Finding an AI-crypto project with bona fide product-market fit is rarer than finding a nun in a whore house. Grifters understand that if they can keep the thin veneer of “quality” on their shitstained vaporware long enough, they can fool enough dumb money to pile in before the truth emerges. And when you get the double whammy of AI grift mixing with crypto grift, we’ve got a multi-level marketing scheme of snake oil on our hands.

So when a project arises from the banks of these degenerate waters with a plan to solve a real-deal, we pay attention. Now, there are going to be one-off projects that create some fun products - whether that be an AI Waifu or voice cloner of Jennifer Lawrence (both real). But, today, friends, we are thinking bigger.

I don’t know about you, but I became immersed in crypto because it dared to solve BIG problems. Bitcoin = digital gold; hard money that you can’t fuck with. Ethereum = the world computer and global settlement layer. This is the future of finance, after all. Set your sights high.

Bittensor - an incentive game for artificial intelligence

*Hits blunt*

“Dude. What if we got Bitcoin miners to build AIs instead of just solving useless math problems?”

The comparisons to Bitcoin seem natural after reading a little about Bittensor:

• 21M total token supply
• Halvening schedule
• Proof of work (kinda)

The whole “closed-source” vs “open-source” development methodology goes back to the Windows-Linux debate and Eric Raymond’s famous Cathedral vs. Bazaar argument. And while Linux is widely used today amongst hobbyists, about 90% of users opt for Windows. Why? Incentives.

Open-source development, at least from the outside looking in, has a lot of benefits. It allows the maximum number of people to participate and contribute to the development process. But in this headless structure, you don’t have a unified directive. You don’t have a CEO who is incentivized to get their product into the hands of as many people as possible to maximize their bottom line. In open-source development, you are at risk of the project evolving into a chimera, splitting off into tangents at every junction in design philosophy.

And what’s the best way to align incentives? Money, of course.

Think of Bittensor as a platform for economic experimentation. The goal is simple - give motivated groups the tools needed to build useful AI models. And the best part? The underlying layer (Bittensor) doesn’t even have to solve the hard problems, they just let intrepid founders figure it out on their own.

Here’s how it works on a high level.

On Bittensor, 32 experiments are running simultaneously. They could be building a language model, or predicting price movements, or generating photo-realistic images of Tom Cruise praising Xenu - it doesn’t matter. These experiments are each run from their own slot, called a subnet (or a subnetwork, all of which make up the Bittensor Network).

To start up one of these subnets, a project fronts the required slot costs (which is quite expensive [6-7 figures] today given the demand) and designs the rules of the “game.” Their goal is to incentivize AI builders to construct useful models - the more useful the output, the more TAO emissions the subnet earns. Now, is “useful” subjective? Absolutely, and we’ll return to that in a bit.

- Subnet Owners (the ones that paid for it) design the incentive mechanisms for the “game” the Miners and Validators will play.

- Subnet Miners - these are the ML/AI builders/engineers/teams, all competing with each other to build the best AI model according to the rules of the game. These are the task-performing entities of the subnet.

- Subnet Validators - think of them like the judges, ranking the outputs of the Miner’s models each time they are challenged within the game. These validators also create the tasks for the subnet (in general).

Just like in traditional machine learning, incentive mechanisms need to be tweaked over time in order to produce the desired output from the subnet. Miners found a way to frontrun one another? Update the incentive mechanism. Validators are colluding? Update the incentive mechanism. Over and over until you have a well-oiled machine that produces great output and earns a bunch of TAO (sometimes shown as the greek letter, τ) emissions.

Yuma consensus and the flow of TAO

For Bittensor, Yuma Consensus (YC) is how each subnet can ensure that all of the Validators and Miners are acting in the subnet’s best interest. It comes in two parts.

First, Subnet Validators rank each of the Miners’ outputs on a scale from 0 to 1 (these are known as “weights”) based on the quality of response, response time, and uniqueness. Since this ranking can be subjective depending on their subject matter, Miners’ scores are averaged across all of the Validators’ responses and their TAO rewards are based on this averaged score.

