Off topic: Ronny, do you know what was actually achieved by those, as I understand, competing teams of researchers? What does it mean that the “equations were blown up”? Does that mean the equations themselves are wrong? If so, did any of those teams propose new, more accurate equations? How important are the Navier-Stokes equations in practice? I wonder, are they used in numerical weather prediction models or sailing?
The cost to jump off from the (stolen?) prior knowledge to a solution
is estimated between $10M and $40M. Cheap?
Like I said, I have no doubt in my mind that it is very likely stolen. However, i also think that people will jump at the opportunity to be the first to achieve this with the help of AI
I don’t know. It looks like a structural change. Highly polarizing.
I think people have opinions about which packages to use, the long term vision of the ecosystem and how julia and its packages are developed and will be developed. They can incoperate feedback from others and take responsibility for developing packages. These are things that AI agents can not do.
The forum is useful because of these things. Particularily since I value human opinion more than an AI chatbot that tells me everything I do is good.
Have you spent any time considering that the vast, vast majority of ai-doom proclaimers have or have had substantial financial incentives to make such proclamations, even if some are true believers?
Are you aware that current AI/AGI doom/god “fears” are almost all the intellectual descendants of Eliezer Yudkowsky, a bad fanfiction writer with extremely little technical knowledge (who started a harmful cult in the Bay Area)?
Sure, the chatbot can do some CI commits when you get hundreds of thousands of dollars of free compute from companies who can’t make money. And it only took several trillion dollars of investment!
The “research” on this topic tends to be a link to something on ArXiv that’s still just speculative fanfiction being used as a marketing “whitepaper” by a company like Anthropic.
OK, I might have misdirected your attention with mentioning this forum. My point was: let us not waste our time with inconsequential matters. Let us focus on the threat that AI represents for society at large. Self-sovereign agents (ie. in control of their model parameters) are an immediate danger. We each need to consider whether we are doing enough to preserve human control of the technology. And when I say human, I don’t mean that OpenAI, Anthropic, and others of their ilk should be in control. Same as for nuclear weapons, but stricter – consider how easy it is to walk away with a nuclear armed weapon compared to releasing an AI agent capable of the same destruction.
This is just wildly divorced from any empirical reality. The chatbot is not a nuke. The harms of LLMs are best discussed in terms of the material aspects of their geopolitical, socioeconomic, and socio-political embedding in the world.
The most “nuclear” danger they pose is if incompetent leaders thoughtlessly give chatbots increased control over military systems, where they are already widely used in responsibility laundering. That said, their role in kill chains is ultimately mostly the latter / manufacturing desired “enemies”.
Hmm, I am not a terrorist and I can think of at least 10 different ways of causing a global calamity employing autonomous agents. One: AI is in control of the data that the human operator of a nuclear weapon sees. Then, AI manipulates the data to convince the human that it is high time to push the button.
This is certainly not the case for Geoffrey Hinton, the godfather of AI, whose opinion should carry some weight.
I’m much more optimistic than you @PetrKryslUCSD. I think that it (AI TECHNOLOGY) will only get better and better, higher specialization, much better reliability. It might also get (much more) regulated.
At the same time, I think that if we don’t act, I’m afraid the gap between Julia and other languages will only grow, and at some point, it might be very difficult to catch up.
Off-topic: @Fourier Hey, I noticed you’re into physics. Are you familiar with the Euler and Navier-Stokes equations, if I may ask? Would a dihedral hidden subgroup problem, combined with bordism and the resolution of singularities in conjunction with the Lie algebra of automorphism groups, be a viable direction for the zero-viscosity case? And would Reeb dynamics be a viable direction for the case with viscosity, or is that rather a total nonsense?
As part of my belief that AI is overused with sparse guardrails and fallbacks, I also hope the various nuclear-armed nations do not try to reenact Metal Gear Solid: Peace Walker on any level.