Career educator, now Lead Media Analyst at MIRI
tanagrabeast
Karma: 2,351
Other MIRI staff report that the book helped them fit the whole argument in their head better, or that it made sharp some intuitions they had that were previously vague. So you might get something out of it even if you’ve been around a while.
Can confirm! I’ve followed this stuff for forever, but always felt at the edge of my technical depth when it came to alignment. It wasn’t until I read an early draft of this book a year ago that I felt like I could trace a continuous, solid line from “superintelligence grown by a blind process...” to “...develops weird internal drives we could not have anticipated”. Before, I was like, “We don’t have justifiable confidence that we can make something that reflects our values, especially over the long haul,” and now I’m like, “Oh, you can’t get there from here. Clear as day.”
As for why this spells disaster if anyone builds it, I didn’t need any new lessons, but they are here, and they are chilling—even for someone who was already convinced we were in trouble.
Having played some small part in helping this book come together, I would like to attest to the sheer amount of iteration it has gone through over the last year. Nate and co. have been relentlessly paring and grinding this text ever closer to the kind of accessibility that won’t just help individuals understand why we must act, but will make them feel like their neighbors and political leaders can understand it, too. I think that last part counts for a lot.
The book is also pretty engaging.
The pitch I suggested we share with our friends and allies is this:
If you’ve been waiting for a book that can explain the technical roots of the problem in terms your representative and your mother can both understand, this is the one. This is the grounded, no-nonsense primer on why superintelligence built blindly via gradient descent will predictably develop human-incompatible drives; on why humanity cannot hope to endure if the fatal, invisible threshold is crossed; and on what it will take to survive the coming years and decades.
Bestseller algorithms are secret and shifty, but hardcover is generally believed to count a little more. And as for overall impact, if either format is good for you, a hardcover preorder helps more because it encourages the publisher to print a bigger initial run of physical copies, which can get pumped into stores and onto shelves where people will see them.
¿Por qué no los dos? Different problems, both worth solving, and not especially connected, in my model.
“Start with an easier problem with similar dynamics to build up expertise” sounds to me like a fully general argument for working on anything other than the actual thing one cares about; it’s a procrastinator’s mantra, in most cases. There are plenty of exceptions, but I don’t think this is one of them.
If I thought there could only ever be one concerted attempt at an AGI ban, and that this was much less likely to fail if the promoters had previously solved some other kind of ban, I could sort of see the logic. But that doesn’t match my model. I’m not even convinced that banning bio gain-of-function is harder, having not tried it, and noticing that I’m confused about the entrenched interests that must exist for it to have continued as long as it has.
I also could see this approach backfiring, as the AGI ban promoters would be seen as the “guys who like to ban stuff” people, and as people who were more concerned about gain-of-function than AI catastrophe.
Glad you got something out of the post! I recognize and appreciate your generous action, and will DM you with regards to your request.
MIRI has its monthly newsletters, though I can tell that’s not quite what you want. I predict (medium confidence) that we will be upping our level of active coordination with allied orgs and individuals on upcoming projects once we ship some of our current ones. I believe you already have channels to people at MIRI, but feel free to DM me if you want to chat.
Me too! I put the AI problem in the broader class of topics where apathy serves as a dual defense mechanism — not just against needing to expend effort and resources, but against emotional discomfort. You can see the same dual barrier when promoting charitable causes aimed at reducing human misery, or when teaching a subject to students who have really struggled with it in the past.
As a teacher, I attacked both of those roots more deliberately as I grew, trying hard to not make my class feel like work most days while building an atmosphere of low-stakes experimentation where failure could be fun rather than painful. (An example of what success looked like: students taking the risk of trying the more advanced writing approaches I modeled instead of endlessly rewriting the same basic meta essay they had learned in middle school.)
One tactic for eroding defensive apathy is therapeutic empathy. You see this both in many good teachers and (I imagine) relationship counselors. It’s much harder in writing, though I suppose I did a little bit of that in this post when I talked about how the reader and I have probably both felt the pull of Apathy with regards to the AI problem. I think empathy works partly because it builds a human connection, and partly because it brings the feared pain to the surface, where we find (with the help of that human connection) that it can be endured, freeing us to work on the problem that accompanies it.
