He described human data as “a kind of fossil fuel that has provided an amazing shortcut,” and contrasted it with systems that learn for themselves as “a renewable fuel.” These are metaphors for a proposed learning paradigm, not measured comparisons of energy use, capability, or training cost.
Silver’s flat-earth example illustrates the point. A language model trained on the beliefs of a hypothetical flat-earth society could inherit that worldview if it had no way to interact with reality. An agent able to act in an environment and observe consequences might, in principle, reach conclusions independently of that inherited account. This is a thought experiment supporting Silver’s argument, not an empirical demonstration that RL agents reliably correct false beliefs or discover truths.
The distinction is about the learning signal, not a proven choice between two systems with equivalent abilities. The sources do not establish that an RL agent can match LLMs at writing, conversation, coding, or other present-day uses. Nor do they show that human-generated information would have no role in a future system built around experience.
Recommended Free Tools
How does the proposed approach compare with LLM-centered training?
The comparison below describes the contrast Silver draws and the limits of the available evidence. It is not a controlled head-to-head benchmark.
| Question | LLM-centered training | RL from experience |
|---|---|---|
| Primary learning signal | Patterns in human-generated data, as characterized in Silver’s WIRED interview, April 27, 2026. | Actions and consequences in an environment, the approach Silver and Ineffable advocate in the WIRED interview and company note. |
| Interaction with the world | Whether a system can act and observe consequences depends on the system; the interview’s contrast concerns learning from human data versus learning through experience. | Interaction is central to the proposed thesis: agents would pursue goals and learn from what happens. |
| Demonstrated scope in the cited evidence | No general capability comparison is provided by the cited sources. | AlphaZero demonstrated strong performance in rule-defined games; that result does not establish open-ended learning in the real world. |
| Evidence status | The interview offers Silver’s characterization, not a controlled comparison of general capabilities. | AlphaZero is a published research result in bounded games; Ineffable’s broader superlearner remains a company ambition. |
What has reinforcement learning already demonstrated?
The clearest precedent in the cited material is AlphaZero, developed by Silver and coauthors. Their 2017 paper reports that it began from random play, received no domain knowledge beyond the rules of the games, and achieved superhuman performance in chess, shogi, and Go within 24 hours.
That is meaningful evidence that self-play RL can produce exceptional performance in environments with explicit rules and outcomes. It is not evidence that the same method has discovered new science, learned broadly useful real-world skills, or produced general intelligence. A game supplies a sharply defined world: its legal actions, objectives, and success conditions can be specified. The company’s proposed ambition reaches far beyond that setting.
Rank #3
What is the hardest step between games and the real world?
Silver’s proposed next step, as described in the WIRED interview, includes putting agents in simulations where they can pursue goals and collaborate. A simulation can give an agent room to act and learn, but the existence of a simulated environment does not by itself show that the agent’s discoveries will transfer to messy, changing real-world settings.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
WIRED identifies the move from confined domains such as Go to real-world complexity as a key challenge. The public sources cited here do not disclose a validated simulation design or show that this transfer problem has been solved. Important questions include how an agent gets useful feedback, how the environment represents the world, and whether skills learned there remain useful outside it.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What does this mean for AI safety?
Learning through action could help a system discover things that are absent from human-written data, but discovery alone does not guarantee that the result is safe or aligned with human interests. The WIRED interview raises the possibility that an agent could find solutions that conflict with people’s goals.
Studying an agent’s behavior in simulations, as Silver proposes, could provide a setting for observing what it does before deployment. Observation is not proof of safety: it matters what the simulation covers, which goals the agent is given, and whether the behavior carries over to other environments. The cited sources do not establish that Ineffable has solved these safety questions.
What can readers conclude about Ineffable’s bet?
Ineffable is betting that reinforcement learning can support a different foundation for AI: systems that build knowledge through experience rather than treating human-produced data as the primary source. AlphaZero shows that self-play can work remarkably well in tightly bounded games. The leap from that success to agents that make open-ended discoveries in the real world remains unproven.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchSequoia Capital, which announced a partnership with the company on April 27, 2026, described the work as hard and the timeline to superintelligence as uncertain. That is an investor’s perspective, not independent validation of the technical thesis. For now, the useful distinction is between a demonstrated game-playing precedent and Ineffable’s much larger goal: an ambition, not an established capability or a demonstrated replacement for LLMs.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
