Over the past few years the industry has built up a strong path dependency: if a problem needs "intelligence", call a large language model.
Classification, call GPT. Extraction, call GPT. RAG reranking, call GPT. Text2SQL, call GPT. Tool selection in an agent, call GPT. Checking whether the result is correct — call GPT again.
The models keep getting better, and we keep getting more comfortable stuffing every problem into the same prompt. I increasingly think this is only a transitional state from the early years of large models.
We are asking the CEO which department gets the email
Take Text2SQL. A user asks: "how many X-700s did Acme deliver in Asia in 2025?" The system has quite a lot to do.
Work out which entities "Acme" and "X-700" refer to, decide which tables to query, choose the fields, determine how the tables join, generate the SQL, then check whether that SQL actually answers the question that was asked.
The easiest thing to do today is hand all of it to a large model. It genuinely can do it. But it is worth asking: do these tasks need the same kind of intelligence?
- Which three of these twenty tables?
- Are Acme and Acme Industries the same entity?
- Which five of these fifty passages are most relevant?
- Should this request go to Search, Text2SQL or the knowledge graph?
- Of SQL A, B and C, which is most likely correct?
Underneath, these are all the same kind of question: judge, choose, rank, score. None of them needs the model to write prose, and none needs a long chain of reasoning.
Using a general model with hundreds of billions of parameters for this is like a company where both "which department should this email go to" and "what is our five-year strategy" go to the CEO. It works. It is certainly not the best way to organise.
What is interesting about Jev is not that it is fast
TypeSafe's recent release of Jev made me take this more seriously. It isn't trying to be another GPT, and generating language isn't even its goal. They position it as "the first System One model" — making fast, structured decisions that software can use directly, rather than producing text for humans to read.
Its interface has just three primitives:
| Primitive | The question | What comes back |
|---|---|---|
| Choice | Choose an option from a list | choice, probabilities, confidence |
| Score | Score the state on a rubric | score, probabilities, confidence |
| Noul | Is this statement true? | a value between 0 and 1 |
You give it a state, ask a set of typed questions against that state, and get structured results back. No text generation, and therefore no parsing. In their words, you stop "coercing a text-generation system into outputting structured decisions, then parsing the results back into something your code can depend on."
At first glance this isn't exciting — classifiers have existed for decades. The real change is in how a task gets defined.
In the classical machine-learning era, deciding whether an article is about a delivery, an acquisition, an accident or a new product meant collecting data, labelling it, training, evaluating, deploying — and possibly retraining when the business rule changed. Now the same model can be asked "which of these categories does this article belong to" today, and "is this article worth pushing to the sales team" tomorrow. The model didn't change. The task did.
Task definition moved from training time to call time. That, to me, is the part that matters.
Their documentation also makes a recommendation I agree with: a complex judgment should be decomposed into atomic questions and composed in code, rather than asking one prompt to walk the whole reasoning chain.
AI is repeating the history of computers
For the past few years we have quietly conflated two things: language intelligence and general intelligence. Because GPT first arrived as a language model, and because language happens to be able to express almost any problem, we slid into assuming that every intelligent task should pass through language generation.
But a system needs several different kinds of intelligence. What is this? Which one is more relevant? Where do I go next? Is this result trustworthy? And — faced with a complex problem I have never seen, how do I understand it, break it down, and solve it?
Only that last one is where a frontier model genuinely earns its place.
Which reminds me of computers. Nobody today asks whether the CPU will eventually kill off the GPU, because the question doesn't hold: they handle different kinds of computation. Graphics and massive parallelism go to the GPU, general control flow to the CPU, video encoding has its own block, networking has its own hardware.
Modern computers are powerful not because we built one processor that can do everything, but precisely because we accepted that different computation belongs on different hardware.
AI will probably take the same road. A mature AI system likely divides the work like this:
| Component | Responsible for |
|---|---|
| Frontier LLM | Complex reasoning, planning, open problems, exception handling |
| Decision model | Classification, scoring, routing, reranking, verification |
| Embeddings | Large-scale semantic recall |
| Vision / speech models | Their own modalities |
| Search engine | Connection to the live world |
| Database / knowledge graph | Deterministic facts and relations |
| Conventional software | Everything that needs no probabilistic intelligence at all |
Added together, that is the complete AI. That is what I understand by heterogeneous intelligence.
GPT sits higher and gets called less
None of this makes GPT less important. If anything I think its position in the system keeps rising — while the number of calls to it does not.
Today an agent completing one complex task might call GPT ten times. In future it might be: twenty calls to a decision model, five database queries, three web searches, a few conventional programs, a few small models — and one call to the strongest model at the end.
And on that one call, its job is to understand the real goal, form a plan, handle the uncertainty, solve what no other module could, synthesise everything the system has gathered, and produce the final answer.
GPT goes from doing everything itself to being the brain in the real sense. Fewer calls does not mean less strategic weight — the CPU didn't disappear when GPUs became ubiquitous. It simply stopped having to perform every calculation personally.
What is under threat isn't GPT, it's "one API call = an AI feature"
So what models like Jev really challenge isn't GPT. It's a development paradigm that formed over the last few years:
An AI feature = one call to an LLM API.
That phase may be ending. When designing an AI product, the first question should no longer be "GPT or Claude here?" but: which kind of intelligence does this step actually need?
Needs retrieval, use retrieval. Needs a judgment, use a decision model. Needs a live fact, use search. Needs an exact calculation, use code. Needs structured facts, use the database and the knowledge graph. Only when the problem is genuinely complex, unknown and open-ended do you call the strongest model you have.
From model intelligence to system intelligence
The question the industry cared about most these past few years was which model is smartest — GPT, Claude, Gemini, whose benchmark is two points higher.
That was never what companies actually care about. They care about what a real task costs, how long it takes, whether the result is reliable, and whether it can run at scale.
At that point the unit of competition is no longer a model but an entire intelligent system. The industry moves from model intelligence to system intelligence.
A genuinely good AI product doesn't necessarily use the strongest model in the world. It picks, at every step, the intelligence whose cost, speed and capability fit best.
We have spent years waiting for an ever-stronger model: more parameters, longer context, deeper reasoning, higher benchmarks, eventually capable of anything. That road will continue. But from an engineering point of view another road matters just as much — not making one model ever more universal, but making the whole system smarter.
The strongest AI of the future may not be a model at all, but a group of very different intelligent modules: some that think, some that judge, some that search, some that remember, some that execute, some that verify. Just as a computer today is not a chip but a whole architecture of computation.
A large model is not AI's final form. It is the single most important component of a heterogeneous intelligent system. The next round of competition may shift from who owns the strongest model to who can assemble the most efficient, most reliable, cheapest intelligent system.
That is probably the point at which agents genuinely start to mature.