TAO is nonsense

Kyle Samani said it. TAO is BULLSHIT.

But is it? Of course there are lengthy papers "debunking the Bittensor myth" out there, but we also have the CEO of Hugging Face publicly shouting out Bittensor.

Who to trust?

Sadly, I have zero knowledge in AI, and I can create a simple query in SQL. All I can say is that BitTensor is controversial, but there are real-world applications proving the skeptics wrong. Meaning, I lean more into the TAO-friendly wing.

Now one might yell at me, "Bro you can't just put your bets like that!!"

I mean, it's pretty simple. Just look at the incentive structure: is the project using tokens to incentivize certain behavior to solve problems in AI development? And is it actually working?

Then fu** yeah let's gooooooo!!

Me when I find a <1M marketcap ZK-Layer 2-Decentralized-AI-Social Platform

What's The Problem Tho?

I found this article from a16z very helpful to understand the main pain points in AI development and what crypto has to do with any of it.

The truth is that today’s neural networks are fantastic interpolators but terrible extrapolators. They are powerful pattern matchers with the ability to contort themselves to fit almost any dataset, but their fitting is blind to the mechanisms that generate the data in the first place.
Neural networks do not yet engage, as humans do, in a creative search for explanatory theories that account for why the data is as it is. They also certainly don’t then go forth, as humans ought to, and steadily strive to falsify every last one of those creative theories until a single one emerges triumphant as the best explanation for the data observed.
A human scientist (like say, James Maxwell) makes predictions about things (like electromagnetism) by building, through an iterated process of conjecture and falsification (i.e. science), a deductive framework (Maxwell’s equations) that generalizes to future situations.
Deep neural networks, on the other hand, take a different approach to modeling reality. They stitch together thousands of linear functions, laboriously shifting each one slightly for each training example, into a kind of high dimensional quilt — a manifold — that fits the training set. In doing so, they cannot help but make predictions by inductively pattern-matching onto what they’ve seen happen before. They mirror (rather than explain, as humans do) the chaos and complexity of the phenomena they observe.
Neural networks arrive at their predictions by induction, not deduction.
This helps explain why today’s deep neural networks require so much data to learn anything useful: They are inductive interpolators and, as such, they require a large number of points between which they can do their interpolating.

To put it simply, AI models are f***ing dumb.

They lack creativity; they only answer your questions by spitting out the most probable output, only after they are fed with twenty dumpster trucks worth of data. If an AI model encounters a novel input, it will still give out the most probable output with such confidence even though it has no clue what you meant by your prompt. We call these AI hallucinations.

AI hallucinations are incorrect or misleading results that AI models generate. These errors can be caused by a variety of factors, including insufficient training data, incorrect assumptions made by the model, or biases in the data used to train the model. AI hallucinations can be a problem for AI systems that are used to make important decisions, such as medical diagnoses or financial trading.

Now imagine if you want to know a specific thing in a specific domain. Will ChatGPT know it? Maybe on a high-level, but it might have problems interpreting details and providing clear-cut answers.

Of course ChatGPT 4 is still a great achievement. But we just can't allow these AI models spewing nonsense, just to be copy-pasted during internet debates that might be referred back as "the truth" in the future somehow.

Imagine everything you knew was all just an AI making shit up

We actually need to improve models, not only for profit but to save the world from false information made out of thin air, coming from something that us, humans created!

But Houston, we have a problem: collecting data is hard.

Collecting very specific types of data requires a lot of effort, and a centralized entity like OpenAI might have problems collecting such data. For example, would they know the Korean internet slang "Jamminnie(잼민이)"? Of course not, because I asked ChatGPT about the word's meaning before writing this article, and it got everything wrong. However, when you actually speak Korean like me, you can simply google the word and find it on a specific wiki page popular in Korea. So it's not that there is not enough data to train ChatGPT, but there are not enough resources, be it manpower/money/time, to feed this type of data when training AI models. Furthermore, there might be little to no incentives to actually feed this data at all because most people don't use ChatGPT to study obscure Korean internet memes, right?

