I guess the problem is that there is a very specious rebuttal that says that plato objected to writing, people flipped out about calculator usage, and computers etc and we are all fine. But we...
I guess the problem is that there is a very specious rebuttal that says that plato objected to writing, people flipped out about calculator usage, and computers etc and we are all fine.
But we aren't fine. I am deeply embedded in high tech stuff and it makes me exceptionally uncomfortable how specialized our thinking has become and how disconnected we are from almost everything we use. LLMs are the next rung on the ladder of delegation, and instead of saying "it's fine, look at all the runs we've already climbed, this the natural progression of things" I feel like we should be saying "we should maybe descend a little?" It is nuts how many things your average Joe interacts with that they have zero understanding of, and it's allowing for all kinds of nasty consequences (algorithmic feeds, dynamic black box pricing with real consequences, i.e. insurance). AI labs can't really explain why models do what they do, and while they are chipping away at the problem, we are sprinting towards using them everywhere.
I feel for the young'uns. I have no idea how they will navigate this without the grace and generosity that was afforded to me when I was a teen to make mistakes without something ubiquitous and available and way smarter than me whispering answers in my ear.
But just look at the history of human civilization and all technological advancement for examples of this over thousands of years. People write without knowing how to make paper, pens, and colored...
But just look at the history of human civilization and all technological advancement for examples of this over thousands of years. People write without knowing how to make paper, pens, and colored inks. Artists paint without knowing how to make their own paints, brushes, canvas, sthretchers and frames. Books are made by authora who don't know how to print the pages, bind them together, or work a leather cover. Bronze workers create things from bronze without knowing how to mine it, or smelt it. We drive our cars without having worked the facotry lines, or the oil rigs, or the refineries.
There is literally too much to know in the world to become an expert in such a wide range in a single human lifetime. The only way to continue pushing the boundaries of human knowledge is through specialization.
I understand the argument that AI may make it so that no human is the specialist any longer, so who do you go to when you want a bespoke frame made as an artist and no one knows how to make them any longer. That is a problem that already happens, though. There are art forms and traditions in Japan and Korea that only a select few people have been trained how to do: traditional ketsugi and soju come to mind for me. I'm not sure I buy the argument that AI will cause this to happen more.
Right: and this is exactly the genre of argument that I say I don't agree with in my opening paragraph. I am saying that all of it is bad. It's also bad by degrees. It is not hard, even for...
Right: and this is exactly the genre of argument that I say I don't agree with in my opening paragraph. I am saying that all of it is bad. It's also bad by degrees. It is not hard, even for someone with middling intelligence to follow that many liquids stain, and that many surfaces can be stained. This the basis of the preservation of human knowledge is learnable with much effort. It's not going to be as good as modern inks, or industrially produced paper, but it can sustain knowledge transfer. Try explaining even an 8bit full adder and the physical circuitry for that to most people, and that's not even the beginning of doing anything useful with a computer.
LLMs and other sufficiently complex systems are bad in the sense that responsibility becomes diffuse to the point that no one will take accountability. LLMs are not special in as much as they lie on a spectrum. We've all effectively been told "computer says no" for years. Before computers decision makers hid behind their lackeys to avoid accountability. LLMs are special in the sense that they currently represent for the tip the same shitty spear, and a step change in capability for continuing this trend. Where I think they differ from what we have seen before is the function that is being outsourced: thinking itself. In this regard I do think the medium is the message and the medium is having a profoundly negative effect on people.
Just as a sidebar, as I'm intruding on your conversation, and this reminded me: I've been feeling less concerned about that recently? A lot of people already go "brain off" in order to survive the...
Just as a sidebar, as I'm intruding on your conversation, and this reminded me:
Where I think they differ from what we have seen before is the function that is being outsourced: thinking itself.
I've been feeling less concerned about that recently? A lot of people already go "brain off" in order to survive the modern world (eg buy the first thing they see advertised that meets a need, do zero research before making an investment, follow trends in their community uncritically, etc.); the LLMs seem to just homogenize that. Also if you've ever seen a DIYer try to fix their house, you may have a different opinion on the inherent value of insourcing thought š
A thought I've had is that a major critique re. climate conscious consumption habits is that the friction is too much for people: they can't plan around a slight inconvenience, do research to find a less damaging alternative, sort their recycling, etc. because they have too much going on. Outsourcing thought (so to speak) seems like a decent means of addressing that, at least.
