3 ms·
The investors are buying into a toolset for what will potentially become open source AGI that is already blowing the competition out of the water in terms of ca
by napier 4y ago
The investors are buying into a toolset for what will potentially become open source AGI that is already blowing the competition out of the water in terms of capabilities, speed of development, and access/reach. Add together large language models with text and code interpretation/generation + diffusion + vision + speech and sound + embodiment = what might you get?
It's unlikely investors care more than an inkling about the cashflow positivity of a (really fun to use, admittedly awesome) stock image synthesis engine other than its demonstrated impact as a stepping stone and buzz generator. I suspect their pitch was a little bit broader in scope than making Getty redundant.
VC exit is over a 10+ year horizon -- people writing cheques are looking to own a piece of the future, as is the case with all catchy narratives, but in this case the future is already here and publicly disclosed progress in the field appears to be accelerating rapidly in no small part due to Stability AI and the research communities they support.
- mark_l_watson 4y agoLeading to AGI? I have been using deep learning professionally for about seven years (and neural networks since the 1980s with a commercial product and I wrote the model for SAIC’s bomb detector they did for the FAA). Personally, as awesomely useful as deep learning is, I don’t see the path to AGI. I think AGI will be a hybrid of DL, GOFAI, and things not yet invented.
- ShamelessC 4y agoSure but the comment is still an appealing pitch to investors, and probably can be read broadly to include "things not yet invented". Out of curiosity - what is GOFAI?
- napier 4y agoGood Old Fashioned (symbolic) Artificial Intelligence. And thanks, absolutely my clumsily worded investor pitch intent was to cover techniques not yet invented, thanks for picking up on that.
- napier 4y agoCouldn't agree more - generalised task capable systems will likely be composable architectures that draw upon a grounding in variety of methods, not just DL by any means. Synthetic sensoria capable of integrating, interpreting and generating general task capabilities are likely to lead us somewhere interesting. DL / LLMs are akin to Wintermutes before waking; with whatever passes for consciousness instantiating and evaporating with each prompt input (which some might argue is a good thing for foom-resistant biospheres) and no continuum vector state pseudoawareness. However, strongly feel we'll have instruct-GPT models on the edge as an interface modality, but they're not the holy grail or be all end all.