What If We’ve Been Building AI Wrong?

One of the people who helped create the technology behind ChatGPT thinks we’ve been building artificial intelligence the wrong way.

So he spent the past two years working in stealth on a radically different approach to artificial intelligence.

Recently, he unveiled the result.

It’s called Jev.

And if it works as advertised, it could make some forms of AI up to 200X faster and 400X cheaper than today’s most powerful models.

But the most interesting thing about Jev isn’t how fast or cheap it is.

It’s why its creator believes we need it.

Why Isn’t AI Everywhere?

Jev was created by Diogo Almeida, who knows a thing or two about how today’s AI models work.

Almeida was part of the OpenAI team that developed InstructGPT, the technology that helped teach large language models (LLMs) how to follow human instructions.

That research became part of the foundation for ChatGPT.

After ChatGPT launched, Almeida watched as LLMs learned to write software, solve complex problems and outperform humans on an astonishing number of tests.

Yet something bothered him.

If these models were becoming so intelligent, why weren’t they producing the massive wave of automation that many researchers expected?

Almeida came to believe part of the problem was hiding in plain sight.

Large language models were built to generate language, which is incredibly useful when you need AI to write an email, explain a difficult concept or create a computer program.

But most business decisions don’t require a conversation.

Think about all the small decisions made every day, like whether a credit card transaction should be flagged as suspicious. Or which warehouse should fulfill an order. Or whether a customer should be transferred to a human.

These aren’t writing problems. They’re decision problems.

Those are the types of problems Jev was designed to solve.

You see, Jev wasn’t built to replace LLMs like ChatGPT or Claude.

Those models were designed to generate language. They produce answers one token at a time, which makes them incredibly useful for writing, coding and complex reasoning.

But Jev was designed to make decisions.

Give it information and a set of possible choices, and it can evaluate those choices at the same time and return a decision, along with a probability showing how confident it is.

For example, a company could use Jev to decide whether a credit card transaction looks like fraud and whether it needs to be reviewed by a person.

TypeSafe describes these kinds of decisions as “smart if-statements.” And there could eventually be billions of them.

And since Jev doesn’t have to generate all that language, it can make those decisions much faster and cheaper.

TypeSafe says Jev can return results in roughly 70 to 500 milliseconds and can be 40X to 200X faster than frontier LLMs on the tasks it was designed for.

The cost difference could be even more significant.

TypeSafe charges just $42 per billion input tokens. For comparison, that much input would cost about $4,000 with GPT-5.6 Sol and $2,000 with Claude Sonnet 5.

And unlike those models, Jev doesn’t charge for output tokens. That’s because it makes a decision instead of generating an answer word by word.

In TypeSafe’s own tests, that helped make Jev nearly 200X faster and more than 400X cheaper than competing LLMs while delivering comparable intelligence on certain tasks.

Now, TypeSafe acknowledges those numbers are probably on the high end of what users should expect in the real world.

But even a fraction of those improvements could change the economics of using AI.

Because the future of AI won’t just depend on making models smarter. It’ll also depend on making intelligence cheap enough to use everywhere.

And that will become especially important as AI agents become even more capable.

An agent can pursue a goal, use tools and decide what to do next. But one complicated task might require dozens or even hundreds of smaller decisions.

Which tool should I use? Did it work? Can I trust the answer?

Relying on a powerful LLM for every one of those decisions takes time and costs money.

Jev could potentially handle many of the smaller decisions while a powerful LLM tackles the problems that require deeper reasoning or language.

In other words, a frontier LLM might perform difficult research and generate an answer, while Jev decides which information matters and whether the result needs to be checked.

TypeSafe demonstrated this speed by connecting Jev to the classic video game Doom, where it makes roughly 10 decisions every second at a cost of about $7 an hour.

It proved that Jev can make intelligent decisions quickly and cheaply enough to operate in real time.

And that’s where Almeida sees a much bigger opportunity.

Here’s My Take

Jev is named after 19th-century economist William Stanley Jevons, who famously observed that making steam engines more efficient didn’t cause Britain to use less coal. It caused people to find more uses for it.

That’s the Jevons Paradox I’ve written about before.

Almeida believes the same thing could happen with AI.

If intelligence becomes dramatically faster and cheaper, companies won’t necessarily use less of it. They’ll find thousands of new places to put it.

And that could finally help answer the question that sent Almeida down this path in the first place.

If AI is already so intelligent, why isn’t it everywhere?