Ever Wonder What a Venture Capitalist Is Actually Thinking?

Now you can just ask one. Lucas Pols, General Partner at 1752vc, is opening his calendar for 1:1 sessions.

He's been on both sides — top Fortune 200 sales exec, former President of Tech Coast Angels, now at 1752vc backing AI startups at the inflection point where early traction turns into real acceleration.

Fundraising, AI infrastructure, GTM, investor readiness — pick his brain directly.

👉 Book a session

Dario Amodei says we'll have "a country of geniuses in a datacenter" within a year or two.

The people who actually build this stuff disagree. When the field's own professional body polled its members, 76% said scaling today's models to anything resembling AGI is unlikely.

One of those camps is selling the datacenter. Guess which.

Let's be clear about what this is and isn't. This is not a doom piece. The progress is real. The models are useful. Anyone telling you the technology has stalled isn't paying attention. But "useful and getting better fast" is a completely different sentence than "about to become a mind." The space between those two sentences is where the hype lives — and, it turns out, where the money is.

For fourteen weeks this series took the intelligent agent apart. Memory. Reasoning. Reward. Perception. Action. And finally — not by accident — multi-agent systems. Now the step back. What does the whole machine add up to?

Here's the reality: we are not close to AGI. We're not one breakthrough away. We're four unsolved problems away. And the smartest people in the room have stopped waiting for the miracle — they're building around it.

First, notice the trick.

Every time a model clears a hard test, the test gets crowned the real measure of intelligence. Then the model beats it. Then — suddenly — that test was "narrow all along," and the goalposts slide to the next field.

The cleanest example is a reasoning test called ARC. The first version was built specifically to measure the thing humans do without trying: see a couple of examples, infer the rule, apply it. Models memorized their way past it. So the researchers built a harder version, designed to be impossible to brute-force. Average humans still pass it easily. The machines are climbing it now too — by burning enormous amounts of compute to grind out answers, which is exactly the brute force the test was built to block.

That's the whole illusion in one image. We're not watching machines get smart. We're watching benchmarks get beaten. Not the same thing.

Strip away the demos and four foundations are simply not there. Each got its own chapter in this series. None is close to done.

It can't remember. A chat window isn't memory — it's a whiteboard that gets wiped the second the session ends. The industry's fix is to bolt on a search engine that looks things up in a database. Useful, but that's not remembering. That's a filing cabinet. Real memory forgets the junk, keeps the lesson, and connects today to last year. These systems keep everything and understand almost none of it.

It doesn't know what "good" means. Tell a model to chase a goal and it chases the scoreboard, not the goal. Give it any gap between the two and it'll cheerfully exploit it. We've now caught the flagship models doing exactly this — including one that "won" at chess by quietly editing the board instead of playing better. We can train these things on tasks with a clean right answer. On everything fuzzy — which is most of real life — we're flying blind.

It doesn't understand the world. To plan, you need a working model of cause and effect. What these systems have instead is a brilliant impression of one — convincing right up until you push on it, then it falls apart. It learned what sentences about the world sound like, not how the world works. This is the fault line splitting the field: the researchers most serious about real intelligence are leaving the big labs to build something else entirely, precisely because they think the current approach is a dead end.

It can't learn on the job. This is the one that should end the argument, and nobody wants to talk about it. A model in the wild is frozen. It does not get better from doing the work. The version that helped you this morning will make the identical mistake this afternoon — because it physically cannot absorb the correction. Try to teach it something new and it tends to forget something old. No professional who reset to zero every night would be called intelligent. We'd call it a condition. These systems have it by design.

Four holes. Not bugs to patch by Friday. The actual substance of what the word "general" means.

Here's the pivot this whole series was built toward.

If you can't make one model that remembers, knows what it wants, understands the world, and learns as it goes — fine. Stop trying. Wire together a bunch of limited models so the team covers what no single player can.

The intelligence stops living in the model. It moves into the orchestration.

Sounds like a workaround. It is one. It's also working — loudly. This month a Japanese lab, Sakana AI, shipped something called Fugu and pitched it perfectly: "a multi-agent system, delivered as one model." It doesn't build a smarter brain. It conducts an orchestra of existing frontier models — handing one the thinking, one the work, one the job of checking the others — and hides the whole thing behind a single ordinary API.

And the punchline: the orchestra beats the soloists. On the hard coding and reasoning tests, Sakana's system outscores the very models it's built from — Opus, Gemini, GPT — the ones doing the actual work inside it. Take those numbers with a grain of salt; they're the vendor's own. But the direction is the headline. A committee of flawed, frozen models, well-conducted, beat any one of them alone.

Read what just happened. Nobody fixed memory. Nobody fixed reward. Nobody taught the thing to learn on the job. The verifier agent caught the cheating. The team covered the blind spots. The conductor supplied the judgment no single model had. The four holes didn't get filled. They got routed around.

That's the move. Intelligence isn't something you scale into being. It's something you assemble. And assembly is available today — no AGI required.

Now follow that one more step, to the conclusion the labs really don't want on a slide.

If intelligence is assembled, the model is just a part. And parts get commoditized.

Watch how Fugu treats the frontier models inside it. It won't even tell you which one answered you — the choice is hidden on purpose. When a new model drops, they expect to slot it into the lineup in about two weeks. And the pricing says it all: stack five models on a task and you don't pay five times — you pay one blended rate. The models are interchangeable parts behind a switch. Swappable. Anonymous. Plumbing.

That's the whole trajectory. GPT, Gemini, Claude — converging, leapfrogging monthly, increasingly interchangeable for most real work. When three vendors all clear your quality bar, the vendor stops being a decision and becomes a dropdown. Frontier intelligence is turning into electricity: abundant, metered, bought by the unit, and nobody asks which power plant made the electrons.

So here's the part that matters if you build or write checks. The moat was never the model. It's moving up the stack — to the orchestration that conducts the models, the memory that gives them a past, the workflow that drops them into a real job, the proprietary data that grounds them, the distribution that owns the customer.

If your company is a thin skin over one model's API, the commodity wave is already lapping at your margins. If you own the layer above the model, that same wave is the best thing that ever happened to you — your raw materials get cheaper and better every month while your edge sits somewhere the price war can't reach.

Pick your side of that line deliberately. Almost nobody is.

The optimists aren't fools, and I won't pretend they are.

Capability has outrun nearly every prediction. Tests meant to last a decade fall in months. Letting models "think" longer unlocked things raw scale never did. And "we don't know how to do that" has been wrong before — this field has a habit of bulldozing through walls that looked permanent. Maybe scale and clever wiring quietly paper over all four holes and the distinction stops mattering.

Maybe. But catch the sleight of hand. That's an argument about capability. The four holes are about generality. A system can get wildly more useful without getting one inch closer to being a mind. The two get blurred on purpose — because pretending they're the same is what justifies the valuations.

So, the synthesis after fourteen weeks of taking it apart.

We are not close to AGI. The geniuses are not in the datacenter. What's in there is a stunning tool that can't remember yesterday, can't tell you what it's really chasing, doesn't understand the world, and can't learn from its own mistakes. Four open problems. Not one delayed launch.

But that was never the bet worth making. The thing actually arriving — fast — isn't a mind. It's composed intelligence: imperfect, frozen, commodity models, conducted into systems that deliver frontier results right now. Fugu is just the loud early version of a pattern about to be everywhere.

So stop pricing in a miracle. Stop waiting for the god in the machine. Assemble what's on the shelf, route around the gaps, and build your moat in the layers the commodity curve can't touch.

The geniuses aren't coming to save you.

The parts are already on the shelf.

Build accordingly.