I saw some captures on Twitter, it looks insanely impressive! This and the recent paper by OpenAI on SAEs makes me think they aren't as far ahead as they were on GPT-4's release. It's still pretty crazy that a lot of the truly impressive modern neural architecture is usually pioneered by them first (probably because they're the only ones actually going back and trying whatever works)
It also gets me thinking about what the ex-employee Leopold talked about. While I'm of the camp that the there should be minimal restrictions on AI to foster the growth of the space, it's pretty clear that any function can be learned sufficiently with a large enough network, the right neural architecuture, and a lot of data and time. I don't think scaling up a neural network will ever lead to reasoning, but do you really need to reason to connect all of the minor associations extracted from every transcripted presidential address and piece together a reasonable approximation of the nuclear launch codes? I've always had this sci-fi idea floating around in my head of a device that you place in a room and remove it after N hours. You can then plug it into some machine and get a reconstructed approximation of what that room looked like up to N hours before you placed it, by using some form of learned model of how the light bounces around the room and some other assumptions. It wouldn't be perfect, but imagine getting even a 90% accurate reconstruction of everyone and everything that happened in a room without ever having been there while it was happening. It sounds ludicrous, but we are steadily approaching the age of being able to approximate nearly anything we can dream of. When put this way, the potential use cases are terrifying. Again, I'm of the opinion that our society will generally adapt to meet the increased potential of individuals - just like how inter-state communication and cross border policing had to rapidly evolve to tackle the 70s/80s drug and serial killing wave. But the thought that in less than 20 years from now, a stranger could just drive by your home, park for a couple of minutes, drive off and have a representation of the interior layout of your house, complete with reconstructions of exactly where everyone was and what they doing via radio/audio data, a laser and an NN, is insane - yet that is the kind of tech that will exist. Then you need to imagine what this looks like when the CIA and other foreign intelligence actors lean into these tools 100%. To have any semblance of privacy, people would need to have a constant persona, a false memetic fabrication meant to trick any would-be NNs that feed off of the data they emit. Worst part is that this is unstoppable at this point, every country is pushing to have better and better versions of this technology, it is inevitable. I am looking forward to be able to generate a proper Men In Black 4, though. On Fri, Jun 7, 2024, 12:37 AM <immortal.discover...@gmail.com> wrote: > https://kling.kuaishou.com/ > <https://kling.kuaishou.com/#expression-body-view> > > Do you see the demo if you are in China? I see nothing yet other than > people saying this ... > *Artificial General Intelligence List <https://agi.topicbox.com/latest>* > / AGI / see discussions <https://agi.topicbox.com/groups/agi> + > participants <https://agi.topicbox.com/groups/agi/members> + > delivery options <https://agi.topicbox.com/groups/agi/subscription> > Permalink > <https://agi.topicbox.com/groups/agi/Ta32348ecc8396e1a-M32d0910dc350650819325320> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta32348ecc8396e1a-Mf860d7385b782efdd194f5a7 Delivery options: https://agi.topicbox.com/groups/agi/subscription
