Better AI Starts With a Better Task, Not a Bigger Computer
Data has to be invented for each new job, which is why adding computing power alone stops paying off.

The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch · 20VC: Jev Creator on Why Data Is More Important Than Compute | Will US Open-Source Models Catch Chinese | The Cost of Intelligence and Tokens: What No One Sees But Needs to Know with Diogo AlmeidaMost people assume smarter artificial intelligence comes from bigger computers. Diogo Almeida, co-founder of AI startup Jev, says that gets the order backward.
His ranking, from most to least important, is task first, then data, then computing power, then algorithms. Almeida says, “you get what you optimize for,” so the definition of the job matters more than the size of the machine.
Before Jev, he says, language models had two main tasks. One was training from human preferences to please people, and the other was reinforcement learning from verified rewards, which means training it to score high on benchmarks.
Neither task trains software to work reliably in the background, making sound decisions without a person checking. A chatbot can be unfailingly polite and fluent, yet still make bad judgments.
Once the right task is set, the data for it usually does not exist yet. When researchers created human feedback training, no internet archive contained comparisons of chatbot answers, so people had to write instructions, hire labelers to pick better responses, and iterate until they learned what good data looked like.
The same was true for verified rewards. Earlier efforts trained models on pass or fail software tests and kept failing. The big part of the shift was that models could reason privately, in a hidden chain of thought, before answering, so the structure of the task mattered more than the verified reward itself.
Almeida doubts that a corporate data lake is a shortcut. The information there, he argues, is probably not organized for the new task, and vendors may need to rework it to make it useful. Data collection only helps if it is judged by whether it moves the chosen task forward.
Questions people ask
What is reinforcement learning from human feedback?
It is a training method where people compare two chatbot answers and say which is better. The model learns to favor responses humans prefer, which makes it helpful and polite, but does not necessarily make its decisions correct.
What is hidden chain-of-thought reasoning?
It means the model writes out intermediate reasoning steps that the user does not see, then gives a final answer. That private draft lets it work through math, code, or logic problems more reliably than jumping straight to an answer.
Why can't a company just train AI on its own stored data?
Most business records were kept to run the business, not to train a model. Emails, invoices, and logs are messy, incomplete, and in the wrong format, so they must be cleaned, relabeled, and reshaped around the specific judgments the model needs to learn.