What's missing in this argument is that calculators do not hallucinate. It's true that fewer people now are able to do addition/multiplication in their heads; but with a calculator they can still...
What's missing in this argument is that calculators do not hallucinate.
It's true that fewer people now are able to do addition/multiplication in their heads; but with a calculator they can still do the arithmetic reliably, and that will remain the case for as long as they have a calculator available to them.
LLMs are not like that. LLM output requires skepticism.
I think paper+ink sits somewhere in the middle. If you just wanna write someone a letter, then paper+ink is sufficiently preserving. But for historical records that are supposed to last centuries, it's a different story. Nevertheless, at least we can tell when the ink has run or the paper has torn; it won't suddenly say something different but with the same level of plausibility. LLMs don't just run the ink, they write something different/wrong and we can't tell.
100% agreed. I'm a little hopeful, though? LLMs present a profound capacity for teaching, beyond what's ever been accessible before. Exceptional people have always been able to self-teach from...
I am deeply embedded in high tech stuff and it makes me exceptionally uncomfortable how specialized our thinking has become and how disconnected we are from almost everything we use.
100% agreed. I'm a little hopeful, though?
LLMs present a profound capacity for teaching, beyond what's ever been accessible before. Exceptional people have always been able to self-teach from books or pure experimentation, but the rest of us could really do with a hand (or two) understanding all those rungs society pulled itself up on. At present, we're given neither time to learn nor appropriate resources to learn from, but -- deployed correctly -- LLMs should hopefully be able to address the latter constraint.
Very few organizations (beyond e.g. Khan academy's attempts) are genuinely trying to make a teaching tool, though. The real money is in a magic box that replaces human competency, not a tool for creating a more capable and educated citizenry -- Anthropic isn't being valued at a trillion USD for teaching Joe Average how retirement funds work, or that the krebs cycle exists. But still, I could (realistically) imagine a future where people are more educated on the facts of reality, rather than whatever was pitched to them in their most recent Facebook ad.
(tbh though the only way that's happening is if a Chinese or European model goes for it)
AI labs can't really explain why models do what they do [...]
Ehhhhh I dunno. I feel like we're still working on the formalizations, but since this is very much a new frontier in computer algorithms, I'm at least cutting them a lot of slack. The discipline of CS has been very heavy on the "Computer" and extremely light on the "Science" for the entirety of its existence, so people are understandably confused by the prospect of any process that requires experimentation, modelling, and rigor to make progress in (i.e. science). We're doing decently imo from that perspective! We used a lot of discoveries before fully understanding the physics underlying them, so this doesn't seem much different.
The problem I see with LLMs for teaching though is I think it takes an exceptional individual to wield an LLM and use it to increase their own understanding instead of letting it carry them. Ofc...
The problem I see with LLMs for teaching though is I think it takes an exceptional individual to wield an LLM and use it to increase their own understanding instead of letting it carry them. Ofc someone could design an LLM experience/ harness explicitly around the learning loop so that it would take significant energy to hold it wrong, but that's not where the money is so I find it unlikely to be as widespread on the same timeline.
I say this as selfishly I use LLMs to do my work more efficiently and pull the ladder up behind me. We aren't hiring grads. We aren't hiring apprentices, we aren't hiring juniors. I don't know who is supposed to replace me when I go. Weird to see my generation (elder millennial) gap those below them just as hard as the ones they were complaining about 10 years prior.
EDIT: on discovering the mechanisms via which LLMs work, I agree there is progress there, I'm just not that bullish on what the possibilities are, and certainly don't think that we will bind our usage of them by our understanding of the underlying mechanics.
I think you might be looking a this slightly askew though. The way you've phrased it here implies that the person who's learning must also understand how to be an effective teacher, and...
The problem I see with LLMs for teaching though is I think it takes an exceptional individual to wield an LLM and use it to increase their own understanding instead of letting it carry them.
I think you might be looking a this slightly askew though. The way you've phrased it here implies that the person who's learning must also understand how to be an effective teacher, and furthermore must also self-educate themselves on how to teach with an LLM specifically.