But what if Validators and Miners try to collude? That’s where the other half of YC comes in. Subnet Validators submit their Miner weights (scores) for review, where these weights are compared to one another and searched for irregularities. Subnet Validators earn a reputation over time (known as a “vtrust score”) which determines their owed portion of the subnet’s TAO rewards. Found colluding? Your vtrust score plummets.

I like to think of YC as a kind of hybrid of Proof of Work (PoW) and Proof of Stake (PoS). Where Miners are “proving” their work by submitting their best response to inputs, and are rewarded for doing the work. And since Validators are required to post a stake of TAO (or be delegated to from others), they have some skin in the game that’s at risk of slashing, similar to other PoS networks. The yin and yang of YC are PoW and PoS.

And, of course - the Owner gets their cut too, equal to 18% of the total subnet rewards. This incentivizes the Owners to optimize the rules of their game to make the best AI possible. Since every subnet competes for a set number of TAO emissions, each Owner has to bring their A-game to hoover all the TAO they can get.

If you’re a visual person, maybe this graphic helps clarify things.

So, how’s it going so far?

By basically all metrics, Bittensor is on a rampage. Even in TAO terms, demand for a subnet slot has never been higher, running between $1-10m at the time of writing; showing just how valuable Bittensor blockspace is today.

Some cool shit is being worked on too. Here are a couple of our favorite subnet projects.

Taoshi - imagine the world’s most overpowered TradingView indicator

Time-Series Prediction (Subnet 8, which the cool kids call “SN8”) is developing quant models, starting off predicting the intraday price movement of Bitcoin. As you can imagine, quant trading is an intensive process, usually kept behind closed doors where a bunch of unshowered and overpaid PhDs are generating millions a day for their firm. Taoshi’s goal is to flip the script.

By contributing their models, ML engineers can take their cut of about $175k per day in emissions - pushing the envelope of what is possible with open-source price prediction. With each improvement, the bar is set higher and higher.

The goal is relatively simple, to reduce the error between predicted price and actual price over time (measured in Root Mean Square Error, RMSE, if you care to know). In theory, as the models improve, the spikes seen below will dampen, showing the models have become more accurate in their price predictions.

It’s a win-win for quant developers, because in the real world, you have real money on the line if you are wrong. But for SN8 miners, they earn TAO for being right, but don’t lose anything when they’re wrong.

Taoshi is partnered with DCG (yes, that DCG) to pump in a ton of extra vetted data for the Miners to use in their models (open interest, exchange volumes, option expiries, etc.) - building towards a common goal.

Nous Research - the big swinging dicks in the room

If you went to an open-source AI conference and mentioned that Nous was building on Bittensor, you’d get the attention of the whole room. These guys are as legit as they come. Their Nous-Hermes-13b is one of the most used open-source fine-tuned language models in existence today.

Nous has been working on LLMs for quite some time, but turned to Bittensor when they noticed an issue in how open-source LLMs were being optimized. In the non-crypto world, open-source LLMs had been judged against static benchmarks with public datasets. This left them easy to game, leading to models that just “studied for the test” instead of truly innovating.

Nous took a different approach, using randomized training data freshly piped in from SN18 (Cortex.t, another cool one), miners’ models on Nous are challenged against an unpredictable examination. SN18 currently leverages GPT-4 to generate a constant stream of training data, but this could be swapped out whenever a superior model emerges. This brings open-source models up to the highest standards, all without the censorship that comes along with centralized options.

Because models on Nous are open-source, the frontrunner can be copied by a trailer, incentivizing the top dogs to continuously improve or be left splitting the pot. And the pot is LARGE. Because Nous wants to incentivize high-performers, distributions are weighted heavily toward the winners, with frontrunners taking home 50-95% of emitted TAO.

Leaders in these categories are taking home $50-100K in TAO per day at the current rate. Money talks - and this could be much more interesting for budding ML engineers (or entire companies) than begging for that sweet VC money.

TAO tokenomics breakdown

As with all projects, there is a push and pull of the supply-demand curve with TAO. Understanding how this delicate dance is orchestrated can give you a hint into where TAO may end up.

