The Future of Everything Is Lies, I Guess: Safety

(aphyr.com)

127 points | by aphyr 2 hours ago

17 comments

  • atleastoptimal 11 minutes ago
    There really are only 3 options that don't involve human destruction:

    1. AI becomes a highly protected technology, a totalitarian world government retains a monopoly on its powers and enforces use, and offers it to those with preexisting connections: permanent underclass outcome

    2. Somehow the world agrees to stop building AI and keep tech in many fields at a permanent pre-2026 level: soft butlerian jihad

    3. Futurama: somehow we get ASI and a magical balance of weirdness and dance of continual disruption keeps apocalypse in check and we accept a constant steady-state transformation without paperclipocalypse

    • operatingthetan 9 minutes ago
      Or we keep building AI and no apocalypse happens.
    • raincole 9 minutes ago
      In other words, only one option.
  • philipkglass 38 minutes ago
    In short, the ML industry is creating the conditions under which anyone with sufficient funds can train an unaligned model. Rather than raise the bar against malicious AI, ML companies have lowered it.

    This is true, and I believe that the "sufficient funds" threshold will keep dropping too. It's a relief more than a concern, because I don't trust that big models from American or Chinese labs will always be aligned with what I need. There are probably a lot of people in the world whose interests are not especially aligned with the interests of the current AI research leaders.

    "Don't turn the visible universe into paperclips" is a practically universal "good alignment" but the models we have can't do that anyhow. The actual refusal guards frontier models come with are a lot more culturally/historically contingent and less universal. Lumping them all under "safety" presupposes the outcome of a debate that has been philosophically unresolved forever. If we get hundreds of strong models from different groups all over the world, I think that it will improve the net utility of AI and disarm the possibility of one lab or a small cartel using it to control the rest of us.

  • jagged-chisel 25 minutes ago
    "Alignment"

    In what world would I ever expect a commercial (or governmental) entity to have precise alignment with me personally, or even with my own business? I argue those relationships are necessarily adversarial, and trusting anyone else to align their "AI" tool to my goals, needs, and/or desires is a recipe for having my livelihood completely reassigned into someone else's wallet.

    • sigbottle 16 minutes ago
      Interesting you single out commercial and government entities but not people. What defines the difference? Bureaucracy? Concentration of resources? Legal theory?

      I guess I'm trying to wonder why this line of thinking (in theory) doesn't turn to paranoia about everybody. I don't know much ethics or political theory or anything.

    • __MatrixMan__ 11 minutes ago
      You could expect such a thing in a world where consent was currency, rather than scarcity.
  • Cynddl 2 hours ago
    > "Unavailable Due to the UK Online Safety Act"

    Anyone outside the UK can share what this is about?

    • 0x3444ac53 1 hour ago
    • starik36 13 minutes ago
      What specifically is unsafe in this article?
    • satvikpendem 21 minutes ago
      Ironic.
    • jazzpush2 2 hours ago
      The Future of Everything is Lies, I Guess: Safety Software LLM The Future of Everything is Lies I Guess 2026-04-13 New machine learning systems endanger our psychological and physical safety. The idea that ML companies will ensure “AI” is broadly aligned with human interests is naïve: allowing the production of “friendly” models has necessarily enabled the production of “evil” ones. Even “friendly” LLMs are security nightmares. The “lethal trifecta” is in fact a unifecta: LLMs simply cannot safely be given the power to fuck things up. LLMs change the cost balance for malicious attackers, enabling new scales of sophisticated, targeted security attacks, fraud, and harassment. Models can produce text and imagery that is difficult for humans to bear; I expect an increased burden to fall on moderators. Semi-autonomous weapons are already here, and their capabilities will only expand.

      Alignment is a Joke Well-meaning people are trying very hard to ensure LLMs are friendly to humans. This undertaking is called alignment. I don’t think it’s going to work.

      First, ML models are a giant pile of linear algebra. Unlike human brains, which are biologically predisposed to acquire prosocial behavior, there is nothing intrinsic in the mathematics or hardware that ensures models are nice. Instead, alignment is purely a product of the corpus and training process: OpenAI has enormous teams of people who spend time talking to LLMs, evaluating what they say, and adjusting weights to make them nice. They also build secondary LLMs which double-check that the core LLM is not telling people how to build pipe bombs. Both of these things are optional and expensive. All it takes to get an unaligned model is for an unscrupulous entity to train one and not do that work—or to do it poorly.

      I see four moats that could prevent this from happening.