I work in higher education (as staff), and I've seen some of the ways faculty are beginning to experiment with LLMs in a teaching context. Some of them are kinda ham-fisted and naieve, but some have interesting potential. I think the educational value of LLMs is that they can engage with students in a way no other technology can possibly do. They are the closest we have ever come to being able to give every single student a 1:1 tutor that specializes in every single subject they have.
There are at least a dozen examples that I'm aware of off the top of my head of faculty attempting in various ways to build a customized LLM that will effectively teach students specific concepts. They are refining and specializing the technology to be an effective teaching tool. Where exactly this ends up and how effective it will be is still a wide open question, but I do agree with @kacey that there is enormous potential.
The reason I am phrasing it this way is because no matter what educators produce, the commercial/freemium models and harnesses are not going away. This is like sending alcoholics to hangout at the...
The reason I am phrasing it this way is because no matter what educators produce, the commercial/freemium models and harnesses are not going away. This is like sending alcoholics to hangout at the bar. It will take tremendous force of will to restrict oneself to making life "harder" for themselves now for a payoff later. I cite the entirety of the brain-rotting internet as my evidence.
I think it's great though that schools are looking at building better tools though!
Oh, yeah, agreed that that's not where the money is (although Khanmingo does exist as a counterexample). I don't think it'd take a hundred million dollars to get this right, though, and what we've...
Ofc someone could design an LLM experience/ harness explicitly around the learning loop so that it would take significant energy to hold it wrong, but that's not where the money is so I find it unlikely to be as widespread on the same timeline.
Oh, yeah, agreed that that's not where the money is (although Khanmingo does exist as a counterexample). I don't think it'd take a hundred million dollars to get this right, though, and what we've seen from open weight model development/the recent leaps in frontier models seems to indicate that better harnesses + training on smaller models is the key to improved benchmark performance anyhow? As it'd ultimately only take one model and harness to nail this UX, then minor updates over time to shove updated info into the weights (either with retraining from early checkpoints or something like ROME to edit memories).
(sidebar: while I still had throughput to think about this, it seemed like distributed training of a ~30b equivalent model might be possible. EleutherAI has been trying to do this for a while, too. Widespread training of small, cheap models might adjust the prior for this prediction, too)
I say this as selfishly I use LLMs to do my work more efficiently and pull the ladder up behind me. We aren't hiring grads. We aren't hiring apprentices, we aren't hiring juniors. I don't know who is supposed to replace me when I go. Weird to see my generation (elder millennial) gap those below them just as hard as the ones they were complaining about 10 years prior.
Hah, fair. You're definitely ahead of me by nearly a full generation, but those're many of the reasons I bailed on the industry. I was always pushing to hire on more interns and expand the breadth of experiences we'd accept into the company, but management hated it.
I wouldn't worry about it. It feels selfish in the moment, but putting your own mask on first seems like generally good advice. I was beating myself up this morning about how I spent a large chunk of my life's earnings, since if only I'd made some different choices, I might've had a chance to build some at-cost housing for friends who dearly need it. But I don't have a crystal ball, and neither do you -- maybe investing in yourself now will present you with an even greater opportunity in the future to contribute, that otherwise you'd be unable to access.
on discovering the mechanisms via which LLMs work [...]
Mmhm, understood. Honestly I'm just glad that I don't need to talk someone down from blindly following the Anthropic's gospel on circuits š
It's funny, as someone from a third world country, I worry about the exact opposite. LLMs will only exacerbate the current global inequality. Someone who can comfortably spend 300 USD on tokens...
if anyone can write papers and proposals and code as fluently as he could, then people like him lose their competitive edge.
It's funny, as someone from a third world country, I worry about the exact opposite. LLMs will only exacerbate the current global inequality. Someone who can comfortably spend 300 USD on tokens will always have an advantage over someone for whom 20 USD is a tenth or even a twentieth of their income, even if their level of understanding is the same.
Edit: I realized that I did the math wrong and didnāt mean to use such low numbers. I recently met a developer who makes 500 USD a month and is subscribed to the 100 USD ChatGPT tier. Somehow, in my mind, āspending a substantial amount of your incomeā automatically translated to the cheapest subscriptions. Regardless, my original point still stands.
