A former Google DeepMind researcher has secured $1.1 billion in funding to build Ineffable Intelligence, a startup pursuing a fundamentally different approach to artificial intelligence — one that forgoes the human-generated training data that underpins today's dominant AI systems in favour of reinforcement learning techniques the founders believe could lead to superintelligence.
Ineffable Intelligence, a newly prominent AI startup founded by a veteran of Google DeepMind, has raised $1.1 billion to develop artificial intelligence systems trained without human-curated data — a direct challenge to the large language model (LLM) paradigm that has defined the industry in recent years.
The substantial funding round, reported by Decrypt, signals growing investor appetite for alternative approaches to AI development at a time when companies like OpenAI, Anthropic, and Google are pouring billions into scaling up LLMs built on vast repositories of human-generated text and images.
A Different Path to Intelligence
Ineffable Intelligence is centring its research on reinforcement learning (RL), a technique in which AI agents learn by interacting with environments and receiving feedback based on their actions — rather than by ingesting and pattern-matching across pre-labelled human data. Proponents argue this mirrors more closely how intelligent beings actually learn, through experience and consequence rather than rote absorption of existing information.
The approach is not new to AI research. Reinforcement learning was central to some of DeepMind's most celebrated achievements, including AlphaGo and AlphaZero — systems that mastered complex games such as chess and Go at superhuman levels by playing against themselves, without being trained on records of human gameplay.
The founders argue that scaling LLMs — feeding them ever-larger datasets of human-produced content — is approaching fundamental limits, and that genuine machine intelligence will require systems capable of generating and testing their own knowledge rather than recombining what humans have already written.
A Crowded and Competitive Field
The $1.1 billion raise is among the larger early-stage funding rounds in the AI sector, reflecting both the enormous capital requirements of frontier AI research and the intensity of competition to identify the next breakthrough architecture. It also underscores how investors are beginning to hedge bets beyond the LLM-centric approaches that have dominated the past several years.
Other research groups and startups have explored reinforcement learning and related self-supervised techniques, but few have attracted funding at this scale specifically premised on moving away from human training data entirely.
The company's name — "ineffable," meaning too great or extreme to be expressed in words — suggests its founders have ambitious expectations for what their systems might ultimately achieve, though no specific products or timelines have been publicly disclosed.
Risks and Open Questions
Reinforcement learning at scale carries its own well-documented challenges. RL systems can be notoriously difficult to train stably, prone to finding unexpected shortcuts, and computationally expensive. Critics of the approach note that while RL has produced impressive results in constrained, rule-based environments, extending it reliably to the open-ended complexity of the real world remains an unsolved problem.
Whether Ineffable Intelligence can translate its founders' DeepMind pedigree and substantial funding into a viable alternative to LLMs remains to be seen. The company has not yet announced a product roadmap or detailed its specific technical methodology.
Analysis
Why This Matters
- The $1.1 billion raise is a significant vote of confidence that the LLM paradigm may not be the final word in AI development, potentially shifting research and capital toward alternative approaches.
- If reinforcement learning without human data proves viable at scale, it could reduce AI's dependence on the enormous — and ethically contested — datasets of human-generated content currently used to train systems.
- The founding team's DeepMind background lends credibility to what might otherwise be dismissed as contrarian positioning, and could attract top research talent away from established labs.
Background
The dominant approach to building powerful AI systems over the past decade has relied heavily on large language models trained on vast corpora of human-generated text scraped from the internet, books, and other sources. This paradigm produced GPT-4, Claude, Gemini, and their contemporaries, and has delivered striking capabilities in language, reasoning, and code generation.
Reinforcement learning, by contrast, has a distinguished but narrower track record. DeepMind's AlphaGo (2016) and AlphaZero (2017) demonstrated that RL-based systems could achieve superhuman performance in games like Go and chess without human training data — learning entirely through self-play. These results were widely seen as proof-of-concept for a fundamentally different kind of machine intelligence.
However, extending RL beyond well-defined game environments to general-purpose intelligence has proven far more difficult. The field has seen periodic waves of enthusiasm and disappointment, and many researchers argue the two approaches — LLMs and RL — are complementary rather than competing. OpenAI itself uses reinforcement learning from human feedback (RLHF) to fine-tune its models, blending both paradigms.
Key Perspectives
Ineffable Intelligence and backers: Believe scaling LLMs is hitting diminishing returns and that reinforcement learning without human data is the more promising route to genuinely novel, general intelligence — potentially superintelligence.
Mainstream AI labs (OpenAI, Anthropic, Google DeepMind): Have largely continued to bet on LLM scaling, while incrementally incorporating RL techniques. Most researchers at these organisations view LLMs and RL as complementary, not mutually exclusive.
Critics and skeptics: Point out that pure reinforcement learning at scale outside constrained environments remains an unsolved engineering and research challenge. Some argue the framing of "no human data" may be more marketing than methodology, and question whether the approach can generalise to the messy, open-ended real world.
What to Watch
- Whether Ineffable Intelligence publishes research or benchmark results that demonstrate tangible advantages over LLM-based systems in open-ended tasks.
- Announcements of further funding rounds or strategic partnerships, which would indicate sustained investor confidence beyond the initial raise.
- How established labs respond — any pivot toward pure RL approaches by OpenAI, Anthropic, or Google DeepMind would signal the industry taking this challenge seriously.