Supply side (distribution, inflation)

A lot of Bittensor’s identity can be traced back to Bitcoin (even its name, obviously). The maximum supply of TAO is 21,000,000 and the token follows an inflation schedule with a built-in halvening cycle (sound familiar?).

New blocks are mined every 12 seconds, freeing up 1 TAO to be distributed amongst network participants (as explained in the YC section). This works out to be 7200 TAO per day, which is set to halve to 3600 TAO/day at the first halvening around October 2025.

One catch - remember how we mentioned that new subnet owners have to pay a registration fee? Where does that fee go? It actually goes back to the bucket of TAO set to be distributed for the current cycle. This pushes back the halvening date.

For example: Say, over the course of a 4 year cycle, that 100k TAO is paid in registration fees. This would push back the date of the first halvening by 13.89 (100,000/7,200) days. Not substantial, but you could imagine a theoretical future after a few halvenings where registration fees start pushing back subsequent halvening dates by meaningful timeframes.

And how are these emissions helping the network?

With OpenAI bringing in roughly $2b per year, many in the TradFi world argue that open-source competitors simply can’t outspend the concentrated capital that centralized players can muster. But with Bittensor spitting off a constant stream of TAO rewards to contributors to the network, this argument may soon hold less water than your uncle’s assless chaps.

At current prices, Bittensor is emitting $1.3b in TAO rewards to market participants. And this is just the beginning of a bull market (we hope); imagine if TAO price increases and AI/ML engineers see just how lucrative it could be to contribute to the network. Of course, this is reflexive and works the other way around too - can’t have your cake and eat it too. Still, getting paid peanuts as a freelance engineer is better than nothing (I would know).

Demand side (how the token plays into the ecosystem)

Being the native currency of its network, all transactions and gas fees are denominated in TAO. Think of TAO’s utility as a combination of BTC (mining incentive) and ETH (gas) and you’ll be 90% of the way there.

On top of Owner’s paying registration fees, Validators must also post a stake of TAO in order to join the network and start judging miners. Holders of TAO can elect to stake with specific validators in order to earn inflationary rewards. The delegate (validator that received delegations) can validate subnet(s) using the total staked. As the TAO rewards roll in, delegates take their cut (default is 18%) for managing the hardware, and pass the remainder back to individuals that staked with them.

As we are in the early days of Bittensor, inflation is decently high (roughly 19%), so holders are incentivized to stake in order to stave off debasement.

Opening the speculation floodgates in 2024

Today’s Bittensor ecosystem is quite limited. Sure, there are subnets doing some next-level giga-brain shit, but there isn’t much to do for your run-of-the-mill degen besides staking your TAO. This is all about to change.

Dynamic TAO, detailed in the BIT1 governance proposal, seeks to dial up the decentralization of stakeholders as well as the competition amongst subnet Owners and Validators.

Today, the largest 64 validators by stake are in charge of assigning weights to each subnet, determining how much TAO they receive. It’s a shaky mechanism at best, since the top validators aren’t taking any risk if they slack off. With Dynamic TAO, holders call the shots, not validators. Let me explain.

If passed, BIT1 would initiate the creation of unique subnet tokens (known in aggregate as dTAO, for “dynamic TAO”). These dTAO would each be paired with TAO in a Uniswap v2-style liquidity pool (TAO-dTAO), one for each subnet. TAO holders can swap into and out of these 32 pools, swapping in if they think a specific dTAO is undervalued, and out if they think it is overvalued when compared to other subnets.

Instead of the validators assigning weights for subnet emissions (SN4 gets 3%, SN19 gets 5%, etc.), now, subnet emissions are determined by the price of their dTAO tokens. This means that if more people are excited about a certain subnet and rush into their pool, this subnet goes on to earn a higher steady drip of TAO. The market is in control, and not a small cabal of validators.

On top of this, Bittensor is heading towards opening the aperture from 32 subnets to a goal of 1024 in 2024. This lower barrier to entry for new ML teams means more competition and a more urgent need for subnet owners to promote their subnet’s work. More attention, more competition, and more benefits of being early to cool projects.

