I would say that vibes are a much worse proxy than overall public opinion - especially if you have partisan leanings and don't attempt to temper their effects.
Some of my impressions come from a private mailing list where conservative lawyers
Since 2016, there has been a small but very vocal contingent of neoconservatives that get trotted around as "conservative <insert profession here>", but whose priors WRT anything Trump does are closer to the leftmost quartile of Democrats than to the median Republican, or even the median Independent. A common drinking game among news-readers is to search for articles of the format "Conservative Commentator says <thing that one would very much not expect a right-wing American to say>", and take a swig if the unnamed conservative commentator turns out to be Erick Erickson, Bill Kristol, or David French. The three of them alone cover about seventy five percent of these articles, and you can pick another three names to cover 75 percent of the rest.
This is to say that it's very easy to fall into the trap of believing that your views are universal because a cherrypicked set of Fox News Liberals (of either party) are serving as your model for the 'reasonable opposition'. Hard data may not be perfect, but it's essential in emotionally fraught domains.
US courts have so far mostly resisted the growing corruption in the other two branches of government.
I'd say that that's a controversial assertion to state as an axiom. The best proxy I can think of is their public support, which, per Gallup, is slightly higher than the Executive, but by nothing near the gap between the Executive and Legislative branches. Even then, the difference is primarily due to a 600% difference in trust among Democrats specifically, with Republicans and Independents rating them similarly to the Executive.
Bubble Tanks is a Flash game originally released on Armor Games, a two-decade-old online game aggregator that somehow still exists.
Love me old Flash game mentions. It was a magical time when any twelve year old with a computer could pirate Adobe Flash and leave a cultural impact. Tens of thousands of other bored twelve year olds leaving feedback, and occasionally you could get a DM years later about how you inspired someone to make their own game or became the official e-sport of his school's computer lab for a month. I hope we someday get something like that again, though, with the modern internet, I can't imagine how.
Something specific that really speaks to me about that era is games where it was enormously clear that the developer had some kind of grand vision for the game universe. Little snippets of plot text that referenced things that had only ever been written down in the margins of the developers' geometry notes, and conversations with his friends during study hall. Engines that were heroically complex for ActionScript 3 so that they could capture a fraction of the game that the young programmer ultimately dreamed of making.
I think, now and again, about the sheer number of top engineers who got their start this way, back in the 00's. Probably accounts for billions of dollars in extra GDP.
at this point, LLMs can handle nearly every sufficiently-chunked-up bit of music production, graphic design, video editing, background illustration, character concept art, voice-acting, essay writing, and a lot more
Insofar as the current paradigm is concerned, I think this selects for "deep" rather than "weird". One of the big tells for AI-generated images is surface-level details without any higher-order planning about their functionality or significance (the interior of a watch, for example). I think this necessarily extends to your thought experiment about generating films from prompts - a paragraph-long prompt is necessarily less information-dense than a film, so its information will naturally either amount to a mass of stylistic fluff around a relatively sparse narrative core defined by the prompt, or it will impose additional information on the prompt that makes an equally significant contribution to the film, in which case the prompt was really more of a suggestion.
As for the broader point, a distribution isn't just a median - a good foundational LLM, even today, can model the weirder aspects of what it's seen during training just as well as the more normal ones, and the 'flavor'/sameness of LLMs' writing comes more from the fine-tuning process. Moreover, GRPO-style reinforcement learning allows LLMs to learn to do out-of-human-distribution things (like solving unsolved mathematical problems more effectively than a human can), so long as there's a good enough way to quantify reward. I think that, should LLM capabilities continue to improve, and should someone apply a sufficiently good engagement-maximization reward function to them, human novelty-seeking will drive them to produce content that is sufficiently weird for anyone's taste.
I think the core counterargument here is that the dictionary definition of a name isn't intrinsically linked to the group, organization, or set of methods that adopt the name.
- Suppose I support the Ruritanian Democratic Movement, which promises to turn Ruritania into a democracy.
- After getting to power, its leader decides to set himself up as king for life, and the good name of the movement comes crashing down around him.
- Supposing this hasn't broken my faith in Ruritanian democracy, should I still call myself a member of the Ruritanian Democratic movement? Definitionally, I am, but, practically, that doesn't sound like the best way to achieve Ruritanian democracy.
- It may be that the name is so tainted that continuing to keep it in the public eye distracts and demoralizes anyone who believes in the cause.
- It may also be that the methods associated with the group were likely to lead to this undesirable outcome, and a clean break from them would better achieve the goals of its true believers.
Dispensing with the metaphors, what is the semantic difference between saying "I support charitable giving" and "I am an Effective Altruist"? To my understanding, the core difference is one of methods - the charitable giver selects charities heuristically, based on what he sees an immediate need for in the world, while the effective altruist poses that there is a more efficient strategy for achieving the same ends, by maximizing each dollar's marginal impact on some objective function. Identifying why EA has seen the problems it has, and how this relates to the difference in methods, would, as best I can tell, be critical to determining whether "I am an Effective Altruist" is more preferable to say than "I support charitable giving".
From an economic standpoint, the case is clear: immigration is very good for economic growth. The macro scale effect is pronounced, sticky, and it takes a spectacularly failed immigration policy to undo it.
I think that this is in question. Even in the short term, many of the employers that rely on low-wage labor encourage employees to seek transfer payments to make up the difference. Moreover, there are substantial costs associated with additional infrastructure use, education of ESL students, support of multiple languages in public and private communications, and enforcing laws on a larger and less homogeneous population.