Whether and how to use authentic human connections in our communications is a topic of ongoing research and debate at MIRI. It has obvious problems with regards to scientific respectability, as there’s this sense in intellectual culture that it’s impossible to be impartial about anything one has feelings about.
And sure, the science itself should be dispassionate. The universe doesn’t care how we feel about it, and our emotions will try to keep us from learning things we don’t want to be true.
But when communicating our findings? To the extent that our task is two-pronged: (1) communicating the truth as we understand it and (2) eliciting a global response to address it, I suspect we will need some human warmth and connection in the second prong even as we continue to avoid it in the first. Apathy loves the cold.
From chatting with those peak students during the experiment, I think their experience is more like being in a cafeteria abuzz with the voices of friends and acquaintances. At some point, you’re not even trying to follow every conversation, but are just maintaining some vague awareness of the conversations that are taking place and jumping in when you feel like it. People can and do think about other things in a noisy cafeteria. Some even read books! The brain can filter out a constant buzz. It’s just wind blowing through the trees.
The upper middle zone where it’s still possible to try to follow everything (and maybe even reply) looked like more of an attention trap, and was where I was more likely to find that handful of students I already knew had a problem. The FOMO is probably more distracting than the notifications themselves.
They should not have been counting pull notifications, as they were instructed to not engage with their phones during the experiment except to maybe see what caused a vibration or ding. I don’t think students think of pull notifications as real notifications the way we were using the word. They were logging the notifications they could notice while their phone flat was flat on their desk not being touched.
No. Everyone seemed to know what they were, because they all claimed to know someone who uses them. But I don’t recall anyone ever admitting to being such a someone. I sense there’s a bit of a stigma around them.
It is credible that eliminating all preventable distractions (phones, earbuds, etc.) wouldn’t improve learning much. As a teen, I bet you were distracted during class by all sorts of things contained entirely within your head. I know I was!
There’s a somewhat stronger case that video games and social media have given students more things to be preoccupied about even if you make these things inaccessible during class. But I also think that just being a hormonal teen is often distracting enough to fill in any attention vacancies faster than the median lesson can.
This is important work.
One suggested tweak: I notice this document starts leaning on the term “loss” in section 4.2 but doesn’t tell the reader what that means in this context until 4.3
Something similar happens with the concept of “weights”, first used in section 1.3, but only sort-of-explained later, in 4.2.
Speaking of weights, I notice myself genuinely confused in section 5.2, and I’m not sure if it’s a problem with the wording or with my current mental model (which is only semi-technical). A quoted forecast reads:
“GPT-2030’s copies can share knowledge due to having identical model weights, allowing for rapid parallel learning: I estimate 2,500 human-equivalent years of learning in 1 day.”
Wouldn’t the model doing the sharing have, by definition, different weights than the recipient? (How is a model’s “knowledge” stored if not in the weights? ) My best guess: shareable “knowledge” would take the form of vectors over the models’ common foundational base weights—which should work as long as there hasn’t been too much other divergence since the fork. Is that right? And if so, is there some reason this is a forecast capability and not a current one?
My apologies for challenging the premise, but I don’t understand how anyone could hope to be “convinced” that humanity isn’t doomed by AGI unless they’re in possession of a provably safe design that they have high confidence of being able to implement ahead of any rivals.
Put aside all of the assumptions you think the pessimists are making and simply ask whether humanity knows how to make a mind that will share our values. It it does, please tell us how. If it doesn’t, then accept that any AGI we make is, by default, alien—and building an AGI is like opening a random portal to invite an alien mind to come play with us.
What is your prior for alien intelligence playing nice with humanity—or for humanity being able to defeat it? I don’t think it’s wrong to say we’re not automatically doomed. But let’s suppose we open a portal and it turns out ok: We share tea and cookies with the alien, or we blow its brains out. Whatever. What’s to stop humanity from rolling the dice on another random portal? And another? Unless we just happen to stumble on a friendly alien that will also prevent all new portals, we should expect to eventually summon something we can’t handle.