How we’re playing it

For now, it pays to keep tabs on the teams building on Bittensor. We mentioned a couple above that are exciting (Taoshi and Nous), but there a few other big names to watch:

Wombo - popular AI art generator that just spun up their own subnet (SN30). Bullish on projects with sleek front ends and checks Grant’s “mobile-first” box.

MyShell - who just raised a $11m Series A from some serious names, launched their own subnet (SN3) which aims to raise the bar of open-source text-to-speech. Pretty cool.

Sturdy - who also just launched a subnet (currently in testnet at the time of writing) aimed at providing single-sided yield for digital assets. The subnet’s objective is to allocate these assets throughout DeFi via a risk-adjusted optimization. What’s cool about this is that the subnet could be used by any other DeFi project looking for yield optimization.

New subnets launching on mainnet/testnet will introduce themselves in the Bittensor discord, which is a great place to hang out and mainline fresh alpha from the source. Say hi when you drop by.

There aren’t a ton of active VCs in the Bittensor space yet, but one I’m watching is OSS Capital. They recently brought on the DePIN Daddy himself to support burgeoning Bittensor talent. Keep notifications on from them and Tommy from Delphi (think he’s TAO-pilled now as well).

Also keep an eye on the shift in power for currently active subnets. Taostats/Taohub provide a snapshot of how much TAO emissions each of the projects is currently receiving. Here’s what it looks like today.

Once dynamic TAO goes live (sometime this year), expect us to be sharing some of our favorite projects - we will grow our TAO stacks together, friends.

The bear case

As a wise man once said, “if you can’t FUD your bags, you don’t deserve to hold them.” No project is perfect, and as bullish as I am on Bittensor, it’s important to remember the following.

1. Bittensor today has a lot of known problems. The most obvious three to me are weight-copying, a high barrier for subnet entry, and a focus on developer docs over user docs.
   1. Weight-copying: there is no outright punishment for validators to copy the weights of other validators. Validators could theoretically ignore the entire ecosystem and be similarly rewarded compared to validators keeping a close eye.
   2. High barrier for subnet entry: for many aspiring subnet Owners, shelling out $1m is too high of a bar. Luckily (a) and (b) could be alleviated by BIT1.
   3. Focus on developers over users: maybe this is just a “me” thing, but I’d like to see the documentation be segmented into “developer docs” and “user docs.” Current documentation could use some work for the everyday person, and I would like to see OpenTensor actually sponsor some 101-style publications to clarify their high-level objectives.
2. Dynamic TAO, in its current form, could be short-sighted. It may turn out that this plan only rewards subnets for real-time usefulness when it ought to have a long-term vision. If a subnet needs a couple months before their miners produce useful output, they could die on the vine (This is actually a common critique and will likely be addressed in the revised version of BIT1).
3. Open-source AI is still an experiment. Literally the largest companies in the world are all jockeying to build ML models. They are pouring billions (trillions?) of dollars into compute, manpower, and lobbying in a race that leads all of us into a dystopian hellscape. Perhaps it’s naïve to assume that a decentralized project like Bittensor has a chance competing against these players.
4. Friction involved with Bittensor may leave room for competition. High subnet registration costs, clunky mechanics, and a rail-less environment could make it difficult for Bittensor to gain mainstream adoption from ML researchers. Other projects have emerged, attempting to swoop in with lower costs and eat up market share (kind of reminds me of the blocksize wars, and we all know what happened to Bitcoin Cash, but I digress). ML researchers are still inclined to work for centralized entities, be it large government labs, academic institutions, or for AI companies that will pay them gargantuan sums for their expertise. Why would they piddle away on some experimental blockchain?

In the end, I like TAO because Bittensor offers an incentive-driven model to drive innovation forward, incrementally. Is there a lot of hype around it? Absolutely. But unlike 99% of other AI projects in crypto, Bittensor’s mechanism might actually accelerate the capabilities of open-source ML models over and above the likes of our soon-to-be technocracy overlords. And that’s something to celebrate.

DISCLAIMER: In case it wasn’t obvious, the author of this report, 563, is very bullish $TAO and thus holds a lot of it. Take this into consideration while reviewing this report. Nothing in this report is financial (or any other kind of) advice. Do your own thorough research before considering any investment.