While the headlines obsess over superintelligence, a quieter but equally well-funded movement is building AI that works with you, not instead of you.
The dominant narrative in AI right now is straightforward: bigger models, more autonomy, less human involvement. OpenAI, Anthropic, and Google are racing to build systems that can write entire software applications from scratch, execute multi-step tasks without supervision, and, if the more bullish forecasts prove right, eventually surpass human cognitive ability across the board.
But a growing cluster of high-profile labs is making a different bet. Their argument isn’t that this race is wrong, exactly — it’s that it’s incomplete. That the most transformative thing AI can do isn’t replace human judgment, but sharpen it.
Three companies stand out in this counter-movement, each attacking the problem from a distinct angle.
Thinking Machines Lab: AI That Understands Intent
Thinking Machines was founded in 2024 by prominent engineers from OpenAI, and quickly raised a record $2 billion seed round, becoming one of the most-watched startups in AI.
What makes Thinking Machines distinctive isn’t raw capability — it’s the interaction model. Most AI systems today operate on a simple turn-taking structure: you write something, the model responds, you write again. The founding team argues this is a fundamental constraint, not just a design choice.
The vision is to build a model that « constantly perceives what you’re doing and is constantly there to be able to reply and give you information or search for information or use other tools. » The critique of current systems is that « the turns in a conversation are determined by a much less intelligent system » — meaning the structure of interaction is designed around technical limitations, not human cognition.
Technically, Thinking Machines is working on what might be called continuous perception models — systems that maintain persistent awareness of context rather than resetting with each exchange. This requires rethinking how attention mechanisms handle streaming input, how memory is structured across time, and how models learn to predict user intent rather than merely respond to explicit prompts. Their first product, Tinker, launched in October 2025, is an API that lets researchers and engineers fine-tune open-source models on custom data, giving an early glimpse of the personalization layer they’re building toward.
The company’s core vision is one of amplification: AI that understands and predicts human intent, shaping itself around individual preferences and values rather than operating as a generic tool.
AMI Labs: Rejecting the LLM Paradigm Entirely
If Thinking Machines and Humans& are reforming the existing approach, AMI Labs is staging a full rebellion against it.
AMI Labs (Advanced Machine Intelligence, pronounced ami — the French word for « friend ») was founded in Paris in 2025 by a Turing Award winner and former Chief AI Scientist at a major technology company. In March 2026, the company announced it had raised $1.03 billion at a $3.5 billion valuation.
AMI Labs’ founding thesis is that LLMs are structurally incapable of true intelligence — not because they aren’t powerful, but because they learn from language rather than from reality. The argument, developed over many years of research, is that language is an impoverished representation of how the world actually works. A model trained to predict the next word will hallucinate because fabrication is, in a sense, baked into its nature.
Technically, AMI Labs is built around JEPA — Joint Embedding Predictive Architecture. Unlike LLMs that operate by predicting the next token in a sequence, JEPA learns in latent space — building abstract representations of reality from sensory data (cameras, sensors, continuous environmental input) rather than from text. Early internal benchmarks reportedly show 2–10x better GPU utilisation compared to equivalent LLMs on physical reasoning tasks. The company’s first applications are targeting healthcare (in partnership with Nabla), robotics, wearables, and industrial automation.
The company is candid about the timeline: AMI Labs is a very ambitious project, starting with fundamental research rather than applied engineering. It could take a year just to get the first things usable in a product — a deliberate trade-off for a team that believes the current paradigm needs replacing, not iterating.
The Intellectual Backdrop: The Case for Augmentatio
This movement didn’t emerge in a vacuum. A prominent venture capitalist and LinkedIn co-founder — an early OpenAI backer and one of Silicon Valley’s most influential thinkers on the future of work — has become one of the most articulate voices for the augmentation thesis.
In his 2025 book Superagency, the argument is made that AI’s most transformative potential lies not in replacing human capability but in extending it. The vision: a future in which every knowledge worker operates with the leverage of an entire team, directing agents that do parallel work while humans handle judgment, creativity, and coordination.
On the enterprise side, the diagnosis is that companies have been implementing AI wrong — treating it like « isolated pilots » rather than weaving it into the coordination layer of work itself. « AI lives at the workflow level, and the people closest to the work know where the friction actually is. » The 2026 prediction: any company that wants to remain a « thriving, growing concern » will need to be running AI across their meeting cadences, their knowledge flows, and their team structures — not just their individual productivity.
The argument is also explicit about the right posture for AI development. As individuals become capable of directing teams of AI agents, the skills that matter most shift: delegation, critical thinking, coordination — once the domain of managers — become foundational across the entire workforce. That’s not a world where AI replaces humans. It’s a world where humans become more capable, not fewer.
Three Bets, One Direction
These three companies are not building the same thing. Their technical approaches diverge significantly:
- Thinking Machines stays within the language model paradigm but rethinks the interaction layer — continuous, ambient, intent-aware.
- Humans& builds a new architecture for social intelligence — multi-agent coordination, group memory, collaborative decision-making.
- AMI Labs rejects language models as a foundation entirely — grounding AI in physical reality via JEPA and sensory data.
What unites them is a shared conviction that the automation-first race is optimising for the wrong thing. The question isn’t only how much work AI can do autonomously. It’s whether the humans left in the loop — or pushed out of it — are more capable, more empowered, and better served than they were before.
That question, more than any benchmark, may turn out to be the one that matters most.
Published May 2026.