Build Mode

Andrew Dai on the Jagged Frontier That Left Vision Behind

Andrew Dai· Co-founder and CEO of Elorian at Elorian
·~35 min·English·TechCrunch
MultimodalAI CompanyBusiness Strategy
TL;DR

Andrew Dai argues AI's progress is a jagged frontier — really strong at math, coding and Go, still poor at counting the glasses on a table — and he raised $55 million to post-train that gap.

01Core Mental Model

The Frontier Is Jagged, Not Even

AI progress is uneven — really strong at math, new physics ideas, coding and Go, and still beaten by a child at reading a photograph.

I think what a lot of people don't realize is that the way AI is progressing is a jagged frontier.

— Andrew Dai, Build Mode
Key Insight
The jagged frontier isn't only an observation, it's a market map. If every capability climbed together, a new lab would have nothing to do but race the incumbents along the axis they already lead. Unevenness means the notches are addressable — and a notch that every frontier model shares is a business rather than a benchmark. That reading is what makes leaving a frontier lab a rational move rather than a romantic one.
02The Bar

Off By One and You're Mad at Your Fridge

Vision models handle the gist of a photo well, but the uses Dai describes turn on exact answers — the count, the connection, the row — and being off by one or two is enough to make you mad at your fridge.

But imagine if you if these models really understood exactly what you had in the fridge, especially the count, right? If you're off by one or two, you would be really mad at your fridge.

— Andrew Dai, Build Mode
Key Insight
Notice which failures he reaches for: counting glasses, tracing a wire to its connector, locating a row. Each has a single correct answer, which is exactly what makes them both hard and commercially legible — you can price a model that is right, and nobody buys one that is approximately right about their inventory. The gist-shaped tasks vision models already do well, like naming a flower, are also the ones where being wrong costs the user nothing.
03The Benchmark

Thirty Pixels Can't Hold a Fridge

The multimodal benchmark that shows visual reasoning improving feeds models 32-by-32 pixel images, a canvas too small to hold the scenes any real use case cares about.

But, if you look at the details, if you really uh try to understand what's going on, that benchmark uh the inputs are just 32 by 32 pixel images. Although, the newest one is uh like 64 by 64 pixel images.

— Andrew Dai, Build Mode
Key Insight
The pixel count is the load-bearing part of his pitch. A benchmark that improves is evidence the field is handling vision; a benchmark whose inputs are too small to hold a scene is evidence the field hasn't measured the thing yet. Same rising curve, opposite conclusions — and which one you believe decides whether his company is early or unnecessary. It also explains the investor sort he describes, where the technically fluent got it straight away and everyone else needed the detail spelled out.
04The Raise

Raise the Number the Math Produces

Post-training is cheaper than pretraining but far more than a few million dollars, and Dai says their compute and data math put the amount they needed at the 50 million dollar mark.

Our goal was to raise the amount of money we believe that we needed to get things done and that was the 50 million dollar mark from our math. That would give us enough compute and data budget to build a world-class model.

— Andrew Dai, Build Mode
Key Insight
The raise is itself a technical claim. Post-training a visual reasoning model has a floor — he says it cannot be done on a few million — and a ceiling, since it doesn't need the billions a pretraining lab spends on compute infrastructure. That band is what makes a seed this size defensible rather than greedy, and it is fundable only if you accept his premise: that this notch can be closed by post-training at all. If that assumption is wrong, the financing logic goes with it. Compensation competes for the same budget — researchers, as he puts it, are very well compensated these days.
05The Cap Table

They Turned Down the Higher Valuation

Offered higher valuations with more capital, Dai took less to get investors who understand that an AI company may be revenue free for a while and won't flinch at how he spends.

So, we did have offers for higher valuations actually. Higher valuations with more capital, but for us it was really important to have the the right people on our cap table, the right investors who can who understand this journey of an AI company

— Andrew Dai, Build Mode
Key Insight
The trade only makes sense if you think the scarce resource is freedom, not money. More capital at a higher valuation buys runway; the wrong investor costs you the freedom to spend it — on GPUs the moment they're available, on an offer above the band. He optimized for the input he can't renegotiate later. The reverse pitch he says caught him unprepared, where the investor has to pitch back once they're interested, is the same instinct arriving as a process: the moment he had a choice of investors, choosing became the job.
06The Clock

The Clock Belongs to the Market, Not the Cap Table

The pressure Dai answers to comes from the market rather than his investors — new models land every two weeks at some companies, and months spent fundraising are months the frontier labs move without you.

definitely there's a ton of pressure uh on all sides, but I think I would say the majority of that pressure isn't from investors, it's from the market, the AI industry, right?

— Andrew Dai, Build Mode
Key Insight
The pressure he names is his own thesis pointed back at him. A jagged frontier moves in ways you can't observe from outside a lab, so every week spent choosing a name, courting investors and negotiating a valuation is exposure — the labs may be advancing in directions he can't see. That's why he compressed the fundraise instead of optimizing it, leaving in November and founding in December, and it reframes the valuation surprise as a byproduct of speed rather than the goal. His one regret fits the same pattern: he started the raise around Thanksgiving, when the people who do diligence are on holiday.
07The Pitch

Don't Pitch the KV Cache

Pitching MQA, MHA or KV caches to investors fails twice over — they may not follow, and the method may be irrelevant in months — so Dai frames the big picture and lets the technique sit underneath it.

Like if you're talking to an investor about like MQA, MHA, uh or like KV caches, then um there's quite a good chance they won't know what you're talking about.

— Andrew Dai, Build Mode
Key Insight
The advice is really about shelf life. Methods are the most legible thing a researcher owns and the least durable thing a company can be built on — he names MQA, MHA and KV caches, the vocabulary of a field that rewrites its own internals fast enough that he expects the terms to age out of relevance in months. What survives a rewrite is the picture, and every moat he lists (data, customer acquisition, verticalization) is downstream of a picture rather than a technique. Worth noticing what he admits: the picture wasn't fully worked out beforehand, and some of it got built during the pitches themselves.
08The Advice

The Recipe Came From Experiments, Not a Grand Vision

The 2015 paper Dai is asked about wasn't the execution of a grand vision but the residue of many experiments, which is why his advice is to tinker first and raise only once the results surprise you.

the original paper that you mentioned from 2015, that paper wasn't designed. It wasn't uh the case that we had this grand vision and then we implemented it and then it worked.

— Andrew Dai, Build Mode
Key Insight
This is the quietly subversive note, given the company he just raised for. The 2015 result came from trying many things with no grand plan; the company was sized, pitched and funded on a very specific one. He resolves the tension by putting the plan downstream of the experiment — fund the scaling once the results are unexpected, not the idea before them. It's also the only advice in the interview that costs nothing to take: the experiments he describes run on open-source models on a laptop, which is precisely the part of his own story that didn't require 55 million dollars.