LLMs: Revolutionizing Text Generation and Understanding, But At What Cost? Bittensor's Decentralized Solution
Andrej Karpathy’s insights reveal something surprising: Bittensor's future role in revolutionizing model training, development and security
This week has been a rollercoaster and I haven't got nearly as much done as I wanted. Some days I just blissfully imagine my days with no responsibilities and commitments and how every day can be approached in the way I want.
✅️ Morning workout
✅️ Protein smoothie
✅️ Omelet breakfast
✅️ Artisan coffee
✅️ 3 hours deep work
All before 11am.
And then I'm woken up at 5am by my 2-year-old shouting
Dada breakbast! Dada breakbast! from his room.
Every second longer I wait the more frantic he gets.
So on with the day and I do as much as I can, even if its hard like this week. But my commitment is to send this newsletter every week and here I am.
As I learn the basics of AI in some more depth, focusing on Machine Learning, Supervised Learning and Classifying, Unsupervised Learning and Clustering, and Ensemble Modeling this week, one of the most interesting things I've encountered lately has been Andrej Karpathy's materials as they are so accessible to a general audience.
You know the saying, if you can't explain something simply to a 5-year-old you don't understand it enough?
He would turn any child into an AI Engineer if you locked them in a room with his YouTube videos.
His talks on LLMs struck me as having parallels to the decentralized vision of Bittensor as he explain them so lets explore this.
Model Training as Bittensor’s Miners
Andrej describes model training as compressing large datasets into model parameters and requires significant computational resources, costing the Open AIs and Anthropics of this world millions. They are basically trying to compress the internet into a massive zip file.
Bittensor miners compress their unique data into tensors, discussed in Issue 2, which validators compress into network consensus via weighting.
Even if I can't train a miner yet, I now understand that training does not need to be centralized, they all contribute their compressed knowledge and the network takes care of the rest.
Just look at Subnet 42: Masa. Millions in training costs saved!
Model Inference as Validator Scoring
LLMs hallucinating text is something every AI user has come across at some point, which at least serves as I reminder that you are dealing with a model and not a real human.
Bittensor validators score miners' tensors like humans score LLM outputs. If your miner's output isn't aligned with the subnet's goals it gets rejected. It doesn't care if I tried hard. It only cares if the tensor shape matches.
Kind of like when my kids throw a tantrum because I didn't put their fruit in their favorite purple and green bowls.
Absolute rejection.
Stages of Model Development as Bittensor’s Consensus Phases
Andrej is able to eloquently break down model development into pre-training, fine-tuning and reinforcement learning. Bittensor's subnets do the same but in a different way.
LLMs learn a vast range of internet knowledge; however Bittensor miners train on niche data and there are many subnets that specialize in such ways.
LLMs also use human-labeled data for fine-tuning; but Bittensor validators fine-tune the network's intelligence via weights.
LLMs use comparisons to improve (like playing yourself in chess) whereas subnets self-improve through validator feedback loops.
It all leads to a conclusion that Bittensor's decentralized direction is just more efficient, as opposed to centralized players brute forcing their models' development through pumping millions into GPUs (just check out Nvidia's share price) to eventually price out the competition.
Security Challenges = Bittensor’s Trust Problem
Security challenges are aplenty when it comes to LLMs and they are evolving all the time.
Imagine a jailbreak attack using a Universal Transferable Suffix appended to the prompt (yes it exists), giving you master key access to use the model for any nefarious purpose.
Bittensor has a parallel here in that the consensus mechanism is the firewall, if a jailbreak attack was attempted somewhere on the network, or anything else like a prompt injection attack hiding a trigger in a miner's tensor, validators have power to simply reject it.
The Future
LLMs are powerful tools for text generation and understanding, but they come with significant challenges in costs, training and security.
Bittensor’s subnets don’t need a single company’s billion dollar training for an LLM. They just need a thousand miners like me—messy, independent, and incentivized to iterate.
Thank you for reading my post and making it this far, even on the messy weeks where its hard to move forward!
If you have any time management tips for learning python and machine learning basics, please do reach out and in any case subscribe for future updates.
Cheers,
Brian