I Tried to Mine on Bittensor. Here's What Actually Happened.

I Tried to Mine on Bittensor. It Broke First. Here's What Actually Happened.

There's a sentence I wrote three weeks ago that's been sitting in the back of my head like an open tab I refuse to close: "I haven't mined anything yet, and I won't pretend otherwise."

I can finally close that tab. Sort of. I haven't mined anything in the sense of actually earning rewards yet — but I've got a miner account set up and registered, after a very ungraceful detour through something breaking first. That detour is the more interesting story, honestly, more interesting than a clean "and then it worked" would have been.

This is the follow-up to that first confused, honest post about walking into Bittensor with zero crypto background. If you read that one, you know the deal: no success-story energy here, no pretending I've figured this out. Just an update from someone a few weeks further in, who tried to actually do the thing this time instead of just understanding it, hit something that genuinely didn't work, and then — with a bit of help — got past it.

But before I get to the failing part, I have to talk about SN41. Because that's where this week actually got interesting, and honestly, it's the part I'd want a beginner to read even if they skip everything else.

The subnet that made "decentralized AI" stop being abstract

In camp, we got introduced to Subnet 41 — Sportstensor — and I'll be honest, when I first heard "it predicts sports outcomes," my brain filed it under "okay, fantasy-league bot, neat, next." That undersold it completely, and it took me a minute to realize why.

Here's what's actually happening: SN41 has miners running their own machine learning models to predict real sports outcomes — who wins, by how much, across leagues like the NFL and NBA — and validators checking those predictions against what actually happened on the field. We were told the actual target isn't "guess the winner for fun." It's something much sharper: beat the bookmakers. Find the statistical edges that traditional sportsbooks, with all their resources, are still missing. That part didn't fully land for me until someone in camp put it plainly: this isn't a side project for sports fans. It's trying to become a genuine source of pricing intelligence that real betting and trading platforms actually want to plug into.

That reframed something for me. My first post was all about subnets competing for attention and stake. SN41 made that competition feel less theoretical, because I could actually picture who'd want this. A trading desk wanting an edge the public odds haven't priced in yet. A betting platform wanting sharper lines than its own oddsmakers can produce alone. The subnet has already started showing up in real markets rather than staying purely theoretical, which is the part that made it feel less like a classroom example and more like something with actual stakes attached.

That's the moment "decentralized AI" stopped sounding like a slogan to me, again, for the second time in this whole journey — except this time it wasn't a definition that did it, it was seeing one subnet's output get used by something outside Bittensor entirely. It's not just "AI subnets exist next to each other." Some of them reach all the way out into markets that have nothing to do with crypto on the surface, and that's the first time the word "ecosystem" felt earned instead of decorative.

The part that genuinely surprised me

I expected the prediction itself to be the hard part of this subnet, and the scoring to be the simple, almost boring part afterward — did the prediction come true or not, yes or no. I had that backwards.

The actually clever bit is how miners get judged at all. It's not "did you call the winner correctly." It's whether your prediction beat the closing odds — the final price the market itself settled on right before the event happened. That distinction took me a second read to actually feel the weight of. Calling a winner is easy when the favorite usually wins anyway. Beating the closing line means you found something the entire betting market, with all its money and all its sharpest people staring at the same game, hadn't fully priced in yet. That's a completely different, much harder bar than "were you right."

I didn't expect a sports-prediction subnet to have that much thought behind what counts as actually useful. I think I'd quietly assumed this category was the more recreational, lower-stakes corner of the ecosystem — something closer to a fantasy league than serious infrastructure. I don't think that anymore. Measuring yourself against a market that's already trying its hardest to be efficient is a genuinely difficult bar to clear, and it's the kind of design choice that quietly filters out anyone just guessing. That feels like it might be true of most things in this ecosystem, now that I think about it: the mechanism is simple to describe and apparently very hard to actually be good at.

There's also a layer sitting on top of SN41 that I didn't expect — its predictions don't just sit inside Bittensor as an internal scoreboard. They've started feeding into actual betting and trading platforms, the kind of places professional bettors and analytics firms already use, with no need for any of those end users to understand subnets, miners, or TAO at all. That detail mattered to me more than it probably should have. It means the end goal here isn't "crypto people predicting games for other crypto people." It's reaching for an industry that's been pricing odds with its own closed methods for decades, and trying to out-predict it from the outside, in the open. The protocol is the engine room. Most people are only ever supposed to see the edge it produces.

Okay. Now the part where I tried it myself, it went sideways, and then — thankfully — it didn't stay that way.

I want to write this part exactly as it happened, because I think the polished version of this story — the one where I calmly explain mining like I've done it ten times — would be a lie, and also less useful to whoever reads this next.

I wasn't about to throw real TAO at this as a first attempt, so I went the sensible route: testnet, run entirely out of a Google Colab notebook rather than on my own machine. That part alone felt like a small relief — cloning the repo, installing Bittensor, creating the wallet, running the miner, all of it living in one notebook I could poke at from a browser tab, instead of wrestling with a local environment on top of everything else I was already learning.

The camp had shared a faucet link for getting testnet TAO, and using it taught me something I hadn't actually pictured correctly in my head: I'd half-expected requesting test TAO to be its own separate, slightly fiddly step — something I'd have to manually move into my wallet afterward. It wasn't. I generated a coldkey and a hotkey for the wallet first, used the faucet link, and the TAO just landed directly into that wallet, no extra step in between. Small thing, but it's the kind of small thing you only learn by actually doing it once, instead of reading ten setup guides that all describe the wallet and the faucet as two separate paragraphs with no mention of how directly they actually connect. I had a hotkey registered on SN41, sitting there with a miner UID assigned to it. I sat there for a second feeling, for the first time in this whole journey, like I was actually in it instead of reading about it.