      First, training and inference hardware could be difficult to access. This clearly won’t last. The entire tech industry is gearing up to produce ML hardware and building datacenters at an incredible clip. Microsoft, Oracle, and Amazon are tripping over themselves to rent training clusters to anyone who asks, and economies of scale are rapidly lowering costs.

      Second, the mathematics and software that go into the training and inference process could be kept secret. The math is all published, so that’s not going to stop anyone. The software generally remains secret sauce, but I don’t think that will hold for long. There are a lot of people working at frontier labs; those people will move to other jobs and their expertise will gradually become common knowledge. I would be shocked if state actors were not trying to exfiltrate data from OpenAI et al. like Saudi Arabia did to Twitter, or China has been doing to a good chunk of the US tech industry for the last twenty years.

      Third, training corpuses could be difficult to acquire. This cat has never seen the inside of a bag. Meta trained their LLM by torrenting pirated books and scraping the Internet. Both of these things are easy to do. There are whole companies which offer web scraping as a service; they spread requests across vast arrays of residential proxies to make it difficult to identify and block.

      Fourth, there’s the small armies of contractors who do the work of judging LLM responses during the reinforcement learning process; as the quip goes, “AI” stands for African Intelligence. This takes money to do yourself, but it is possible to piggyback off the work of others by training your model off another model’s outputs. OpenAI thinks Deepseek did exactly that.

      In short, the ML industry is creating the conditions under which anyone with sufficient funds can train an unaligned model. Rather than raise the bar against malicious AI, ML companies have lowered it.

      To make matters worse, the current efforts at alignment don’t seem to be working all that well. LLMs are complex chaotic systems, and we don’t really understand how they work or how to make them safe. Even after shoveling piles of money and gobstoppingly smart engineers at the problem for years, supposedly aligned LLMs keep sexting kids, obliteration attacks can convince models to generate images of violence, and anyone can go and download “uncensored” versions of models. Of course alignment prevents many terrible things from happening, but models are run many times, so there are many chances for the safeguards to fail. Alignment which prevents 99% of hate speech still generates an awful lot of hate speech. The LLM only has to give usable instructions for making a bioweapon once.

      We should assume that any “friendly” model built will have an equivalently powerful “evil” version in a few years. If you do not want the evil version to exist, you should not build the friendly one! You should definitely not reorient a good chunk of the US economy toward making evil models easier to train. ...

      • jazzpush2 1 hour ago
        To be clear, that's not the full article, just the intro (though the whole thing isn't too long)
    • sieabahlpark 37 minutes ago
      [dead]
  • macintux 1 hour ago
  • nzoschke 37 minutes ago
    Excellent articles as expected from aphyr.

    I'm seeing that these tools are extremely powerful the hands of experts that already understand software engineering, security, observability, and system reliability / safety.

    And extremely dangerous in the hands of people that don't understand any of this.

    Perhaps reality of economics and safety will kick in, and inexperienced people will stop making expensive and dangerous mistakes.

    • mursu 26 minutes ago
      The future is happening. Instead of trying to raise awareness about evil AI... I think it would be more healthy if we could direct this energy to ways of improving the situation without condemning the unknown of AI evolution. As with anything.. there will be a bad side.. The bad guys will always be there.. be it AI or soccer matches.. should we stop developing nuclear energy because nuclear weapons are developed?
      • fmbb 7 minutes ago
        There is no natural law saying the good sides of any kind of tech will outweigh any bad sides.

        ”The future” is happening because it is allowed in our current legal framework and because investors want to make it happen. It is not ”happening” because it is good or desirable or unavoidable.

  • jbreckmckye 19 minutes ago
    Optimists would argue that the answer to bad actors using AI is good actors using AI. I think this adds marginal costs that eat into the efficiency benefits of digitisation, and instead we will simply see certain things de-digitise

    An example is interviewing and jobs. A fully digitised recruitment pipeline - Zoom calls, CVs, GitHub profiles - is too easy to defraud. I remember even before the pandemic, any kind of remote role would attract hundreds of applications of dubious quality.

    The most likely outcome as I see it, is companies will simply demand in person interviews. Probably only at the final stage, but they will want that in-person verification.

    This was a domain which we digitised to make efficient: lower friction, more accessible, more standardised. It worked until it invited fraud. Detecting fraud is complex and expensive, so, it's easier just to reverse some digitisation

    Another example is education. Universities teach using AI generated scripts and they grade AI generated essays. Students are under too much financial pressure to pay for slop, and, institutions that are subject to fraud will end up trashing their reputation.