Also, there is the problem where OpenAI employees don't even know such words like "Jamminnie" even exists. Therefore, the only solution to this problem is actually having Korean memelords to provide, filter, label, and validate the data when training a model tailored for Korean internet memes. Is that viable? I don't think so. Because demand for explanations on Korean internet memes in the English speaking world is very low; no company will want to develop a model for this task. In other words, there is a lack of incentives to solve this problem.

Now substitute "Korean internet memes" to any other topic you can think of right now. There will be so much ground to cover that a single model cannot handle, and there will be so many fields that would require specific domain knowledge to gather the data to train a model. You see where I am heading to? We have encountered the long-tail problem.

A single AI model will not be able to satisfy all niche subjects and all subjects will be covered by different people, at least for now. A single entity like OpenAI alone, regardless of its size will not be able to solve this problem. Not only because of the sheer amount of data it must process, but also because it will be a daunting task to prioritize what data has the most demand and respond accordingly. Plus they have no clue what data even exists.

Most of all, these models created by tech giants are siloed; they don't share their secret sauce. So we need a system where:

Innovation can be shared
Arbitrary participants can contribute to the system as a supplier or consumer
Create an incentivized structure for demand response

To my understanding, this is what Bittensor does.

Bittensor aims to create a peer-to-peer marketplace that incentivizes the production of machine intelligence. Through cutting-edge techniques like mixture of experts and knowledge distillation, the platform establishes a collaborative network where knowledge producers can sell their work, and consumers can purchase this knowledge to enhance their own AI models. By fostering a “cognitive economy,” Bittensor seeks to facilitate knowledge sharing among researchers, encouraging the development of more powerful AI models. The ultimate goal is to open new frontiers in the advancement of AI by aggregating individual contributions of knowledge. As computers contribute their AI models and training to the Bittensor network, they are rewarded with Tao. - bittensor.org

Since I have limited knowledge on how Bittensor exactly works, I will not go into the details, but if Bittensor is what it says, and if it does turn out to be a remedy to the long tail problem and centralization problem inherent in current AI development, I think it is promising.

Nous Research, who launched a fine tuning subnet on the Bittensor network, also happens to be the creators of the most powerful open source model in the market.

Bindu Reddy

@bindureddy

Nous Hermes 2 beats Mistral-Instruct MOE and becomes the best open-source model.

Open-source AI continues to make strides on a daily basis. The latest release from the Nous team beats the best open-source model!

10-shot MMLU is better, but it would have been for them to…

Their subnet also has a leaderboard where you can see people are actually making (a lot of) money for fine tuning models. So I guess this is no joke.

https://huggingface.co

Finetuning Subnet Leaderboard - a Hugging Face Space by NousResearch

Discover amazing ML apps made by the community

Mind Share & Narrative

According to Kaito, mindshare of AI is about 30%. Considering that the ETF is a macro event rather than a sector in crypto, the mindshare of AI might even be bigger, or at least the AI narrative will benefit from future ETF approvals.

Sentiment analysis is overwhelmingly positive.

Price versus mindshare is also super high, as I am writing this OpenAI just released Sora, an AI model that can create realistic and imaginative scenes from text instructions, causing a huge spike in AI and related tokens.

Final Verdict

Both tech and narrative are very convincing. I think I will regret not riding this one. Yes, there are other small cap tokens out there, but sometimes it's just better to bet on the biggest winner.

Current market cap of TAO is $4.2B, but given that we have only started a generational bullrun fueled by boomer money, let me say it..... We're still early™

Check out more awesome articles from DeSpread Research

https://research.despread.io

DeSpread Research

DeSpread Research · Meet our people's insight, our capabilities, our research, and the ever-changing face of our firm.

Also, subscribe to my blog to get invite codes for Thruster testnet, a Blast-native DEX.

Once you subscribe the codes will be revealed below.