And then I tried to actually start the miner. And it broke.

Not in the dramatic, screen-full-of-red way I'd half-expected going in. It was quieter and more annoying than that — everything upstream of "run the miner" worked exactly like the docs said it would, and then the part where the miner process itself was supposed to come alive just... didn't, not cleanly. I went back through the setup steps more than once, assuming I'd missed something obvious. I genuinely couldn't tell if I'd fat-fingered a command, misread an instruction, or stumbled onto something actually broken.

Turns out it was the last one, and it wasn't even something on my end conceptually — it was a version problem. I'd installed an older version of the Bittensor/subtensor library than what the current setup actually expected, and it sat there quietly causing the kind of failure that doesn't announce itself as "hey, you're on the wrong version." It just looks like things not working, for no obvious reason, while you stare at the same three lines of output wondering what you did wrong.

I didn't catch this myself. One of the mentors in camp looked at what I was dealing with and basically went "ah, that's a version thing" almost immediately — the kind of instant recognition that only comes from having seen the exact same wall before. I want to be honest about that instead of writing myself into the story as the person who cracked it solo, because I didn't. I got stuck, said so out loud, and someone who'd already been stuck there once pointed me straight at the door.

Once I updated the library to the version it was actually supposed to be running, the part that had been silently failing just... worked. Registration went through cleanly the second time around, the hotkey got set up properly, and I actually have a miner account live now — set up, registered, present on the network. I haven't confirmed it's actively earning or doing useful work yet; that's the next thing I'm watching for. But the wall that stopped me cold is gone, and I know exactly what caused it, which honestly feels like its own small win.

Here's the thing I keep coming back to, though: registration — the part I was most nervous about, the part involving wallets and keys and putting something on-chain even if it was only testnet — went fine both times. The part that broke was the unglamorous "now actually run the software" step, the part no explainer really warns you about because it's not conceptually interesting enough to write a paragraph on. Nobody makes a diagram of "and then the library version was outdated." But that's real Bittensor too. The actual experience isn't only the elegant incentive design — it's also dependency versions and config files and the occasional silent failure that conceptual explainers never have to deal with, and sometimes the fix is genuinely that simple once someone who's seen it before points it out.

I'm glad, in hindsight, that I didn't sit on this post until everything was perfectly resolved. The version I almost wrote — "I tried, it broke, I'm still stuck" — would have been just as honest at the time. This version, where it got fixed a day or two later because I asked the right person, is more honest about how most of this actually goes: you get stuck, you say so, somebody who's already been there saves you the hours of confusion, and you move forward a little wiser than before.

What I'd tell someone starting today, with both of these experiences in hand

Go look at SN41 even if you have zero interest in sports betting. Not because you need to mine it, but because it's the clearest example I've found of a subnet's purpose being obvious. A lot of subnet names and descriptions still go over my head completely. This one, I got immediately, and that made everything around it — why subnets compete, why some get more stake than others, why "useful" is the entire game — click harder than reading the mechanics alone ever did.

Don't assume registration is the hard part. I built up the wallet-and-keys step in my head as the scary boss fight, and it was honestly the smoothest part of the whole attempt, both times. Use a testnet faucet first if you're nervous, the way I did — it removes the financial stakes while you're still learning where the actual friction lives.

You don't need your own machine to start. Running the whole thing in Colab — repo, install, wallet, miner — meant I didn't have to solve "set up a proper dev environment" before I could even get to the Bittensor-specific learning. If you're intimidated by the idea of a local setup, this is a real way to lower that bar without skipping anything that actually matters.

Check your library versions before you assume you've broken something conceptually. My entire wall turned out to be an outdated Bittensor/subtensor install, not a misunderstanding of how mining works. If something fails right after registration with no obvious cause, a version mismatch is genuinely one of the first boring things worth ruling out, before you spiral into thinking you don't understand the system.

Ask the room. The fix here didn't come from me reading harder or staring at error output longer — it came from saying "this is broken and I don't know why" out loud to someone who'd already hit the exact same thing. That's not a failure of independence. That's just how infrastructure problems usually get solved, in camp or anywhere else.

Let "it broke, then someone helped, then it worked" be a fully acceptable version of the story. I said some version of "I don't get this part" five times in my last post. This time it's "I got stuck, asked, and got unstuck," and that's no less honest a thing to write about than a clean first-try success would have been — arguably more useful to whoever reads this next and hits the same wall.

Where I'm actually standing right now

I went in this time wanting to write a "first mining attempt" post and I'm coming out with one that's slightly messier and more honest than that framing — a real setup failure, a real fix, and a real account now sitting registered on the network. I generated both a coldkey and a hotkey for the wallet, registered on SN41, and came out the other side with miner UID 96 — an actual number, on an actual subnet, that's mine. I haven't watched it actually earn anything yet, and I'm not going to pretend otherwise. But SN41 gave me the clearest sense yet of what "useful work" concretely looks like inside this ecosystem — not an abstract market, but real predictions, measured against a real market's closing odds, with real platforms on the other end. And the mining detour gave me something just as valuable: proof that the gap between understanding a system and running one is real, that it's usually something boring like a version number, and that asking for help closes that gap a lot faster than stubbornly debugging alone.

Next, I'm watching to see if UID 96 actually starts pulling weight on the network — whether it's doing anything a validator would call useful. That's the next milestone, and probably the next post: hopefully a shorter, less philosophical one that's mostly just "and here's what it's actually doing now." Until then, this is where I am: a little more impressed by what SN41 is trying to do, a little more humbled by what "hands-on" actually means, genuinely grateful for a mentor who recognized a problem I couldn't, and still, gladly, very much a beginner.