For a few years, every new AI model felt like a jump. You'd try the latest one and something that was impossible last spring just… worked. It held a long argument together, and it stopped making the mistakes that used to give it away. It was easy to assume the pattern would hold forever: wait nine months, get a smarter machine.
Lately it feels different. The new models are better, but you have to squint. Benchmark scores creep up a point or two, and the labs are honest that each increment costs far more than the last. For most everyday work, the difference between this year's best model and last year's is small, and much smaller than the difference the year before that.
The comfortable read is that AI is slowing down. We think that's wrong. Progress didn't end; the bottleneck moved, and almost nobody is spending their effort where it moved to.
The models are already smart enough
For the overwhelming majority of everyday tasks, today's models are not failing because they aren't clever enough. They're failing because they know nothing about today and can't touch anything outside the chat window.
A model is trained once, on a snapshot of the world, and then sealed. It's like hiring a brilliant graduate who has read every book ever written, locking them in a room with no phone and no colleagues, and sliding questions under the door. They answer everything, and they answer beautifully. They also have no way of knowing that the price changed on Tuesday or the item sold out an hour ago.
Worse, they have no way to tell that they don't know. A model's job is to produce the most plausible-sounding answer. When it has the facts, plausible and true are the same thing. When it doesn't, plausible is a confident guess wearing the same voice as a correct answer. That last part is what breaks trust. People get things wrong too, of course, but a person's voice usually changes when they're unsure. A model's doesn't.
A bigger brain does not fix this. You cannot reason your way to a stock level, and ten times the training will not tell you what's in the warehouse right now. This is a whole category of failure that model size has no grip on, which is why more of it stopped feeling like magic.
Where it breaks in practice
Put a smart model on ordinary questions and watch where it breaks. It sails through the hard reasoning and then trips on the simple, checkable things.
| You ask | A sealed model answers | A connected one answers |
|---|---|---|
| Is this in stock? | "It should be available." Guessing from a snapshot. |
"In stock, 3 left." Checked against the live source. |
| What's the rule this year? | "As of my training data…" Honest, and still useless. |
"Here's the current rule, and where it came from." Read today, with a source. |
| Book it / add it / send it. | "Here's a link, I think." Can describe the action, can't take it. |
"Done, and confirmed it went through." Acted, then verified the result. |
Notice that the right-hand column doesn't need a smarter model. It needs a model with a phone line to the real world. Every one of those failures is an information and access problem dressed up as an intelligence problem.
The next leap is the room, not the brain
So here's our bet. For the last few years, "make AI better" meant "make the model bigger." From here it mostly means building the room around it: the tools it can reach for, the live information it can look up, the actions it can take on your behalf. And the part everyone skips, which is a way to check that the answer is true before it's spoken.
Think about what makes someone good at their job. A great doctor is not the one who memorised the most of the textbook. They're the one who orders the right test and reads the chart before deciding. Take away the lab and the records and you don't have a slightly worse doctor, you have someone guessing from memory, however brilliant. We have spent years hiring the brilliant part and none of the lab.
This is why we think "which model is smartest" has stopped being the interesting question. An ordinary model with the right tools and current information will beat a far more advanced one working from memory alone, at least on questions whose answer lives in a live catalog or a price that moved this morning. It doesn't think better. It just isn't guessing.
General got good. Specific is where the headroom is
There's a second half to this, and it's the part we find most exciting.
General intelligence is now something you can buy. Several excellent models, broadly comparable, priced lower every quarter. That's wonderful, and it means general reasoning is turning into an ingredient rather than an advantage, the way bandwidth or storage did. When everyone's ingredient is the same, nobody wins on the ingredient.
What's still scarce is depth. Knowing exactly how this business works: what it sells, what it has in the warehouse, what its rules are, what changed this week, what happens when you press the button. That knowledge is in no training set, because it's in no book. It's alive, and it has to be learned from the thing itself and kept fresh.
So the frontier splits in two. The labs keep making the general part better, which is extraordinary and we want them to keep at it. The other frontier is narrow and deep: turning the messy, living particulars of the world into something a model can look up and act on reliably. That one is nearly unbuilt. It's also where the next big jump in useful intelligence is hiding.
Correctness is the new scoreboard
Once you accept that, the thing you optimise for changes. Fluency is solved. Every model sounds credible now, so sounding credible stopped being evidence of anything. The open problem is being right, and being able to show it.
Which makes the most valuable sentence in the product a strange one: "I checked, and I can't confirm that." An honest refusal beats a confident guess, because a guess you can't tell apart from a fact poisons everything near it. One wrong price quoted with total conviction costs more trust than ten "let me verify that" answers ever will.
This is the discipline we think the next decade of AI runs on. Don't answer from memory when you can check. Check against the live source, not a copy of it from last month. Take the action, then confirm it happened instead of assuming. And when you can't verify, say so plainly rather than filling the silence with something plausible.
Everyone is racing to build a smarter guesser. The opportunity is to build the layer that knows.
What we're building
That layer is what poneva is. We don't train models; we're happy customers of the ones that exist. What we build is the room they work in, a tooling layer that turns a working business into something an AI can use.
Point us at a site and we learn how it works, then turn what it can do into a set of dependable tools an assistant can call: search this catalog, check whether that size is in stock, add it to a cart. All of it grounded in live information rather than a snapshot. The step we care most about is the last one. Before a tool is allowed to serve customers, it has to prove it works against the live site, and if it can't prove that, it declines instead of guessing.
The bet underneath all this is that the layer gets cheaper to extend with every business we learn, in a way that training a model doesn't. It's a bet you can only win the unglamorous way, one business at a time, rather than by training something larger.
The curve turned
We don't think AI is plateauing. The easy axis is plateauing, and the hard one just opened.
The next time AI feels like a leap, our bet is that it won't be because someone trained something larger. It'll be because the thing you're talking to finally had access to what it needed, checked before it spoke, and was simply right about your order, today. That's a less dramatic story than a bigger brain, and we think it's the one that ends up counting.
That's the layer we're building. If it's the problem you're wrestling with too, we'd like to hear from you.