That's been on my mind a lot too. I'm pursuing study in pure math, which as a field has always prided itself on being uniquely accessible in the sense that all you need to do math research is pen...
That's been on my mind a lot too. I'm pursuing study in pure math, which as a field has always prided itself on being uniquely accessible in the sense that all you need to do math research is pen and paper. It's seeming like lately you also may need a GPT Pro subscription...
Totally valid, I've spent a lot of time looking at how to increase "value per token" ratio. While it's a way to approach it, most the effort seems to be on "accuracy per request" because accuracy...
Totally valid, I've spent a lot of time looking at how to increase "value per token" ratio. While it's a way to approach it, most the effort seems to be on "accuracy per request" because accuracy needs to be higher most the time to have actual vs perceived value since hallucination rates are still high and can be deceptive.
Trying Deepseek v4 Pro with a token-efficient harness called Reasonix (built by the same company) is great because you can go really, really far on no money, but it creeps up often enough that it might be cutting corners and not giving accurate information and that requires MORE tokens to verify claims (go slow: measure twice, cut once) and so you lose token efficiency for confidence in accuracy. It cost me time and pre-paid usage, so there's a higher cost at the end of the day.
Between this and smaller local models, I can only hope more accessibility finds its way into non-frontier labs as a focus so edge devices (phones, personal machines) can achieve value that matters to those individuals.
I think we will be seeing more price competition soon. Open weights models are getting pretty good, and so are cheaper proprietary models. I sometimes use Luna instead of Terra and rarely try Sol...
I think we will be seeing more price competition soon. Open weights models are getting pretty good, and so are cheaper proprietary models. I sometimes use Luna instead of Terra and rarely try Sol or Opus.
In a bizarre and disturbing twist, it is convincingly argued by https://boxobark.ing/3mj42airv3s2o that this entire blog post may itself be written by AI. Given that I thought I could pretty...
In a bizarre and disturbing twist, it is convincingly argued by https://boxobark.ing/3mj42airv3s2o that this entire blog post may itself be written by AI. Given that I thought I could pretty reliably detect AI writing, and that I thought the blog post in question was generally cogent and interesting, I'm not really sure how to react. At the very least, it would be remiss of me not to include this additional context for any readers later on in this increasingly macabre world.
I spend an embarrassing amount of time talking to Claude and the whole essay screamed Claude to me. I didn't want to say anything because it felt like turning the discussion into a bun-flight...
I spend an embarrassing amount of time talking to Claude and the whole essay screamed Claude to me. I didn't want to say anything because it felt like turning the discussion into a bun-flight about authorship was less interesting that the discussion about the underlying arguments themselves.
I've seen the original post floating around but wasn't really interested. The idea that LLMs can contribute to skill athrophy, or not building skills in the first place, is low hanging fruit...
I've seen the original post floating around but wasn't really interested. The idea that LLMs can contribute to skill athrophy, or not building skills in the first place, is low hanging fruit that's been discussed from every angle over the last couple years.
But after I saw your comment I gave it a skim to see if it was LLM generated and, yes, it definitely is. The author gave it at least one human editing pass after generation which makes it a bit harder to detect. And they did a better job than many do of editing, but as always happens the human editing gets progressively lazier as you go.
If you wonder if something is AI written, ignore the first two paragraphs. They'll often be the only human written part, or the most carefully post edited. Instead start halfway through.
Anyway the tells in the post are so numerous and frequent that anyone who's used frontier agents in any volume will recognize the tropes.
I imagine the "author" convinced themselves they were being profoundly ironic by using AI to write the post. But I suspect that really they're just lazy and cynical.
In the context of the point their LLM is trying to make for them: The prompter robbing themselves of the opportunity to develop skills and deep mental models around writing. They're also missing out on the rewarding dynamic created by respecting your audience.
Its that obfuscation part that disappoints me the most, and its pretty damning to me. Mainly because it goes against the entire premise of the author's intent, a very Bob move. So does the author...
Its that obfuscation part that disappoints me the most, and its pretty damning to me. Mainly because it goes against the entire premise of the author's intent, a very Bob move. So does the author not really believe what they write, or do they make exceptions for themselves because its just that easy, and they don't want the backlash?
In any case you know they will update claude's memory to not ever touch quotes going forward.
