David Silver and Ineffable Intelligence

Who he is

Silver was one of DeepMind’s first employees when the company started in 2010, and later led its reinforcement learning team. He is still a professor at University College London. He has been a friend of Demis Hassabis since they met at a chess tournament as kids.

He was the lead researcher on AlphaGo, the program that beat the world’s best Go players in 2016. The same research line produced AlphaZero, MuZero, AlphaStar, and later AlphaProof. He is also a co-author on the original 2023 Gemini paper.

He left Google DeepMind in January 2026. Fortune reported that staff were told earlier that month, after a sabbatical, and that U.K. filings show Ineffable was formed in November 2025, with Silver appointed a director on January 16, 2026.

Ineffable Intelligence

Ineffable is a London AI lab founded in late 2025 (November 2025). Silver is founder and CEO. The public mission is to “make first contact with superintelligence” by building a superlearner: a system that discovers knowledge and skills from its own experience, through trial and error, rather than from human data.

The company site says that capability will be driven by reinforcement learning, and that a single superlearner should apply to any digital or physical machine and any stream of experience, then keep learning. Silver wrote that other kinds of AI (language, video, code) are already in good hands, and that this experiential path is the work he wants a dedicated lab for.

In April 2026 the lab raised a $1.1 billion seed at a $5.1 billion valuation, reported as Europe’s largest seed round. Sequoia and Lightspeed co-led, with Nvidia, Google, DST Global, Index, and the U.K. Sovereign AI Fund among the participants. Silver has said any money he makes from Ineffable will go to high-impact charities.

How this differs from the LLM approach

Most of the field is building large language models, then adding extra training and tools on top. Silver’s bet is that this stack is bounded by human data, and that the next jump comes from agents that act, get rewards, and learn from consequences. The two paths already mix. The difference is which one is the engine.

How today’s language models are built

Modern LLMs (GPT, Claude, Gemini, Grok) sit on the transformer architecture from Attention Is All You Need (2017). In pretraining, a model reads a huge amount of human-written text from the internet, books, and similar sources, and learns to predict the next token. The 2020 GPT-3 paper showed that this next-token objective, at scale, produces a general text model that can be prompted into many tasks without being rebuilt for each one.

Scaling laws made that bet operational. Kaplan and colleagues (2020) found that loss improves in a regular way as you add parameters, data, and compute. Hoffmann and colleagues (Chinchilla, 2022) argued that many large models were undertrained relative to their size, and that data and parameters should be scaled together. The practical lesson labs took is simple: more high-quality tokens and more compute still buy capability.

Raw next-token models are not yet helpful assistants. Post-training reshapes them. Supervised fine-tuning (SFT) shows the model example conversations and answers. Reinforcement learning from human feedback (RLHF), described in OpenAI’s InstructGPT paper (2022), trains a reward model from human preference rankings, then optimizes the policy against that reward. Anthropic’s Constitutional AI (2022) and later RLAIF work replace some of that human ranking with AI feedback against a written constitution. The learning signal is still “what a rater, or a constitution written by people, prefers,” not “what happened in the world.”

A newer layer is reasoning models trained with reinforcement learning on checks you can score automatically: a math answer that is right or wrong, a program that passes tests. OpenAI’s o1 (2024) is an LLM trained with large-scale RL to use a long internal chain of thought before answering. The follow-on o3 models (2025) add tool use inside that reasoning loop. DeepSeek’s R1 paper (2025) showed that a base model can grow reasoning behaviors from RL on verifiable rewards, even without a big supervised “how to think” dataset first (DeepSeek-R1-Zero), then added a cold start to make the traces more readable (DeepSeek-R1).

Labs also generate synthetic data: model-written text, solved problems, and self-play traces used as extra training fuel. That can stretch a fixed pile of human text. Silver and Richard Sutton argue it is still a static recipe. In Welcome to the Era of Experience (2025) they write that data has to keep improving as the agent gets stronger, and that any fixed synthetic generator will be outstripped. The data they want is experience the agent creates by acting in an environment.