    So the opportunity emerges for more competitive universities to differentiate themselves by assessing on blue-book exams. Students won't like it, but they'll respond to incentives when employers filter out anyone with a highly digitised degree

    Ultimately I suspect AI will to some extent corrode the value of digital information, just by generally producing distrust.

    In some ways "slop" is not actually a new problem: we have had people generating spam, scams, and algorithm-bait for many years. But the volume of fraud and the cost of suffering it was acceptable enough. That may have changed

  • Imnimo 1 hour ago
    >Unlike human brains, which are biologically predisposed to acquire prosocial behavior, there is nothing intrinsic in the mathematics or hardware that ensures models are nice.

    How did brains acquire this predisposition if there is nothing intrinsic in the mathematics or hardware? The answer is "through evolution" which is just an alternative optimization procedure.

    • fmbb 1 minute ago
      Well, through natural selection in nature.

      Large language models are not evolving in nature under natural selection. They are evolving under unnatural selection and not optimizing for human survival.

      They are also not human.

      Tigers, hippos and SARS-CoV-2 also developed ”through evolution”. That does not make them safe to work around.

    • order-matters 42 minutes ago
      natural selection. cooperation is a dominant strategy in indefinitely repeating games of the prisoners dilemma, for example. We also have to mate and care for our young for a very long time, and while it may be true that individuals can get away with not being nice about this, we have had to be largely nice about it as a whole to get to where we are.

      while under the umbrella of evolution, if you really want to boil it down to an optimization procedure then at the very least you need to accurately model human emotion, which is wildly inconsistent, and our selection bias for mating. If you can do that, then you might as well go take-over the online dating market

    • Terr_ 50 minutes ago
      > just an alternative optimization procedure

      This "just" is... not-incorrect, but also not really actionable/relevant.

      1. LLMs aren't a fully genetic algorithm exploring the space of all possible "neuron" architectures. The "social" capabilities we want may not be possible to acquire through the weight-based stuff going on now.

      2. In biological life, a big part of that is detecting "thing like me", for finding a mate, kin-selection, etc. We do not want our LLM-driven systems to discriminate against humans in favor of other systems. (In some cases they already do.)

      3. The humans involved making/selling them will never spend the necessary money to do it.

      4. Even with investment, the number of iterations and years involved to get the same "optimization" result may be excessive.

      • fweimer 17 minutes ago
        While I don't disagree about (2), my experience suggests that LLMs are biased towards generating code for future maintenance by LLMs. Unless instructed otherwise, they avoid abstractions that reduce repetitive patterns and would help future human maintainers. The capitalist environment of LLMs seems to encourage such traits, too.

        (Apart from that, I'm generally suspect of evolution-based arguments because they are often structurally identical to saying “God willed it, so it must true”.)

    • pants2 49 minutes ago
      This Veritasium video is excellent, and makes the argument that there is something intrinsic in mathematics (game theory) that encourages prosocial behavior.

      https://www.youtube.com/watch?v=mScpHTIi-kM

    • almostdeadguy 34 minutes ago
      There’s a funny tendency among AI enthusiasts to think any contrast to humans is analogy in disguise.

      Putting aside malicious actors, the analogy here means benevolent actors could spend more time and money training AI models to behave pro-socially than than evolutionary pressures put on humanity. After all, they control the that optimization procedure! So we shouldn’t be able to point to examples of frontier models engaging in malicious behavior, right?

    • miltonlost 46 minutes ago
      "just" is doing a lot of lifting here
    • cowpig 1 hour ago
      There are also many biological examples of evolution producing "anti-social" outcomes. Many creatures are not social. Most creatures are not social with respect to human goals.
      • nyrikki 53 minutes ago
        There is a reason we don’t allow corvids to choose if a person gets a medical treatment or not.
      • b00ty4breakfast 48 minutes ago
        Luckily, this is a discussion of humans.
        • fmbb 5 minutes ago
          This is a discussion about large language models.
  • simianwords 28 minutes ago
    The author is still grieving by watching a civilisation changing technology just passing by. Every single one of the problems they note applies to any technology that existed.

    The internet produced 4chan. Produced scammers. Produced fraud. Instrumental in spreading child porn. Caused suicides. Many people lost their lives due to bullying on the internet. Many develop have addictions to gaming.

    To anyone who has given it some thought, any sufficiently advanced technology usually affects both in good and bad ways. Its obvious that something that increases degrees of freedom in one direction will do so in others. Humans come in and align it.

