One of the lessons we've had as the Federal, multi-branch nature of the US governmennt has frustrated Trump is that the government may be fucking us over, but it's not doing it in *unison*. It's doing it piecemiel, on the initiative of many interests working against each other, just as the framers intended. The motto on the Great Seal notwithstanding, there are myriad roadblocks to consolidating power in the hands of a single individual. It takes time and repeated failures. This is why the second Trump Adminsitration is worse than the first; they've figured out ways around things like Congressional power of the purse, put more of their henchmen in the judiciary, and normalized Congress lying down and letting the president walk all over them. It's a serious situation, although fortunately Trump isn't long for this world.

Yeah, but other than that, what has Elon Musk ever done for us?

While that's true, a responsible generation aims to boost the next generation to a *higher* level than the education they received. The world has become more complex and faster-paced, and even if that weren't true, the consequenes of aiming high and falling short are better than the consequences of aiming for the status quo and falling short.

So while I'm 100% onboard with skepticism that technology will magically make education better, I think the argument that "the education I got worked for me should be good for them" isn't a strong argument. What we need is a better ecducation that would have been a better education fifty years ago: stronger math, science, and language skills, general knowledge, and, I think critical thinking and media literacy. Possibly emotional intelligence -- it's kind of pointless to teach people critcial thinking skills if they are carried away by emotions.

There are no economic or security reasons to blockade Cuba, so that leaves *political*.

It used to be believed that bullies were low status individuals who are lashing out out of frustration. But research has shown that bullying is an effective strategy for achieving and maintaining social status. In other words it's a political winner. So the focus of research has shifted from the bully to the people around him who enable the bullying. The inner circle are the henchmen -- people without the charisma and daring to initiate the bullying, but join in when the bully gets things started. Around them are the audience, the people who wouldn't risk participating but enjoy the bullying vicariously. And around them are the much larger group of bystanders, who don't approve but are waiting for someone else to stop the bullying. Then off to the side are the defenders, who stand up to the bully.

Perhaps the least appreciated supporting factor in the phenomenon of the high-status bully is the silence of the bystanders, which is dependent upon the perception of widespread approval. Since you can't visibly see the the line between the approving audience and the apalled bystanders, the silence of the bytstanders is absolutely essential in sustaining the bullying.

Lot's of Americans are apalled at the idea of using military force to inflict suffering on the Cuban people. But that's only politically advantageous *because* of *them*. Tney are indistinguishable from the relatively small number of people who are thrilled when Trump announced he can do anything he wants wtih Cuba. The gap between actual approval and *perceived* approval is absolutely critical in establishign and maintaining any kind of authoritarianism. This is why would be authoritarian leaders are so focused on punishing and marginalizing any kind of expression of disapproval.

To add to the parent post, the paper appears to be the first step in the scientific method: "Notice a trend".

The next steps will be "form a hypothesis", "construct a test to confirm or deny the hypothesis", "perform the test"... and so on.

In this specific case, "perform the test" might be impossible to do for ethical reasons - you can't take people at random and sit them down in front of a LLM and test their level of psychosis before and after, because of that pesky "do no harm" rule.

But we might be able to find people who have had their psychosis levels measured before LLMs became available, and whose LLM accounts will accurately show how much LLM usage they have, and we can then remeasure their levels of psychosis and see if this correlates with LLM account usage.

Or some other test like that.

The paper appears to be an attempt to raise the issue and start a conversation. From the abstract:

From the parent post:

I'm also heavily influenced by TV, and have spent a lot of time trying to sort out beliefs that come from TV from beliefs that come from experience or research.

I'm constantly presented with a situation or belief and have to pause to reflect and say "I believe that because it was on TV, it's probably not real". Many of my opinions on the police, government agencies, other countries, world events, and social constructs come not from experience, but on how they were portrayed on TV.

We're hard-wired to believe what people tell us, it's a cognitive shortcut in an environment where you can't know anything, but lots and lots of what we think today are only dramatic choices intended to provoke emotional response. (Compare with news reporting today. On both sides.)

For example, I've met people who won't go hiking because of all the bugs, skunks, poison ivy, and bears.

Assuming that LLMs are content neutral, I think in 10 years or so we're going to find people whose worldview is a greatly amplified version of random events that were highlighted when they were kids.

In 1790, the US population was 94.9% rural. There is no country. in the world today that rural -- Burundi, which looks like blanks spot in the world at night satellite picturs, is 88% rural.

The largest city at the time was New York, with a population of 33,000. Northern Manhattan was near-wilderness, mid-town was farms and country houses.

In 1790 the US was. country you could "police" with sheriffs and volunteer posses, largely to keep the peace. If you got robbed, you hired a private thief catcher. This works in a 95% rural country with just 3.4 million inhabitants. It would be chaos in a country 87x larger.

8GB is plenty of RAM for what most users use their laptops for, if the OS doesn't gobble it up with background tasks. Probably be a good idea to steer clear of Chrome as a browser on these things.

It's actually more like an iPhone 16 Pro runing MacOS in a laptop form factor. Apple basically rummaged through their parts box and pulled out a mobile CPU that'll deliver 50% more single core performance than what's in a high-end Chromebook with only 80% of the power draw. And Apple's got *massive* economies of scale on those parts, so they can afford to deliver a lot of bang for the buck.

The only place the Neo appears to falls short is in RAM, but this is *not* a power user machine, it's for basic office tasks and multimedia consumption. Realistically 8GB is plenty for many users.

In any case, the desktop isn't the center of most users's universe anymore; the switchboard of their life is their smartphone. This is a gateway drug to MacOS IOS integration, and eventually onto the upgrade treadmill. Users will switch seamlewssly between their iPhones and Neos all day long, with data on iCloud and iMusic etc., and when it comes time to upgrade their phone or their laptop, they won't be *stuck* exactly, but if they leave the reservation they lose a lot. But they certainly could upgrade to a *much nicer* Macbook....

It's no wonder the other laptop makers are sitting up and taking notice. Apple has set up a one way conversion ratchet for people tempted by a really nice and perfectly adequate entry level machine at an entry level price.Nobody else has the vertical integration -- chip foundries to device manufacturing, to software platform -- spanning desktop and phones that's needed to do this.

Anyone who's watched a house go up has marveled at how quickly the framing goes up, then how long it takes everything else to get done.

Framing is about 1/l4 of the build time for a house. The *labor* for framing is less than 10% of the build cost. If the machine cost *nothing*, and framed the building *instantaneously*, those are hard limits on how much faster and cheaper the house building robot could make the process: about 25% faster with about a 10% cost reduction. But the machine wouldn't work instantaneously, nor would it be free.

There already is a better way of doing this. You prefabricate the house in units, ship them to the site, then bolt the units together. The modules could be completely finished at the factory. Savings over traditional construction would be substantial -- 40%. The problem is, can you build houses people want to buy and which local building codes will allow you to live in. If you throw out expectations that a house looks like a house a child would draw with crayons, you can build a really nice. So with prefab houses you either have things that look like mobile homes; or things that look like they were designed by a scandanavian architect. Houses that *look* like mid-range, hand-built homes are a tough nut to crack.

There was a movement among architects to use pre-fabricated construction to solve the problem of housing returning GIs after WW2. It didn't catch on as the kind of democratizing mass produced housing the movement envisioned because people wanted a house that looked hand-built. But if you can get over that, it produced some really great houses. One of the more famous examples (although not completely pre-fabricated) is the Eames House. There's a company from that period that's still in business, but they pre-fabricate million dollar luxury homes, not mass produced housing.

The obstacles to prefabricated houses are regulatory, which is why it can't reach the middle of the market. Anti-mobile home rule discourage really cheap pre-fabricated houses, but high end producers can afford to jump through the regulatory hoops. For mid-range houses, the regulatory burden outweighs the economic advantage of prefabrication. This could allow a framing robot to have a niche, although as I pointed out it won't save much money on the build cost.

LLAMA was [illegally] released into the public three years ago (to the day - March 3, 2023), and it's estimated that ten years of AI improvements happened in the subsequent 6 months. People were doing all sorts of things with LLMs that meta hadn't thought of, or didn't have time to develop. Such as text-to-audio, local LLM use, and automated manuscript generation.

All these attempts at monetizing the LLMs are, at the same time, holding back the progress of AI development. If OpenAI wants to leap ahead of the competition, they should put their language model online and see what the community comes up with.

I get it - training a LLM takes roughly $100 million for the initial dataset, and companies need to recoup this expense.

Still, I'm saddened that I can only use the system for purposes that the company approves of, and in ways that they have already thought of.

There's a lot of potential there, and we're not making good use of that.

THis is true. You also get a better understanding for AI's limitations, familiarity of where to be skeptical, and a feel for when a context is losing its mind. I like to put it this way: nobody who can't outthink an AI should be using an AI for anything important.

It wouldn't be news if you looked at their terms of service -- which you should. The ToS explicitly say they use a combination of automated systems, human review, and reports to identify and investigate violations of their usage terms, including violence, abuse, fraud, impersonation, disinformation, foreign influence campaigns , abusive sexual content, and academic dishonesty. This includes "anonymous" sessions that are saved for a minimum of 30 days. You have no expectation of privacy from the provider's compliance teams.

This is *absolutely* standard among the major online players. So why not use a local AI workstation with a couple of big-ass GPU cards in it to run the campaign? That's what they *should* have done. But the major online players like ChatGPT and Claude are much better at realistic content generation than the widely available local models you can run.

What they should have done is design and run the compaign on a local AI workstation, and used the local workstation to generate prompts they could feed into burner accounts on public services like ChatGPT and Claude. But they got lazy and ran the *whole* operation in ChatGPT, right in plain fiew of the OpenAI compliance teams the ToS they evidently didn't read would have told them were there. They even did *performance reviews* in the same account.

Remember folks, these "spooks" are just mid-level paper-pushers in an opaque communist bureacuracy. You can never discount inertia in such an environment. Because this was something new, they might even have had trouble getting the purchase of some high end GPUs approved.

I don't hate Trump. I pity him. No matter how much shit he slaps his name on, it won't fill the hole is daddy left in him. But people that damaged shouldn't be let anywhere near power.

Yes, because those are totally a real thing.

Well, no. It's true you can't buy books for the purposes of scanning them *and then making them available online* (Hachette v. Internet Archive). Scanning them for AI training is not settled law in every Federal District, although in at least one that has been ruled transformative and therefore allowable (Bartz v Anthropic, Northern District of California).