Thank you for including it, I have had a quick skim, and I'm torn about whether I think it is AI generated text or not. I don't think I am that good at detecting AI text (or at least I'm not...
Thank you for including it, I have had a quick skim, and I'm torn about whether I think it is AI generated text or not. I don't think I am that good at detecting AI text (or at least I'm not according to the the wikipedia article on it), I will see if the author writes a rebuttal (maybe they already have, but I don't have time to search for it atm).
Detecting LLMisms is something you learn from practice and you get better at it. There are patterns that are invisible at first, but you start seeing them after repeated exposure. And I wonder if...
Detecting LLMisms is something you learn from practice and you get better at it. There are patterns that are invisible at first, but you start seeing them after repeated exposure. And I wonder if there are more patterns that I donāt see yet?
I enjoyed this read and I think it sums up my thoughts about AI: if you donāt do the grunt work you canāt understand what you are doing. If you canāt understand it, then you canāt check it, and...
I enjoyed this read and I think it sums up my thoughts about AI: if you donāt do the grunt work you canāt understand what you are doing. If you canāt understand it, then you canāt check it, and you lose half the reason of doing it in the first place (that deep understanding).
I tend to avoid LLMs in the same way I avoid using th car for trips I can walk- I donāt want those muscles to atrophy, and learning to struggle sometimes makes you strong and competent.
Thanks for sharing. This is exactly my biggest issue with AI. I'm grumpy about all sorts of other issues that it brings, but this is the one that can't really be "fixed" in my mind. I'm in an...
Thanks for sharing. This is exactly my biggest issue with AI. I'm grumpy about all sorts of other issues that it brings, but this is the one that can't really be "fixed" in my mind. I'm in an interesting position that I'm somewhat new to my field but not new to learning technical skills. The balance of doing grunt work to strengthen my chops and just getting things done or avoiding annoying work is ongoing. I don't have answers, but more and more I've been doing the grunt work. One, because I've found some flagship models less useful lately, not sure if they've been detuned or just tuned away from how I like to us them. And two because I've found that using AI is more friction over the long term than doing the work to memorize or master the boring or annoying thing. There's joy for me in knowing things and as the article points out: you can't nod along with an answer and call it knowledge gained.
I guess the problem is that there is a very specious rebuttal that says that plato objected to writing, people flipped out about calculator usage, and computers etc and we are all fine.
But we aren't fine. I am deeply embedded in high tech stuff and it makes me exceptionally uncomfortable how specialized our thinking has become and how disconnected we are from almost everything we use. LLMs are the next rung on the ladder of delegation, and instead of saying "it's fine, look at all the runs we've already climbed, this the natural progression of things" I feel like we should be saying "we should maybe descend a little?" It is nuts how many things your average Joe interacts with that they have zero understanding of, and it's allowing for all kinds of nasty consequences (algorithmic feeds, dynamic black box pricing with real consequences, i.e. insurance). AI labs can't really explain why models do what they do, and while they are chipping away at the problem, we are sprinting towards using them everywhere.
I feel for the young'uns. I have no idea how they will navigate this without the grace and generosity that was afforded to me when I was a teen to make mistakes without something ubiquitous and available and way smarter than me whispering answers in my ear.
But just look at the history of human civilization and all technological advancement for examples of this over thousands of years. People write without knowing how to make paper, pens, and colored inks. Artists paint without knowing how to make their own paints, brushes, canvas, sthretchers and frames. Books are made by authora who don't know how to print the pages, bind them together, or work a leather cover. Bronze workers create things from bronze without knowing how to mine it, or smelt it. We drive our cars without having worked the facotry lines, or the oil rigs, or the refineries.
There is literally too much to know in the world to become an expert in such a wide range in a single human lifetime. The only way to continue pushing the boundaries of human knowledge is through specialization.
I understand the argument that AI may make it so that no human is the specialist any longer, so who do you go to when you want a bespoke frame made as an artist and no one knows how to make them any longer. That is a problem that already happens, though. There are art forms and traditions in Japan and Korea that only a select few people have been trained how to do: traditional ketsugi and soju come to mind for me. I'm not sure I buy the argument that AI will cause this to happen more.