Feel free to place wagers on whether humanity can figure out alignment before getting a bad roll. You might decide you like your odds! But don’t confuse a wager with a solution.
This is about where I’m at, as well. I’ve been wrestling with the idea of starting a run myself, but one of my qualifying traits (I teach creative writing) also means I work full time and have little hope of beating out ten people who don’t. So much the better, I say, so long as the work gets done well and gets done soon...
...but if, eight months from now, much of the budget is still on the table because of quality issues, it may be because people me sat on our hands.
Hopefully, someone will emerge early to work around this issue, if it turns out to be one. I, for one, would love to be able to turn in a sample and then be offered a credible good-faith assurance that if my run is completed at same quality by such and such date, a payment of x will be earned. But as it stands, the deadline is “whenever that fastest mover(s) get there”. Who knows when that will be? Any emergent executive candidate making me a deal might be made a liar by a rival who beats them to the jackpot.
My questions are mostly about the player side, and about how deeply the DM should model the player:
Should the player be assumed to be implicitly collaborating towards a coherent, meaningful narrative, as is necessary for a long-lived TTRPG? Or might they be the type of player you often find in AI Dungeon who tries to murder and/or have sex with everything in sight?
Should players ever try to steer the story in a genre-breaking direction, like erotica or murder-hobo power fantasy? Should DMs resist these efforts or play along? If the latter, should the DM go a step further to actively intuit what this particular player would like to see happen?
Should players provide input that might be more sweeping than usable in narrative? (e.g. Take over the world!) If so, on what sort of level should the DM respond to these?
Should players be assumed to be ready to end the narrative at the ~1,000-step point?
I don’t see as much disagreement between us as you might be thinking. Precisely because I agree with your numbered points 1 and 2, I suggested it could be beneficial to compress most of our 12 years of math instruction down to a more intensive 2-3 years. That doesn’t mean we couldn’t instill useful basic arithmetic in lower grades. If we chose a smaller set of core basics, it could be quite practical to retain them over long summers and breaks—at least for the students who stay in our system for the long haul.
I’m also glad you brought up the fact that spaced repetition doesn’t have to involve software. I should have done more to remind readers of this. I weave the spacing and testing effects into the fabric of my course in many ways that have nothing to do with software.
Carefully engineered homework assignments are great if you have motivated students. Take-home SRS could even work for that. Those students are usually fine, though. It’s the apathetic middle I have to fight for, and they won’t do homework regardless of how I try to incentivize it.
Moreover, I don’t feel good about assigning to students who would hate to do it. School is already prison for those kids. I don’t want to send prison home with them. As both a child and a parent, I have been too familiar with the toxic effects homework—especially math homework—can have on family relationships. Let kids have a light at the end of the daily tunnel, I say.
Is homework vital to a successful math program? I don’t know. But I’m glad I don’t teach math.
Did you get IRB approval for these human studies on children?
I’m not sure which is more absurd: the IRB approval process or the very idea of high school. I’ve often asked people to consider a thought experiment where everyone on Earth suddenly forgets that our educational system as we know it ever existed. Would we really reinvent it just like it is now? Hearing how it worked, would we scream in terror and cancel anyone who had taken part? (Status quo bias much?)When I was studying stand-up comedy, I actually developed a bit in which I play-acted a researcher proposing high school to an ethics board. It went like this:
RESEARCHER: “I was thinking we could stick 35 sleep-deprived teenagers in a room for an hour and expose them to academic stimuli. After that, we’ll do some tests on them.”
BOARD: “I see. Tell me more about your subjects.”
RESEARCHER: “Well, they’re minors, obviously.”
BOARD: “Okay…”
RESEARCHER: “And most of them will be enrolled against their will.”
BOARD: “And how long will you need them?”
RESEARCHER: “6 sessions a day for four years.”
BOARD: “Wait, hold on. Sample size? How many kids are we talking about, here?”
RESEARCHER: “All of them.”
