The Final Frontier of Intelligence

AI is not a bubble, LLMs are. Language is a lossy compression of reality. You can predict the next word better and better, but eventually you have extracted every bit of structure that exists in text, and there is no more text to read. The internet was a one-time windfall, and it is nearly spent. The scaling laws make it concrete: every 10x more compute buys less than the last 10x did.

Synthetic data does not save you. Distilling a model from its own distribution only recombines what is already there. LLMs and agents are enormously useful, which is why they are worth hundreds of billions, but they cannot generate genuinely new knowledge about the physical world. They are a finite resource dressed up as an infinite one.

From Reading to Experience

You can already see the turn in where the compute goes. Models have started learning from their own experience: the model acts, the outcome is scored, and the trajectory becomes training data that never existed before. In modern reinforcement learning systems, most of the compute is spent generating these rollouts, not updating weights.

Today that learning hides behind scaffolding: long chains of reasoning, agent harnesses, elaborate prompts. The scaffolding works, but not well, and it is not the destination. Each generation, what the scaffolding did gets baked into the model itself as better priors, the scaffolding shrinks, and the loop starts again on harder problems. The costumes are temporary. The pattern underneath them is not: act, score, learn, repeat.

The fuel is no longer text humanity wrote. It is experience the machine makes for itself. And experience, unlike text, is not a finite reserve. It is manufactured, and the price of manufacturing it is compute and the energy that feeds it.

The frontier is moving from language to reality.

Recursive Self-Improvement

The interesting question is what AI does once LLMs and agents become commodity infrastructure. It stops consuming human knowledge and starts producing it. Google DeepMind called this a new golden age of discovery, and it is already underway: in late 2025 the U.S. Department of Energy launched the Genesis Mission to put AI to work across all 17 of its National Laboratories.

The mechanism is the same experience loop, pointed at the physical world:

1. The AI proposes something new: a molecule, a material, a design, a configuration.
2. A simulator models how it would behave in the real world.
3. The result tells the AI how to improve its next proposal.
4. The best candidates are checked against reality, in the lab or at higher fidelity.
5. Those real results make the simulator itself more accurate.
6. Repeat.

Early versions of this already exist. The crux is the fifth step. The system is not just searching, it is improving the thing that does the searching. Each turn produces better data, better data produces a better model of reality, and a better model produces better proposals on the next turn.

This is recursive self-improvement. It applies to the system itself, not just its model of the world. By mid-2026 Claude was writing most of the code at Anthropic, and its engineers had shifted from writing software to choosing which experiments to run. The same loop that discovers a new material improves the system doing the discovering. That is intelligence.

LLMs are bounded by how much humanity has written down. This loop is bounded only by how fast it can check its ideas against reality. That is still a limit, but a vastly larger one, and it grows every time the loop runs.

The Value of Computing

An LLM is consumption: a user asks a question, gets an answer, and the value ends there. An agent is automation: it does the work, and as that work gets cheap, things become abundant. Recursive self-improvement is production: AGI, ASI, whatever the label, it produces knowledge that did not exist before. Each cycle can yield a new drug, a new material, a new process, with enormous and lasting downstream value.

The world already spends nearly $3 trillion a year on research and development. As that abundance spreads, the figure climbs, because the only scarcity left is new knowledge, and new knowledge is what this compute produces. Every serious discovery problem becomes a market for it, and those markets compound as each discovery opens the next.

The constraint is mostly hardware. The chips we run AI on were each optimized for one kind of work, and this is not it.

These chips sit on a spectrum from flexible to fixed. The FPGA is reconfigurable down to the wire, able to become almost any circuit, but slow and inefficient at heavy math. The GPU runs anything and wins for that reason, though its generality leaves most of the chip idle on any workload that is not a steady stream of large, regular operations, and nearly a third of GPU users run below 15% utilization. The TPU is built from big systolic arrays optimized for big matmuls, which is what AI currently is, at high throughput and low precision (bf16, fp8, int8), fast but rigid. The ASIC burns a single model paradigm into silicon, fastest of all and useless for anything else.

This is not about what intelligence must be. It is about the shape of the work in front of us: three properties, at once.

1. Dense math, shrinking precision. GEMMs are still the workhorse of neural networks, and the precision they run at keeps falling: frontier models now ship natively in 8-bit and even 4-bit formats. Silicon spent on wide arithmetic is silicon the workload never touches, and the rare kernel that wants wider math can compose it from narrow, exact arithmetic in software.
2. Sparsity. Models get fatter in parameters and thinner in activity: DeepSeek V4-Pro has 1.6 trillion parameters and fires 49 billion of them per token. Intelligence is conditional and irregular, closer to a brain than a dense matrix.
3. State. The durable workload is the long, stateful trajectory: an agent working for hours, a rollout being scored, a simulation stepping through time. Every step depends on everything the task has seen so far, and carrying that state is memory traffic, not math.

No single point on the spectrum fits it. The TPU and the ASIC freeze yesterday's dense assumptions into silicon and struggle with the sparse, stateful rest. The GPU can run all of it, and pays for that generality in idle silicon and energy. FPGAs are out of the question.

Underneath all three sits the physics. On any modern chip, moving a byte costs orders of magnitude more energy than computing with it. The arithmetic is nearly free; the distance the data travels is the bill. And because datacenters are capped by power, energy per token is the number that decides what intelligence costs. A chip built for this era is designed around data movement first and arithmetic second.

There is also a bet hiding in every chip: which assumptions it freezes into silicon. Model families churn every few months, and a chip takes years. What persists through the churn is the shape of the work: dense math at shrinking precision, sparse access, and state that lives as long as the task does. The right chip bets on the shape, not the costume.

The bottleneck is not a missing idea or a lack of data. It is the hardware. The work needs dense throughput, sparse execution, and living state, at the lowest energy physics allows. That range is the "lagom" between a GPU and a TPU. Until the hardware catches up, the ceiling on what AI can discover, AGI included, is set by silicon, not ideas.

The Final Frontier of Intelligence

Get past the hardware and there is no ceiling on this. It is not bounded by how many researchers exist or how big the simulation market is today, but by how much is left to discover, which is effectively unlimited. Every problem solved opens new ones.

This is a self-improving loop. AI produces new knowledge, new knowledge produces better models, and better models produce deeper discoveries.

The question is not whether this becomes a trillion-dollar market. It is whether AI learns to make real discoveries. If it does, everything else follows. If it does not, none of this matters anyway.

We are building the machines that get us there. If that excites you, come build them with us: