Bertrand

Bittensor is flawed. I'm fascinated by the idea of decentralised collective intelligence and was exploring the BT model to bootstrap AI networks in specific domains. But I found several things that gave me pause: 1. Incentive-driven competition WITHIN subnets to produce the best outputs can make sense IF implemented properly. There are challenges like gaming of incentives, relay mining & assumptions that validators know the correct outputs, but I believe these are solvable. The economic model starts to fall apart when competition extends horizontally and subnets must compete to get a larger allocation of inflation rewards (esp when dynamic TAO is implemented). How do $TAO holders compare the economic value to the network between: (1) a text-to-speech model (subnet 3), vs (2) a vision model that screens X-rays for disease? Value generation is entirely different. Relying on market forces to determine the value contribution of each subnet is not optimal for long-term strategic development. It’s like having every AAPL shareholder vote on what the next product Apple should build is. 2. I don’t think subnets will eventually be competitive in generalised tasks like image generation, text generation, etc. Many parties' incentives are not necessarily financial—e.g., driven by reputation, research contributions, or public service. Example: Meta releases powerful LLMs like Llama-3 open source and is available for anyone. I haven’t seen subnets do things that existing open-source projects can't, nor have I seen any data showing that they outperform existing benchmarks. 3. @opentensor is somewhat decentralised, but it is not open source. There’s a clear distinction here. While anyone can use a subnet, what actually goes on is not transparent. What are the miners doing? What algorithms are they running? How are the results generated? The user has little visibility; it's a black box. 4. The demand side for TAO is contingent on (1) Speculators (2) More subnets spinning up, and/or more demand to be miners/validators. However, more subnets/miners/validators dilute the incentives for existing subnets. A drop in token price leads to a reflexive drop in network activity as the unit economics for miners/validators fall under pressure. The network's operating cost is 7,200 TAO (daily inflation). That's ~$3.2M/day or $1.2B/year. Today, this funding largely comes from speculators. Is Bittensor producing $1.2B of value in a year? There is value, but we are probably too far ahead of ourselves. Some musings: - There will be more synergy if all subnets in a network are aligned towards a particular sector rather than on disparate tasks. For example, a new “Bittensor for genomic research/drug discovery”, or a “Bittensor for finance/trading”. Helps with economic alignment, direction and business development. - BT’s economic model today might allow it to thrive in certain niche areas of intelligence that don’t receive much general interest but where incentives could drive model advancement in those areas. Perhaps crypto-specific use cases, like predicting token prices. I believe that decentralisation in AI has an extremely important role to play. Bittensor is an interesting experiment. Still consolidating my thoughts. Would love to hear any counterpoints or informed views on these.