TheEzEzz

> The "top researchers" in AI are Chinese. And I am skeptical that they even remotely have the philosophical or political alignment you are attempting to project on to them.

I assure you that Chinese researchers have a diversity of philosophical and political alignment, much the same as other researchers. I also assure you that top researchers as a whole are not all Chinese, though the ones that are that I know are all very thoughtful.

The problem with censorship isn't that it degrades performance. The problem is that if the censorship is unilaterally dictated by a government then it becomes a tool for suppression, especially as people use AI more and more for their primary source of information.

A company might choose to avoid erotica because it clashes with their brand, or avoid certain topics because they're worried about causing harms. That is very different than centralized, unilateral control over all information sources.

I wrote my own dynamic keyboard layout to optimize typing speed while procrastinating on my dissertation.

15 years later I'm still using it. My dissertation not so much.

Procrastination is (sometimes) awesome.

I could easily see this going the other way. Life long single people develop strong social networks that they keep investing in into old age. Married (and with children especially!) couples invest less time in their social network, in old age they then have many fewer friends when their children leave and their spouse passes.

(I'm not sure this is true or not, but seems plausible. I agree with the author that we should get better data to resolve these questions).

You're basically taking the model "off policy" when you bias the decoder, which can definitely make weird things happen.

LeCun is very simply wrong in his argument here. His proof requires that all decoded tokens are conditionally independent, or at least that the chance of a wrong next token is independent. This is not the case.

Intuitively, some tokens are harder than others. There may be "crux" tokens in an output, after which the remaining tokens are substantially easier. It's also possible to recover from an incorrect token auto-regressively, by outputting tokens like "actually no..."

Super cool. When I think about accelerating teams while maintaining quality/culture, I think about the adage "if you want someone to do something, make it easy."

Maintaining great READMEs, documentation, onboarding docs, etc, is a lot of work. If Auto Wiki can make this substantially easier, then I think it could flip the calculus and make it much more common for teams to invest in these artifacts. Especially for the millions of internal, unloved repos that actually hold an org together.

Good question. Perhaps depends on the type of warning shot. Plenty of media has an anti-tech bend and will publicize warning shots if they see them -- and they do this already with near term risks, such as facial recognition.

If the warning shot is from an internal red team, then higher likelihood that it isn't reported. To address that I think we need to continue to improve the culture around safety, so that we increase the odds that a person on or close to that red team blows the whistle if we're stepping toward undisclosed disaster.

I think the bigger risk isn't that we don't hear the warning shots though. It's that we don't get the warning shots, or we get them far too late. Or, perhaps more likely, we get them but are already set on some inexorable path due to competitive pressure. And a million other "or's".

I agree -- the risks are bigger, the rewards larger, the variance much higher, and the theories much less mature.

But what's striking to me as the biggest difference is the seeming lack of ideological battles in this Moon story. There were differences of opinion on how much precaution to take, how much money to spend, how to make trade offs that may affect the safety of the astronauts, etc. But there's no mention of a vocal ideological group that stands outright opposed to those worried about risks -- or a group that stands opposed to the lunar missions entirely. They didn't politicize the issue and demonize their opponents.

Maybe what we're seeing with the AI risk discussion is just the outcome of social media. The most extreme voices are also the loudest. But we desperately need to recapture a culture of earnest discussion, collaboration, and sanity. We need every builder and every regulator thinking holistically about the risks and the rewards. And we need to think from first principles. This new journey and its outcomes will almost surely be different in unexpected ways.

A good analogy for AI risk. We'd never visited the Moon before, or any other celestial object. The risk analysis was not "we've never seen life from a foreign celestial object cause problems on Earth, therefore we aren't worried." The risk analysis was also not "let's never go to the Moon to be _extra_ safe, it's just not worth it."

The analysis was instead "with various methods we can be reasonably confident the Moon is sterile, but the risk of getting this wrong is very high, so we're going to be extra careful just in case." Pressing forward while investing in multiple layers of addressing risk.

Supernova are powerful enough that even a star in a different solar system going nova can kill you, if it's a "nearby" system. But I believe there aren't any stars close enough that would go supernova any time soon.

I watched this yesterday and got the feeling something big was happening. At one point he says "This is actually a very inconvenient time for me [to be here]." At the end of the session when they're wrapping up, he begins to stand up to leave the instant that the moderator starts wrapping up.

Anyway, I suppose we're reading tea leaves and engaging in palace intrigue. Back to building.

I'm sympathetic to your take on how overly grandiose the language is, but I also think you're being too harsh here.

The idea that the universe is discrete/computational is a fine idea, but underspecified and useless on its own. There's an infinite array of computable rules to choose from. But the fact that with a few assumptions on the rules you can then limit to both GR and QM is very non-trivial and, in my opinion, pretty surprising.

To your point, does it prove that this is _the_ correct theory? Definitely not, and metering language around the claims is important. Still, the result feels novel, surprising, and worthy of further investigation, alongside the other popular models being explored. I think it's a shame that Wolfram's demeanor turns people off from the work.