BOARD: (mutterings among themselves) “Well, it sounds like everything is in order...”
Are you familiar with Direct Instruction, which is reminiscent of the Mennonite school?
Someone (probably on LW) pointed me to Direct Instruction a few years back, so yes, I’m acquainted with it. Because of the emphasis on staying fully reviewed on all relevant prior knowledge, I saw it as having obvious promise for technical subjects like math, in the hands of the right teacher. I was less convinced it made a good fit elsewhere, perceiving (perhaps unfairly—I didn’t dig too deeply) some big negative trade-offs:
Like with my whole-class Anki, it seems heavily reliant on the teacher’s high-energy snake-charmer charisma. This makes it difficult to sustain for much of a class period and demands a great deal from a teacher who tries to do it all day long, day after day. This also makes it difficult to broadly among teachers with different personalities.
It sounds brittle with regards to roster variance. Specifically, it seems pretty insistent on having everyone in the room up to speed. With careful tracking/grouping of students, this can be achieved, but in practice, kids move in to your school part way through the year and aren’t on the same page. Or you only have the one or two teachers for that grade level math, so the slowest kids are in the same boat as the sharpest. I would think that one or two stragglers would grind the class to a halt, and that this would be statistically inevitable in larger classes. (I don’t know if this makes DI math worse than the status quo, where plenty of students are fall behind and get lost, but with less fanfare and hold-up for everyone else.)
Have you ever tried SRS for muscle memory?
No. I’m not seeing how that would work, or how that would be relevant to what I do, but I’m certainly curious. Do you have examples?
Do you think that instinctive drive to listen to experts “talk shop” applies to apathetic students, though?
That’s definitely the right question. If you and another expert leap straight into fluent French, no, I don’t think your apathetic students will try to keep up—especially if they are early beginners. More helpful might be a Franc-lish hybrid conversation where you swap stories of embarrassing errors and insights largely in English while sprinkling in French words and expressions, reenacting parts of colorful encounters from your combined French-speaking experience.
I also think one of the difficulties in modeling language fluency is that the whole point of being fluent is to not need to think about the language, but to simply think in it, so I’m not sure what your vocalized monologue would be about...unless...
Ok, here’s a thought: I and the other motivated folks I learned Spanish with sometimes found ourselves slipping into a Spanglish patois outside of class where we spoke English with Spanish syntax. It felt like silly play at the time, but I now think it was an instinctive intermediate step to thinking in that language.
“It makes rain.” (It’s raining.)
“To me pleases the rain!” (I like rain.)
Perhaps you could try fostering a Franc-lish dialect in your classes by thinking out loud in that style and inviting others to join you in banter, patiently nudging them to get the grammar right instead of just talking like Yoda. From there, substituting actual French with increasing frequency could feel very natural.
You may not have immersive environments, but I imagine you’ll be creating simulated immersion: play-acting situations that give you a chance to think out loud as though you are navigating the moment for real. (Example: Going to the produce section of the store and seeing what looks good, what you could make with it, etc.) How much of that you should do in English, Frank-lish syntactic patois, or French will probably be something you will develop an expert instinct for as you become skilled at reading the room. Along the way, developing an entertaining stage presence for this play-acting would give you a powerful weapon against apathy.
Yes, yes… and you would be randomly involving students in your little improvised plays, assigning them roles, keeping them on their toes, making the non-participants want to get called on.
Yep, it sounds pretty awesome from the comfort of my not-having-to-teach-French perch :)
Oh, wow. Yes. That. Looks like there’s another book I don’t need to write.
The fact that the concept was so fleshed out thirty years ago kind of pisses me off. My teacher training was so the opposite of that (a bunch of student group work nonsense). And I’m not finding apprenticeship familiar to new teachers currently, though strong veterans often seem to have at least a half-baked version they’ve derived from experience. I get a lot of wide-eyed “Yes!” when I share it with them.
Experts talking shop with other experts is one of my favorite finds when I study!
During my dive into stand-up comedy, I came across this video of some top comedians talking shop. Especially from about the 30 minute mark, when they seem less concerned with entertaining their audience, they get into some juicy minutiae of why a joke might work or not. It really expanded my thinking on the subject.