    There's some social credit to gain by being cynical and by signalling this cynicism. In the current social dynamics - being cynical gives you an edge and makes you look savvy. The optimistic appear naive but the pessimists appear as if they truly understand the situation. But the optimists are usually correct in hindsight.

    We know how the internet turned out despite pessimists flagging potential problems with it. I know how AI will turn out. These kind of articles will be a dime a dozen and we will look at it the same way as we look at now at bygone internet-pessimists.

    This is response not just to this article, but a few others.

    • raincole 3 minutes ago
      I think you underestimate people's grievance with technology. If you make a poll my guess is more than 50% of people will say the world was a better place pre-social media.

      If the AI tech keeps going at the direction it's going now, more and more people will start believing the world would be better if the internet and computer had never been invented.

  • cowpig 40 minutes ago
    > I think it’s likely (at least in the short term) that we all pay the burden of increased fraud: higher credit card fees, higher insurance premiums, a less accurate court system, more dangerous roads, lower wages, and so on.

    I think the author is brushing against some larger system issues that are already in motion, and that the way AI is being rolled out are exacerbating, as opposed to a root cause of.

    There's a felony fraudster running the executive branch of the US, and it takes a lot of political resources to get someone elected president.

  • themafia 13 minutes ago
    > They also build secondary LLMs which double-check that the core LLM is not telling people how to build pipe bombs

    Such a fear mongering position. You can learn to build pipe bombs already. Take any chemical reaction that produces gas and heat and contain it. Congratulations, you have a pipe bomb.

    Meanwhile.. just.. ask an LLM if you can mix certain cleaning chemicals safely.

    > I see four moats that could prevent this from happening.

    Really? Because you just said:

    > human brains, which are biologically predisposed to acquire prosocial behavior

    You think you're going to constrain _human_ behavior by twiddling with the language models? This is foolishly naive to an extreme.

    If you put basic and well understood human considerations before corporate ones then reality is far easier to predict.

  • imbus 1 hour ago
    [dead]
  • throwway120385 2 hours ago
    At scale I think our society is slowly inching closer and closer to building HM.
    • nine_k 1 hour ago
      What is HM here?
      • throw4847285 35 minutes ago
        A Hidden Machine. That's right, a being that can cut, fly, surf, strength, and flash! Terrifying.
      • derektank 1 hour ago
        Maybe they meant AM (Allied Mastercomputer) from “I Have No Mouth, and I Must Scream“
      • zackmorris 1 hour ago
        Hacker Mews
        • throw4847285 36 minutes ago
          Sounds like a lovely co-op building, or perhaps a retirement community for aging hackers.
        • throwaway27448 1 hour ago
          Looksmaxxing really has gone mainstream huh
          • bitwize 52 minutes ago
            Thought it was all the Rust catgirls.
      • Sardtok 49 minutes ago
        Hennes & Mauritz is a Swedish clothing retailer.

        On a serious note, I think they meant TN, as in Torment Nexus, but I could be wrong.

  • dgfl 48 minutes ago
    The issue with most of these articles is that they seem to demonize the technology, and systematically use demeaning language about all of its facets. This one raises a lot of important points about LLMs, but the only real conclusion it seems to make is "LLMs are bad! We should never build them!". This is obviously unrealistic. The cat is out of the bag. And we're not _actually_ talking about nuclear weapons here. This technology is useful, and coding agents are just the first example of it. I can easily see a near future where everyone has a Jarvis-like secretary always available; it's only a cost and harness problem. And since this vision is very clear to most who have spent enough time with the latest agents, millions of people across the globe are trying to work towards this.

    I do think that safety is important. I'm particularly concerned about vulnerable people and sycophantic behavior. But I think it's better not to be a luddite. I will give a positively biased view because the article already presents a strongly negative stance. Two remarks:

    > Alignment is a Joke

    True, but for a different reason. Modern LLMs clearly don't have a strong sense of direction or intrinsic goals. That's perfect for what we need to do with them! But when a group of people aligns one to their own interest, they may imprint a stance which other groups may not like (which this article confusingly calls "unaligned model", even though it's perfectly aligned with its creators' intent). People unaligned with your values have always existed and will always exist. This is just another tool they can use. If they're truly against you, they'll develop it whether you want it or not. I guess I'm in the camp of people that have decided that those harmful capabilities are inevitable, as the article directly addresses.

    > LLMs change the cost balance for malicious attackers, enabling new scales of sophisticated, targeted security attacks, fraud, and harassment. Models can produce text and imagery that is difficult for humans to bear; I expect an increased burden to fall on moderators.

