Ditto on the hate, technical, but important distinction here, they support open-weight ML. They do not release training source code or data sets to actually make your own (granted you’d need millions in video cards to do it, but still). Open-source gets thrown around a lot in AI, presumably virtue signalling, but precious few walk the walk.
Never underestimate the value of getting hordes of unpaid workers to refine your product. (See also React, others)
Agreed, and the chance of it backfiring on them is indeed pleasingly high. If the compute moat for initial training gets lower (e.g. trinary/binary models) or distributed training (Hivemind etc) takes off, or both, or something new, all bets are off.
The compute moat for the initial training will never get lower. But as the foundation models get better, the need for from-scratch training will be less frequent.
I think unlike Google, there are still many pure engineers that need to contribute to open source to be motivated and are still have some power.
I feel, but I am not sure, that for Google, thing have switched more and faster to the side of Big soulless corps.
Generally speaking my experience is that even in these big soulless corps there are positive and passionate people. But quite often they do not have enough decision power to have a positive impact.
I hate Meta and never use their products. But I have to give them credit for their support of open-source ML: first pytorch, then llama.
Ditto on the hate, technical, but important distinction here, they support open-weight ML. They do not release training source code or data sets to actually make your own (granted you’d need millions in video cards to do it, but still). Open-source gets thrown around a lot in AI, presumably virtue signalling, but precious few walk the walk.
Never underestimate the value of getting hordes of unpaid workers to refine your product. (See also React, others)
I understand the distinction, but it’s still waaay better than what
OpenIAIClosedAI is doing.Also people are really good at reverse engineering. Open weights models can be fine tuned or adapted. I am trained a Llama 3 Lora not that long ago.
Agreed, and the chance of it backfiring on them is indeed pleasingly high. If the compute moat for initial training gets lower (e.g. trinary/binary models) or distributed training (Hivemind etc) takes off, or both, or something new, all bets are off.
The compute moat for the initial training will never get lower. But as the foundation models get better, the need for from-scratch training will be less frequent.
I think unlike Google, there are still many pure engineers that need to contribute to open source to be motivated and are still have some power.
I feel, but I am not sure, that for Google, thing have switched more and faster to the side of Big soulless corps.
Generally speaking my experience is that even in these big soulless corps there are positive and passionate people. But quite often they do not have enough decision power to have a positive impact.