Are such chats more insightful than an expert teacher would be in a lesson on that same topic? Not necessarily. But you might not find a skilled teacher ever teaching a lesson on that exact topic. I think humans are naturally primed to closely observe expert-expert chats for a few reasons:
• Social proof. We instinctively want to be able to talk like the experts do so we can blend in with them. So we listen carefully to how they talk.
• Authenticity. If this is what experts actually talk about, we feel like it must really matter. It’s not just the lesson of the day.
• The overhearing effect. This is a term I’m making up, but I’ve found it to be an important one exploited by storytellers. We naturally want to deduce the context of overheard language, so we listen extra carefully, trying to fill in the blanks. I suspect this is down to humans’ highly evolved appetite for gossip. The fact that the experts aren’t talking to us is essential for exploiting this effect.
Although… I find that an expert talking to himself, seemingly unguarded, seemingly without conscious awareness that he is being overhead… can also trigger the overhearing effect. When I model a skill to my students, I try to verbalize my inner monologue in a way that will be intriguing to overhear and carry that essential whiff of authenticity.
I’m not sure what expert self-talk looks like in foreign language instruction, but I would be interested to find out. (Any ideas?)
But from my time becoming a reasonably fluent Spanish speaker (since lost), I can describe a few language dimensions I found interesting but neglected by all but the nerdiest supplemental books.
Sentence-level inflection patterns vary, and it helps to be aware of them. For instance, the musicality of typical question sentence is different in American English than in Castilian Spanish. If you can pick up on the melody earlier in the sentence, you can better contextualize what is being
saidasked.The way speakers in different languages produce what seems, on the surface, to be identical phonemes, can be quite different, and understanding this is essential to actually sounding like a native. There can be hours of fun trying to practice a Castilian ‘toh’ sound (as in toma), with its thicker top-front palette tongue contact, vs. the American English cousin equivalent (as in tomato).
Native speakers of language A learning language B often end up predictably adopting many of the same idioms and juicy words from language B into their language A conversations with each other, and they find themselves saying or thinking in those patterns even when their brains are mostly running language A. It could be fun to introduce some of these to novices and make it part of the language A classroom slang—a kind of introduction to thinking in language B.
Looking back, I was surprised by the (unflattering, in my opinion) degree to which LWers saw this data as strong confirmation of their hypotheses about phones being the source of the ills they see in our schools and our young.
I thought it was much more of a mixed picture despite—or perhaps because—the numbers were significantly higher than I, a veteran teacher, had expected: If I had been told the previous summer that “next year, we’re giving you a cohort of students with this phone usage profile”, I might have braced for a crop of students that came off as especially scatterbrained. But overall, these students—even most of the high-receivers on the left of the chart—seemed to me about as well adjusted as teenagers of yesteryear and, I would say, better adjusted than the students I remember from the early days of pre-smartphone texting going mainstream. (There’s probably a thread to pick there about messages being more alive among teens with vibrant in-person social lives.)
The update for me from this experient was more like, “Huh, I guess notifications are just sort of a background hum of modern existence now. If adults can largely manage to be productive in this environment, maybe I shouldn’t be so surprised that young people who grew up in it can, too.” This it’s-just-a-different-world-now feeling was accentuated by the fact that the school itself was now behind a high fraction of notifications for the median student.
I’m wary of anecdotal clickbait-ish tales of “This school locked away phones, and something miraculous happened” because I would expect a honeymoon phase where an engaging teacher can step in and fill any real attention void before the more traditonal expressions of mental exhaustion and apathy can assert themselves.
To change my mind about this, I would need to linger in the counterfactual world where the students I had known from the start to be actually, problematically distracted by their devices were permanently deprived of them. My mental simulation of that world sees about half of these students gradually becoming traditional excessively-chatty-in-class types, and most of the rest daydreaming, sleeping, or checking out in some other way. To the extent that a busy life on a phone keeps a teen’s mind active, it might actually be an improvement?