The next scaling law should be about cost.
Modern AI has learned to do more by becoming larger. The next step is to make capability grow faster than computation, memory, latency, and financial cost.
Read the articleEssays & research notes
Thinking about AI,
in public.
Articles is where we explain what we think modern artificial intelligence is getting right, what it is still missing, and why efficient intelligence matters for the path ahead.
- Point of view
- Direct · technical · independent
- Primary question
- What should modern AI become?
- Topics
- Efficiency · reasoning · agents
- Audience
- Researchers · builders · the public
Featured article
Perspectives
Ideas we return to.
Short positions that shape our research questions, experiments, and collaborations.
Bigger is not the same as better.
Scale can hide weak structure. A capable system should use its parameters, memory, and time deliberately.
Lab perspective · 4 minCompression is a way to study intelligence.
What a model can lose without changing what it knows tells us something important about how its capability is organized.
Research note · 5 minAgents need loops, not only prompts.
Useful agency depends on memory, feedback, verification, and the ability to revise a plan after the world answers back.
Systems note · 6 minThe cost of intelligence shapes who can use it.
Efficiency is not only an engineering metric. It determines which researchers, institutions, and communities can participate.
Lab position · 4 minFounding note
The next scaling law should be about cost.
The most important question is no longer whether scale works. It does. The question is how long capability must remain tied to ever larger budgets.
Modern artificial intelligence is unusually good at turning computation into capability. Larger models, larger datasets, and longer training runs have produced systems that can write, code, reason, and use tools. This progress is real. So is its cost.
When each improvement requires more memory, more energy, more specialized hardware, and more capital, intelligence becomes concentrated. Only a small number of organizations can train the frontier, reproduce its results, or test a genuinely different idea.
A better model is not only one that knows more. It is one that does more with what it has.
Efficiency is part of capability.
We do not see efficiency as cleanup after a large model has already been built. It is a first-class research problem. A system that reaches the same result with less depth, lower precision, or a better learning rule has discovered a more useful form of intelligence.
This is why compression, pruning, optimization, and efficient architectures matter. They force us to ask what the model actually needs. They reveal which layers are redundant, which gradients carry information, and which parts of a reasoning loop create real improvement.
Reasoning is a systems problem.
A language model is only one part of an intelligent system. Memory, tools, feedback, verification, and coordination determine whether an agent can turn a plausible answer into reliable work. The quality of the loop matters as much as the size of the model inside it.
Our goal is Fast Lightweight AGI. That means systems that are computationally practical, financially accessible, and strong enough to learn, reason, and coordinate across difficult tasks. We want to build the methods, models, and open evidence that make this path credible.
Editorial direction
What belongs here.
We write when a research result changes our view, when a common assumption deserves another look, or when technical work has a broader consequence worth explaining.
01
Clear positions supported by technical reasoning.
02
Research notes that explain what an experiment taught us.
03
Practical views on models, agents, and the cost of intelligence.
04
Ideas we are willing to revise when the evidence changes.