The “data wall” is the worry that public human text is a finite stock. Epoch AI estimated the useful public human-text supply at around 300 trillion tokens, and projected that current scaling trends would use that stock between 2026 and 2032, earlier if models are heavily overtrained (Villalobos et al., 2024). At NeurIPS 2024, Ilya Sutskever put it more bluntly: pretraining as we know it will end, because “we have but one internet” and “we’ve achieved peak data” (The Verge, Reuters). Silver uses a similar picture. In WIRED he called human data “a kind of fossil fuel that has provided an amazing shortcut,” and systems that learn for themselves “a renewable fuel.”

How the major labs frame the path

The large labs are not identical, but most describe a path that still starts with scaled pretrained models, then adds alignment, reasoning RL, tools, and agents.

OpenAI. Planning for AGI and beyond (2023, updated 2025) defines the mission as AGI that benefits humanity, and treats scaled, deployable models plus better alignment as the near path. InstructGPT and ChatGPT were the first public alignment step. o1 and o3 are the reasoning step: RL on the model’s chain of thought, then agents that can use tools. The stack is still an LLM, made more agentic.
Anthropic. Claude is a scaled language model shaped by Constitutional AI and related preference training. Dario Amodei’s Machines of Loving Grace (2024) talks about “powerful AI” as something that could arrive soon and then change science and the economy. The method in public is still: train a large model on human data, then steer it.
Google DeepMind. DeepMind is the lab Silver just left, so it already contains both traditions. Gemini is a frontier multimodal LLM (2023 paper). The same organization built AlphaGo, AlphaZero, MuZero, AlphaStar, and AlphaProof: agents that learn from interaction and search. The public product path still leans on Gemini-scale models. Silver’s complaint, in WIRED, is that an experience-first lab should not be “a corner of another place dedicated to LLMs.”
xAI. The company mission is to build AI to understand the universe. In public that has meant Grok: large models, then reasoning variants and agents (Grok 3 was framed as “the age of reasoning agents”). The ingredients look like the rest of the LLM field, with a first-principles brand around them.
Meta. Meta’s open Llama line is a classic pretrain-then-fine-tune stack. In June 2025 Mark Zuckerberg created Meta Superintelligence Labs, led with Alexandr Wang, and wrote that “developing superintelligence is coming into sight,” aimed at “personal superintelligence for everyone” (CNBC). The organization still houses foundation-model, product, and FAIR teams. That is a superintelligence goal on top of the Llama-style path, not a restart from self-play.

So the mainstream picture is: scale next-token models, spend the remaining human data carefully, synthesize more, add RL where rewards are cheap to check, and wrap the result in tools and agents. Silver thinks that can be very useful and still cap out near what humans already know.

Critics and other bets

Not everyone in the field thinks “bigger LLM, plus RL on the side” is the whole story. Three other public alternatives help place Ineffable.

Yann LeCun has argued for years that autoregressive LLMs are the wrong end state. In A Path Towards Autonomous Machine Intelligence (2022) he proposes agents built around a predictive world model, trained with self-supervised Joint Embedding Predictive Architectures (JEPA): learn abstract representations that are useful for predicting the next state, then plan against a cost. The critique is sample efficiency and common sense. Animals learn a model of the world from observation, not from predicting the next word. LeCun later left Meta’s AI-chief role and started AMI Labs; that is a world-model bet, not Silver’s RL-from-experience bet, even though both reject “LLM as the only path.”

Richard Sutton is the other pole. His 2019 essay The Bitter Lesson says the methods that win, over decades, are general ones that scale with compute (search and learning), not ones that bake in human knowledge of the domain. “We want AI agents that can discover like we can, not which contain what we have discovered.” Sutton and Andrew Barto received the ACM Turing Award for reinforcement learning (announced in 2025). That is the experience-learning tradition Silver comes from. In 2025 Sutton and Silver wrote Welcome to the Era of Experience together. The essay is the intellectual backbone of Ineffable: agents in long streams of experience, observations and actions grounded in an environment, rewards from outcomes rather than human prejudgment, and planning about those consequences rather than only in human language.

Ilya Sutskever’s Safe Superintelligence (SSI), announced June 19, 2024, is a different kind of alternative. SSI is a “straight-shot” lab with one product, safe superintelligence, insulated from ordinary product cycles (Reuters). Sutskever has said pretraining will end because the internet is a finite dataset. SSI has not published an experience-first recipe the way Silver has. Fortune grouped the two as researchers who left big labs to chase superintelligence directly. The shared piece is focus. The technical thesis is not the same, at least not in public.