To my knowledge, nobody has done the exact math yet, but given the fairly consistent unpopularity of increased migration rates and the relative public apathy towards corporate subsidies, I would not be surprised if it were less costly, in both economic and political terms, to simply provide a government-backed discount to the cost of domestic labor directly.
In the longer term (12 years, as opposed to 12 months), I think the stability and human capital costs of current youth unemployment alone vastly outweigh the costs of paying 20-30 percent more for labor. Moreover, quality of work is a factor - I've already seeing pretty substantial tech debt at major companies in the wake of mass H1B visa abuse (the causation is anecdotal, but consistent with the experience of every other programmer I've spoken to on the subject), and newer-build homes have a reputation for vastly poorer quality than Americans are used to.
Edit: It is worth noting that even AEI, a Bush Conservative think tank with a generally pro-migration tack, acknowledges the above (see section "Why not just bring in more immigrants?").
I think your argument applies well to misinformation/deepfakes, but not to sex/relationships/pornography, on the basis that they are targeted towards two different orders of needs. When seeking information, we care a lot about veracity and usefulness, and the value gained by small tweaks to an existing story is counteracted by the value lost by that story moving from 'entirely true' to 'not entirely true'. As an example, if there was a terrorist attack and twelve people died, then an AI edit of the story and related media that claims thirteen people died would not be straightforwardly better than the original in terms of ability to propagate or be capitalized upon. AI provides the ability to A/B test at scale and cheaply personalize, which is of relatively limited usefulness in that space.
In contrast, if I'm in the adult video business, I'm targeting the lizard brain. If an AI generated video has some immediate visual features that hit better than real ones, then my video will get more clicks, as users look for the content that elicits the strongest reaction. There isn't a fundamental limit to this, either - plenty of people already make money by producing completely animated AVs, for example, which definitely don't pass for real footage. Moreover, there's currently a tradeoff between actor quality and novelty - 'weirder' fetishes that are specific to small groups are less able to attract the most or the best performers, which serves to mitigate the "novelty spirals" that we constantly hear about, hitting a fundamental upper bound when a fetish is so specific to a given individual that nobody at all has produced a video that services it.
At a bare minimum, I think that AI-generated videos will outcompete real pornographic content and take the limits off of divergence from "normal" sexual tastes among heavier consumers. The gradient towards weirder content can become infinitely small, mitigating the shock of a piece of content that goes "too far" for someone's current preferences, and every performer can be a flawless match to the viewer's exact tastes.
He suggested the same number of people, around ten percent, are still going to real university,
That's an interesting position. It makes sense to me that that's the number that'd have the qualifications to do so, but are they still getting the same quality of education today?
I'm not sure what it's like for American high schools, but in Alberta (Canada) we had different streams for different kids. Dash-1 courses were for academic stream kids, dash-2 was either non-academic or for someone who wants a college certificate or a trade, dash-3 focuses on employability, and dash-4 is for students with learning disabilities.
That's a very common-sense system, seems like a gentler version of what Germany and Korea do. Unfortunately, the U.S. system doesn't look anything like it. We have only one track, with differentiation delivered in theory through AP courses (advanced kids) and special education (kids with severe learning disabilities) courses. Unfortunately, the former are a constant political target, and the latter are constantly in the process of "mainstreaming" students that are unsuited to standard classes by dropping them into gen. ed.
Bill Clinton (1993-2001), George W. Bush (2001-2009), Barack Obama (2008-2012), and Joe Biden (2021-2025) were U.S. Presidents with the authority to launch nuclear weapons. Despite the unilateral ability to launch nuclear weapons, they did not.
Why are Obama's second term and both of Donald Trump's terms missing?
Great article - provides clear examples of a trend that we've all quietly suspected. There is, however, one thing that I think it misses:
If the average American has barely improved, what about the intellectual class? That is, those Americans who have at least attended some college?
In the American 1950's (and in virtually every other country with a globally-respected education system today), college attendance was something for the top ten percent of the population. At present, 39 percent of 25- to 29-year-olds have a Bachelor's degree, and nearly half have a degree of some kind.
While the dumbing-down of college can be somewhat ameliorated by the stratification of the college experience that we've seen (a State U degree used to be prestigious!), with 'top-college education' replacing 'college education' in our culture, in practice, it's very hard to make a clean break, especially seeing as the U.S. education system is allergic to the idea that one person might ever be inextricably more qualified than another on the basis of their intelligence or natural work ethic. Proposals for a German - style path system, in which students that aren't likely to thrive in higher education are directed towards careers they're more likely to do well with, have been repeatedly shot down on this grounds.
Worth pointing out that high school has had a similar issue. NCLB made a lot of people feel better about themselves, but it essentially destroyed the last vestiges of rigor in the public school system. You can't build a meaningful HS degree around the -2sd kid that isn't allowed to wait a year and go over difficult material again, and is strongly discouraged from dropping out at 17 and getting an early start on his career (which would be significantly kinder for him - the extra money helps, and not everyone can learn Calculus).
1
I think this is a useful goal. I'd pose that quantifying human research taste seems like the best starting point. Can a human researcher achieve a high score on this metric?
A metric I would propose is somewhat different, and I think potentially less vulnerable to noise and measurement error:
For example, if I were evaluating an LLM that had stopped training just after DiffPure released, I might ask it how best to combat this defense from the perspective of the attacker. I'd then compare its proposal to the methods demonstrated by DiffAttack, and rank LLMs (and humans - either queried before the cutoff or from separate subfields without knowledge of the ground truth outcome) by having an evaluator repeatedly decide which of two models' proposals best match what ended up working. This might look like: