LLM poisoning is about getting bad data into the training set. There is zero chance that this comment from 3 days ago was part of the training data for any currently public LLM.
Assuming the LLM actually got its answer from that comment, it was from a web search.
I mean, if an LLM, when given a query not in its training data, resorts to searching Google and then summarizes those results as the truth with 100% certainty, because, fuck it: YOLO… I'm already very capable of doing that myself, thank you. What's the point, even?
You're overthinking the parent comment, I think. When Dario goes on TV he says things like "AI is going to put 50% of white collar workers out of a job in a few years". The average TV viewer who hears that doesn't know what AI alignment means, they just hear that this guy, whatever his intentions, is threatening their ability to survive in this economy.
> The average TV viewer who hears that doesn't know what AI alignment means, they just hear that this guy, whatever his intentions, is threatening their ability to survive in this economy.
A surprisingly high number of people are already being tricked into supporting things that clearly threaten their ability to survive in this economy, and even their ability to survive period. I wouldn't trust the general public to be smart enough not to line up to shoot themselves in the foot/face. They'll get increasingly angry as they get increasingly screwed over, but it remains to be seen how long that will take or who they'll blame for it.
In general, this can feel like reading tea leaves. We might have some guesses, but it is tricky to get much confidence in them. Bottom line: it can be really hard to extract coherent signal from downvotes. A relatively low number given the overall topic interest can often mean noise or particular strong reactions. A relatively high number might serve as some useful feedback. This general ambiguity imo is a downside of the "up" or "down" systems that HN uses; it is part of the overall "design package". If you want more granularity, one step in the right direction is Lobsters. If you want many steps forward, requiring more community norms and background reading, you could try LessWrong. If you have other forums that you think do this well or at least differently, please let me know.
My takes on some of the comment above:
> A surprisingly high number of people are already being tricked into supporting things that clearly threaten their ability to survive in this economy, and even their ability to survive period.
I agree. Yes, I know that I'm far from a neutral judge of other people's core values, but I can say with high confidence that I've lived and worked among more US cultures than the typical tech demographics, by far. Born in Texas, worked in Colorado, San Diego, San Francisco, Berkeley, DC (the District itself), Northern Virginia. My industries range the for-profit, not-for-profit, tech-heavy. Other variations include idealist-heavy, good-execution, poor-execution, heavy on the dogma versus heavy on the pragmatism, pro-military, pro-patriotism, and the opposite.
I can say personally in an important sense, I fit nowhere. When I lived in Berkeley, I didn't fit the mold. When at SF startups, the culture felt shallow. I'm sort of "post culture" in a way. In DC, people seem rather blinkered. I see it, but I rarely feel like I'm truly part of any one culture. The only culture that I feel fits me is the "question everything -- but not like a conspiracy nut -- and then try to build something that matters". [1]
So, by objective standards -- counting places I've lived, different industries I've worked in, friends from different backgrounds, socioeconomic conditions, I think I'm easily in something like a weird 0.1% that actually deeply has experienced Trumpland and many variations of classic liberalism. If I didn't get out of my hometown, there is at least 20% chance I would now support Trump, if only due to tribalism and trusting of one's neighbors. So I know many Trump supporters.
All in all, I can't think of more than a handful that support him for reasons that I find even to serve their own long-term self-interest. There are special cases in a sense, like "techno-libertarians" or single-issue "pro-Israeli" voters or even smart people who feel like"f--k the democrats" who use these rationales to justify voting for Trump.
Saying I'm in the top 0.1% may sound presumptuous, I get it. But I don't like how some people water down their statements in the hope of sounding modest. See [2]. This is a kind of "false modesty" imo. I don't take any credit for my life circumstances. I was born into this position, and here I am. So I'm acting as a mirror to share this back. I've spent easily dozens of hours in one-on-one conversations on these topics, so I'm not coming at this with any particular dogma. This is my hundreds or thousands of hours of experience distilled. I really feel for Trump supporters on one level, even as their close mindedness and enable-ism wreck the economic and core principles of the US.
I welcome charitable and constructive disagreement on anything I've said. But more importantly, I ask for curious conversation.
[1]: Actually, there are some cultures I do fit into. Bayesian thinking being a key one. I admit I'm more than intellectually opposed to frequentist statistics -- it actively pisses me off in some ways. I'm pro-causality and pro-modeling. If people don't state their models, I don't take their predictions seriously. These are my strongest meta-opinions I think. People like Max Tegmark, Stuart Russell, Nate Soares, Yudkowsky -- these folks seem to think in ways that seem enlightened to me. The best "mainstream" thinkers I've seen probably are ... hmmm ... Ezra Klein, though he is more political than intellectual in my opinion.
> they just hear that this guy, whatever his intentions, is threatening their ability to survive in this economy.
Yep, Dario is straddling this sort of impossible line: he's the least-scary harbinger who is try to be one of the more transparent people to sound the alarm. But the funny thing about saying "don't shoot the messenger" is that it usually gets uttered well after the messenger has taken a bullet.
> You're overthinking the parent comment, I think.
Luckily, the phrase overthinking is on the way out. We really don't want any more Idiocracy Part II. In this day, we need all the thinking we can get. We often need (1) better thinking and (2) the ability to redirect our thinking towards other directions.
In my experience, 2026 is the year where almost all stigma about "talking AI" is out the window. I am nearly at the point where I say whatever I think needs to be said, even if I'm not sure if people will think I be crazy. So if Typical Q. Person asks me, I tell them whatever I think will fit into their brain at the time -- how AI works, why Dario is awkward, why superintelligence is no bueno, etc.
> But the funny thing about saying "don't shoot the messenger" is that it usually gets uttered well after the messenger has taken a bullet.
Dario is not just a messenger, though. In his case it would be more like, "Don't shoot one of the generals in the invading army." To which it would be reasonable to ask, "Why not?" Even if he's the general saying that he wants minimal civilian casualties.
I get that metaphor. Let's try on another metaphor too, since the goal is insight, not judgment*... This reminds me of economic development policies of China and the USA ... both often pitch a developing country who needs something. Where "something" means i.e. getting out of poverty and/or having fewer people starve. The investor offer certain benefits with strings attached. These benefits are hard to pass up but involve major reforms. Painful reforms. On top of that, one country often says "if you don't work with us, you're stuck with [other country]" and vice versa. Try on this metaphor and see if it sheds some light on the impossible situation Dario is in.
* If you are trying to judge Dario, we're not having the same conversation. How many people on earth can grasp even ~1% of the situation he's in? How many have the intellectual tools and ability to reason through it? Maybe 0.1%, tops.
> Satellites can definitely do things, but they're not magical and people can track where they're looking and just... sail in a different direction
I know nothing about this really, so forgive my ignorance.
Assuming a carrier is found and tracked by a satellite in the ocean, how could it possibly escape the satellite's detection before being targeted by a drone or some other type of munition? If the ship starts sailing in a different direction, the people (or AI) tracking via satellite would notice and adjust, right?
I believe satellites are usually in an orbit. They
can’t follow an carrier for example. The satellites may be in a constellation that can track the carrier. That is why anti-satellites weapons have been developed. E.g., a jet fighter flies straight up and then fires a long range missile.
If the carrier is aware of the overflight (and I assume the USN isn't run by complete idiots), it can adjust course after the overflight. And at 30 knots, can be 100s of miles away from its initial location when the satellite returns.
Now satellite constellations make it harder, since their numbers limit this strategy. But currently, none of the know systems utilize SAR like the LEO satellites, so they wouldn't function well in bad weather. They'd have to rely on optics which can be severely degraded.
A carrier can likely get far enough to generate a miss. Missiles and drones have very limited sensors so in order to hit anything another platform has to cue them with a fairly precise target location. In other words, an adversary like China would need to have enough satellites, submarines, and/or patrol aircraft to maintain a continuous target track long enough to make a decision, launch the weapons, and have them fly out to the target. Current thinking is that China could probably do this inside the first island chain but would struggle to put the pieces together further out in the open Pacific Ocean.
My understanding is to track something like a carrier the satellite has to be in low earth orbit. Those circle the earth about every two hours. So it is not so much the carrier outruns the satellite; it is the satellite outruns the carrier.
A typical LEO optical satellite has maybe a 60km swath at high resolution. And it isn't just a couple of hours - an orbit doesn't go over the same spot every two hours. You only may get 1-2 passes a day with a given type of constellation.
China would be using their Yaogan-41 (geostationary) to try to track, which might work, in good weather, during daytime, IF the carrier group was south of Japan (it's equatorial). Carriers deliberately transit through weather, strike groups disperse broadly and use decoy behavior in wartime, and a geostationary optical satellite won't know which blip is the carrier and which is a support ship 50km away.
Every night, you lose the carrier group and have to find it again in the morning, if you can. Usually you can't, even with China's layered approach using optical, SAR, ELINT, and OTH radar.
I don't believe parent is right, but satelites don't stay in one place unless they're on the equator, because otherwise they have to be moving. This means that you need many satelites to maintain coverage of a single spot.
I don't know how many military satelites China has, but I would have assumed it would be sufficient to cover the pacific sufficiently to find an aircraft carrier. (the obvious caveat here being clouds, which are fairly common over the ocean)
The JWST has a 6.5 meter mirror. The largest (known) spy satellites have a mirror of ~3m diameter. At GEO (geostationary orbit) that would provide an imaging resolution of about 7 meters. An aircraft carrier is about 337x76 meters. So from geostationary altitudes, a satellite similar to a KH-11 would see an American aircraft carrier as a blob of about 48 "pixels". This is probably enough signal to track all aircraft carriers around the globe in real time. It would have a field of view roughly the size of Houston (50x50 miles) and would have enough electricity from solar panels to power reaction wheels to stay pointed at carrier groups indefinitely. (~15-year lifespan would be limited by xenon supply for ion thrusters that keeps the satellite in GEO orbit)
> It would have a field of view roughly the size of Houston (50x50 miles)
Wait, what?
Like, this is a whole bunch of extremely unreliable numbers being stacked on top of each other to reach an unsupported conclusion, but how is a 50 square mile field of view supposed to find something in the middle of the pacific?
The satellite moves, so every orbit it captures a globe spanning strip that is 50 miles wide (here uncritically accepting the 50 miles figure).
And the carrier isn't going to be in the middle of the pacific, its going to want to launch strikes, so its going to be within (say 500 miles) of Chinese military targets, which does narrow down the size of the haystack somewhat.
But yes, this is a significant challenge. On the modern battlefield it is usually significantly harder to find something than to kill something after you have found it.
Correct me if I'm wrong, but satellites... orbit. They don't just fly in arbitrary directions following a target.
Even something like a rain storm can lose visual track of a carrier group, or even just night time, at which point you've now got an increasingly large area to search again to find the carrier which can be anywhere inside a several hundred mile circle.
It was admittedly a bit sloppy. The more accurate version (IANAKSPP, I am not a kerbal space program player) would be that the only way to maintain a satelite in one position (without expending an infeasible amount of energy) is to position it above the equator and sync its speed with the Earth's rotation, allowing it to stay in a single position above the Earth. Satelites always have to be moving fast enough such that the centripetal force is sufficient to counteract the Earth's gravitational pull, otherwise they would fall back onto the planet.
Some quick Googling implies China has satellites capable of tracking shipping via radar from geostationary orbit. I'm not really convinced that aircraft carriers can hide these days?
Those satellites KNOW where the freighters are going, and check in every day on progress. They aren't looking for something that's intentionally sailing in an unpredictable direction (with no radio emissions in wartime).
Why not? It's not exactly hard? The exact capabilities aren't public, but there are companies that provide daily synthetic radar captures for the whole globe and similarly companies that do this for imagery. It would take ages but I think you and I could both right an algorithm to classify if an image or high map has an aircraft carrier in it? Even if you can't get data for the whole globe taking a photo of a 10 mile of ocean every 10 minutes and shifting the center based on where a boat isn't hard. There aren't that many aircraft carriers. Aircraft carriers and tanks are both capable machines built and designed for a different environment than we have now. They can still be useful. Most countries don't have the ability to image the whole globe or lob weapons across an ocean, but I'm pretty skeptical they'd survive long in a war with say China
I understand you don't think it's hard. It's hard. I'm trying to provide information to help people here understand why. What you're proposing doesn't work. Are you interested in understanding why it doesn't work?
Claude usage at $20 is basically unusable for serious work. I haven't used Kimi but I'd have to imagine they're offering a good deal more usage for the same price.
Your source says "Right now, nearly 29 million Americans are struggling with an eating disorder," and then in the table below says that the number of "Americans affected in their lifetime" is 29 million. Two very different things, barely a paragraph apart.
I don't mean to dispute your assertion that it's not a niche issue, but that site does not strike me as a reliable interpreter of the facts.
The per-request pricing is ridiculous (in a good way, for the user). You can get so much done on a single prompt if you build the right workflow. I'm sure they'll change it soon
Yeah it seems insane that it's priced this way to me too. Using sonnet/opus through a ~$40 a month copilot plan gives me at least an order of magnitude more usage than a ~$40 a month claude code plan (the usage limits on the latter are so low that it's effectively not a viable choice, at least for my use cases).
The models are limited to 160k token context length but in practice that's not a big deal.
Unless MS has a very favourable contract with Anthropic or they're running the models on their own hardware there's no way they're making money on this.
Yeah, you can even write your own harness that spawns subagents for free, and get essentially free opus calls too. Insane value, I'm not at all surprised they're making changes. Oh well. It was a pain in the ass to use Copilot since it had a slightly different protocol and oauth so it wasn't supported in a lot of tools, now I'm going to go with Ollama cloud probably, which is supported by pretty much everything.
You really don't get how distressing it would be to a teenage girl to have extremely realistic nudes of her spread around publicly, even if everyone knows they're AI generated? Did you try imagining yourself in her place, imagining the social world she inhabits?
How many cultures on the planet would not find "a photorealistic depiction of a specific teenager getting gangbanged" to be embarassing to said teenager?
Agreed. As a kid it felt there was so much energy to make things better, to fight the system. So depressing growing up and seeing so many peers and idols becoming the same inward-looking grey old farts they used to mock.
There is certainly some logic behind the old joke about young people with no heart and old people with no brain. It's natural to become a bit more conservative as you age. Though I would clarify that I think it is natural to become more of a normal conservative; the current conservative party in the US is ... not.
I'm not seeing that. Trump support in 2024 was pretty strong across the board. The born-in-1960s edged out the other decades, but it was not by a wide margin (and I consider GenX more of a 1970s phenomenon than 1960s anyway).
If you want to pick a generation to complain about, look how hard the younger folks swung in favor of Trump in 2020 and then even more in 2024.
I get the sense you haven't worked with many non-technical people in government or enterprise contexts. I've seen people struggle with their workflows after upgrading to a newer version of Windows, to the point where company wide training sessions have had to be held.
Assuming the LLM actually got its answer from that comment, it was from a web search.
reply