Why Ineffable’s superlearner is a different bet

Silver is not saying RL is unused today. He is saying the field is using it as a finisher on a human-data model. Ineffable wants the learner itself to be an experience machine: act, see what happens, update, forever.

AlphaZero is the clean example. It learned chess, shogi, and Go from self-play, without human games, and found strategies people had not played (Silver et al., 2017). AlphaGo’s Move 37 against Lee Sedol in 2016 looked wrong to human commentators and later looked decisive. Fortune’s point, quoting Silver’s circle, is that an RLHF-style human rater might have downranked that move for looking like a mistake. Human preference can block the very discoveries you wanted.

In a WIRED thought experiment, Silver asks what happens if you drop an LLM into a world that believes the earth is flat. Without acting on the world, a better language model stays a more fluent flat-earther. An agent that can look, measure, and be rewarded for being right can change its mind. That is the claim about grounding and ceiling.

	Mainstream LLM stack	Ineffable-style superlearner
Data source	Human text (and images, code), plus synthetic text from other models	The agent’s own interaction with environments and simulations
Learning signal	Predict the next token; then imitate demos; then human or AI preference; then verifiable scores on narrow tasks	Reward from outcomes: win, prove, reach a goal, get a grounded environmental signal
Grounding	Mostly language about the world	Actions and observations in the world (or a simulation of it)
Ceiling	Strongly shaped by what humans wrote, labeled, or preferred	In principle, whatever the environment can teach, including strategies no human has seen
Continual learning	Usually a training run, then a frozen model, with occasional post-training	Keep learning from a lifelong stream of experience
Generality	Broad language and knowledge now; weaker where text is thin or feedback is slow	Claimed across any stream of experience, if the environment and reward can be defined
Strengths today	Chat, coding, writing, and a lot of world knowledge, immediately useful	Discovery in domains with clear feedback (games, formal proof, some simulations)
Hard parts	Hallucination, finite human data, preference hacking, weak real-world grounding	Reward design, building rich enough environments, sample efficiency, safety of open-ended goals

They can be combined, and they already are. o1, o3, and DeepSeek-R1 are language models using RL on checkable rewards. AlphaProof, in the Era of Experience essay, starts from a human formal-proof corpus and then generates a much larger set of proofs by interacting with a proving system. Silver’s claim is not “never use human data.” It is that human data is a bootstrap, and that the system which keeps going will be the one that can generate its own curriculum.

That is what Ineffable means by experience-first. The lab is trying to make the AlphaZero move in the open world: not a chatbot that sometimes uses RL, but a learner whose main food is consequences.

Open challenges

This is a harder engineering problem than “get more tokens.” The open questions are old RL problems, now aimed at everything.

Reward design. Games have a score. Formal math has a checker. Open-ended science, economics, or “be beneficial” does not. A misspecified reward can be hacked. Sutton and Silver talk about rewards grounded in experience rather than human prejudgment, but someone still has to choose which signals count.
Environments. Go is closed and cheap to simulate. The real world is not. Silver told WIRED he wants agents inside simulations that can learn goals and cooperation. He has not published what those simulations are. The jump from board games to open-ended reality is the whole product risk.
Sample efficiency. Self-play can burn enormous numbers of games. Language models get a dense signal from every token of human text. An experience-first agent needs either huge compute, a very good world model, or both. LeCun’s JEPA argument is partly that pure trial-and-error is too wasteful without prediction.
Safety. An agent that can find “alien” strategies can also find strategies people would not endorse. Silver’s public answer is to watch behavior in simulation, including how a stronger agent treats weaker ones. That is a research program, not a finished control method. WIRED notes the alignment worry directly. Ineffable’s site says superintelligence “can and must be built in a way that is beneficial to humanity.” How, in detail, is still open.

None of this makes the LLM path obviously right either. Hallucination, sycophancy, and the data wall are real. The honest read is that the field is splitting into a human-data stack that works now, and an experience stack that has already beaten humans in closed worlds and has not yet shown the same thing in the open one. Ineffable is a large bet that the second stack is the one that goes past human knowledge.