I Bet Everything on Eight Weeks

Eight weeks ago I was on the wrong side of a question.

The question was: Do I actually believe what I’ve been saying out loud for years, that I can build a better embedding engine than the teams shipping the ones the world uses, or am I a person who says that and works somewhere safe?

I’d been working in enterprise software for decades, as an architect. The kind of role that pays well and lasts a long time if you don’t make any sudden moves. I had a stable salary, a family relying on it, a life that fit. And I had this idea (a specific, technical idea about how text embeddings should work) that I could not stop turning over in my head.

You can carry an idea like that quietly for a long time. Most people do. I did, for longer than I’d like to admit.

What changed is hard to collapse into one sentence. Part of it was watching the field: teams at the largest labs in the world were scaling the same dominant approach larger and larger, adding parameters and compute and data, because that is what their incentives reward. Look at the English MTEB leaderboard. It is almost entirely populated by institutional research groups with unlimited budgets. I felt, with a certainty I couldn’t argue myself out of, that they were grinding past the place where the actual leverage was. Part of it was the cold math of age. You only get so many five-year windows where you can do something this physically and mentally demanding. I could feel one of mine starting to close.

So I decided. Not dramatically. Just on a Tuesday.

I cut everything I could cut: subscriptions, comforts, the small things you don’t notice adding up. I sold what I could sell. I took the money and bought hardware. RTX PRO 6000s. DGX Sparks. Not metaphor hardware. Actual silicon, sitting in my house, drawing power on my electric bill.

I want to be clear about what that looked like from my family’s side, because if I skip past it the rest of this story isn’t honest. The conversation I had with the people who love me wasn’t “I have a great opportunity.” It was closer to: I am about to be much harder to live with for a while. The thing I’m chasing might not work. I’m asking you to stay with me anyway.

That is a real thing to ask of someone.

They said yes. I don’t take that lightly, and I never will.

Building something extraordinary is not free. You can be the best in the world at balancing the load, and a two-hundred-pound pack still changes how you walk. I knew that going in. I picked it up anyway, with my eyes open.

Then I started building.

I’m not going to walk you through the architecture. We have filed utility patents covering our runtime dynamic routing and our training data-generation pipeline, and I’d rather those do their work than give away the tricks in a blog post. What I can tell you is what eight weeks felt like from the inside.

One person. One cluster. Completely alone.

No institutional compute. No team to divide the work. Just the hardware I paid for, running on my own home electrical bill.

Some days the work was clean: the next idea already queued by the time the current run finished. Other days a training run cratered in the third hour and I sat in the dark trying to figure out whether I’d lost a week or lost the whole bet. Both happened. The people who leave those nights out of their stories are leaving out the only part that matters.

I had three real advantages, and none of them were money.

The first was knowing the field cold. Decades of building enterprise systems teaches you to smell a wrong abstraction from across a room. The dominant approach in modern embeddings has several. The large-lab strategy is to scale a single model harder: more parameters, more compute, more data. It works, slowly, and it’s what the field rewards. It’s also what you’d do if every incentive in your career had pointed that way. I had no such constraint. I took a different path.

The second was being entirely alone. This sounds like a weakness. It is a weapon when speed is the only variable that counts. No meetings. No alignment. No decks. The decision and the implementation were the same physical act.

The third was an absolute constraint. I had this cluster for a fixed, uncompromising window. There was no option to spend a year on it. The clock made me ruthless about what to try and what to throw away.

The leaderboard is where that bet either works or it doesn’t.

Today the leaderboard is public, and it doesn’t lie.

Ingot-8B-R3 is #1 on MTEB(eng, v2), the standard benchmark for English embedding quality: forty-one tasks across eight categories.

One person. Eight weeks. One cluster. The teams with unlimited compute are still there. They just aren’t at the top right now. Go look.

I want to be careful about what that means and what it doesn’t.

What it means: the bet worked. The specific technical idea I quit my job to chase produced a result that beats the giants on the hardest available objective test of embedding quality. That part is settled. I can stop wondering whether I’m a person who says he can do it, or a person who actually does.

What it doesn’t mean: a benchmark is not a product. Topping MTEB proves I can build embedding systems at the frontier. It does not solve real-world retrieval.

The retrieval problems that actually break in production (large corpora with complex structure, tables, raw code, abstract syntax trees) are things the benchmark does not fully measure. I know that better than the people who are about to argue with me about it on Hacker News.

What I’m actually building is Forge.

The benchmark is the proof. Forge is the point.

You can try it right now, no signup, at voxell.ai/forge. Paste your own text. Upload your tables. Ask it anything. If you want to build on it, logging in starts you with ten million free tokens. I’d rather you find out what Forge does with your own data than read me describe it.

I don’t know what happens next.

A leaderboard #1 doesn’t pay rent. It doesn’t fund the next cluster. I’m pre-revenue, and what you’re reading is, in part, the founder-bet phase of every honest startup story: someone went all in, the bet produced something you can verify, and now they’re finding out whether anyone cares enough to use it.

But here’s what I didn’t know going in, and what I’d want you to take from this if you take nothing else.

The hardest thing about a bet like this isn’t the money. It isn’t the GPUs. It isn’t eight weeks of staring at training curves at three in the morning.

The hardest thing is much earlier. It’s the moment when you have to decide whether you believe yourself: whether the thing you’ve been turning over in your head for years is real, or a story you tell yourself to feel like more than you are.

I decided. The leaderboard is the answer.

If you’ve been carrying something like that, the question doesn’t go away.

It just gets older.

Jonathan Corners · Founder, Voxell Inc.
voxell.ai · voxell.ai/forge

Verify what’s in this post:

• MTEB Leaderboard: look for Ingot-8B-R3
• Ingot-8B-R3 Model Card on Hugging Face
• The Voxell Technical Case for Forge