    What about the new scales of sophisticated defenses that they will enable? And for a simple solution to avoid the produced text and imagery: don't go online so much? We already all sort of agree that social media is bad for society. If we make it completely unusable, I think we will all have to gain for it. If digital stops having any value, perhaps we'll finally go back to valuing local communities and offline hobbies for children. What if this is our wakeup call?

    • throw4847285 37 minutes ago
      Thanks LLM!
      • eks391 1 minute ago
        Which LLMisms are you seeing in their post? Their grammar, word choice, thought flow, and markings all denote a fully human authorship to me, so confidently that I would say they likely didn't even consult an LLM.
      • dgfl 20 minutes ago
        lol. I did use a lot of short sentences, that’s my bad. But please read through [1] and compare my text onto it, it may enlighten you on how to actually spot llm writing.

        [1] https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing

        • throw4847285 10 minutes ago
          Oh no, I'm sorry to hear that.

          For the future, try to avoid prevaricating when you actually have a clear sense of what you want to argue. Instead of convincing me that you've weighed both options and found luddism wanting, you just come off as dishonest. If you think stridently, write stridently.

  • jazzpush2 2 hours ago
    Every one of these posts is immediately pushed to the front page, this one within 4 minutes.
    • aphyr 1 hour ago
      It's been weirdly uneven. Sections 1, 3, and 5 did well on HN; 2, 4, and 6 sank with essentially no trace. The distribution of views is presently:

      1. Introduction: 33,088 (https://news.ycombinator.com/item?id=47689648)

      2. Dynamics: 3,659 (https://news.ycombinator.com/item?id=47693678)

      3. Culture: 5,914 (https://news.ycombinator.com/item?id=47703528)

      4. Information Ecology: 777 (https://news.ycombinator.com/item?id=47718502)

      5. Annoyances: 7,020 (https://news.ycombinator.com/item?id=47730981)

      6. Psychological Hazards: 199 (https://news.ycombinator.com/item?id=47747936)

      Feedback from early readers was that the work was too large to digest in a single reading, so I split it up into a series of posts. I'm not entirely sure this was the right call; the sections I thought were the most interesting seem to have gotten much less attention than the introductory preliminaries.

      • simoncion 59 minutes ago
        I'm not sure that HN vote count is a good indicator of interest? HN alerted me to the existence of the intro post. I read the intro, noticed that it was one in an ongoing series, and have been checking your blog for new installments every few days.

        I suspect that if you'd not broken up the post into a series of smaller ones, the sorts of folks who are unwilling to read the whole thing as you post it section by section would have fed the entire post to an LLM to "summarize".

    • acdha 1 hour ago
      That’s unsurprising given the author’s long history in the tech community. A ton of people see that domain and upvote.
      • jazzpush2 1 hour ago
        Sure, but 4 front-page posts from the same url in 4 days surely sits at the tail of the distribution. (I guess they all capitalize on the same 'LLM-is-bad' sentiment).
        • zdragnar 1 hour ago
          It's also aphyr, who is incredibly popular. Take one very popular author, have him write a series of posts on the zeitgeist everyone can't help but talk about, and yes, the outcome is that his posts are extremely popular.

          I still remember his takedown of mongodb's claims with the call me maybe post years and years ago filling me with a good bit of awe.

          • macintux 1 hour ago
            When I worked for Basho, aphyr was highly respected by some of the smartest people I’d ever worked with. Definitely no slouch.
        • borski 1 hour ago
          It’s because it’s aphyr.

          If ‘tptacek posts a blog post, I bet it similarly does well, on average, because they’re a “known quantity” around these parts, for example.

    • tptacek 1 hour ago
      A statement broadly true of most things this author writes.
    • stronglikedan 1 hour ago
      that's just, like, how HN works. people post, people like, people upvote, people discuss
  • ibrahimhossain 1 hour ago
    Alignment feels like an arms race that favors whoever spends the most on RLHF and red teaming. If even friendly models keep leaking dangerous capabilities, the real moat might be making systems that are fundamentally limited rather than trying to patch every possible failure mode. Interesting piece.
  • conquera_ai 34 minutes ago
    Feels like we’re repeating classic distributed systems lessons: assume failure, constrain blast radiusand never trust components that can’t explain themselves reliably
    • ibrahimhossain 20 minutes ago
      Exactly assuming failure and constraining the blast radius feels like the only reliable path when the models themselves are black boxes. Patch based alignment starts looking fragile pretty quickly