Right: and this is exactly the genre of argument that I say I don't agree with in my opening paragraph. I am saying that all of it is bad. It's also bad by degrees. It is not hard, even for someone with middling intelligence to follow that many liquids stain, and that many surfaces can be stained. This the basis of the preservation of human knowledge is learnable with much effort. It's not going to be as good as modern inks, or industrially produced paper, but it can sustain knowledge transfer. Try explaining even an 8bit full adder and the physical circuitry for that to most people, and that's not even the beginning of doing anything useful with a computer.
LLMs and other sufficiently complex systems are bad in the sense that responsibility becomes diffuse to the point that no one will take accountability. LLMs are not special in as much as they lie on a spectrum. We've all effectively been told "computer says no" for years. Before computers decision makers hid behind their lackeys to avoid accountability. LLMs are special in the sense that they currently represent for the tip the same shitty spear, and a step change in capability for continuing this trend. Where I think they differ from what we have seen before is the function that is being outsourced: thinking itself. In this regard I do think the medium is the message and the medium is having a profoundly negative effect on people.
Just as a sidebar, as I'm intruding on your conversation, and this reminded me:
I've been feeling less concerned about that recently? A lot of people already go "brain off" in order to survive the modern world (eg buy the first thing they see advertised that meets a need, do zero research before making an investment, follow trends in their community uncritically, etc.); the LLMs seem to just homogenize that. Also if you've ever seen a DIYer try to fix their house, you may have a different opinion on the inherent value of insourcing thought š
A thought I've had is that a major critique re. climate conscious consumption habits is that the friction is too much for people: they can't plan around a slight inconvenience, do research to find a less damaging alternative, sort their recycling, etc. because they have too much going on. Outsourcing thought (so to speak) seems like a decent means of addressing that, at least.
What's missing in this argument is that calculators do not hallucinate.
It's true that fewer people now are able to do addition/multiplication in their heads; but with a calculator they can still do the arithmetic reliably, and that will remain the case for as long as they have a calculator available to them.
LLMs are not like that. LLM output requires skepticism.
I think paper+ink sits somewhere in the middle. If you just wanna write someone a letter, then paper+ink is sufficiently preserving. But for historical records that are supposed to last centuries, it's a different story. Nevertheless, at least we can tell when the ink has run or the paper has torn; it won't suddenly say something different but with the same level of plausibility. LLMs don't just run the ink, they write something different/wrong and we can't tell.
100% agreed. I'm a little hopeful, though?
LLMs present a profound capacity for teaching, beyond what's ever been accessible before. Exceptional people have always been able to self-teach from books or pure experimentation, but the rest of us could really do with a hand (or two) understanding all those rungs society pulled itself up on. At present, we're given neither time to learn nor appropriate resources to learn from, but -- deployed correctly -- LLMs should hopefully be able to address the latter constraint.
Very few organizations (beyond e.g. Khan academy's attempts) are genuinely trying to make a teaching tool, though. The real money is in a magic box that replaces human competency, not a tool for creating a more capable and educated citizenry -- Anthropic isn't being valued at a trillion USD for teaching Joe Average how retirement funds work, or that the krebs cycle exists. But still, I could (realistically) imagine a future where people are more educated on the facts of reality, rather than whatever was pitched to them in their most recent Facebook ad.
(tbh though the only way that's happening is if a Chinese or European model goes for it)
Ehhhhh I dunno. I feel like we're still working on the formalizations, but since this is very much a new frontier in computer algorithms, I'm at least cutting them a lot of slack. The discipline of CS has been very heavy on the "Computer" and extremely light on the "Science" for the entirety of its existence, so people are understandably confused by the prospect of any process that requires experimentation, modelling, and rigor to make progress in (i.e. science). We're doing decently imo from that perspective! We used a lot of discoveries before fully understanding the physics underlying them, so this doesn't seem much different.
Agreed with that, though.
The problem I see with LLMs for teaching though is I think it takes an exceptional individual to wield an LLM and use it to increase their own understanding instead of letting it carry them. Ofc someone could design an LLM experience/ harness explicitly around the learning loop so that it would take significant energy to hold it wrong, but that's not where the money is so I find it unlikely to be as widespread on the same timeline.
I say this as selfishly I use LLMs to do my work more efficiently and pull the ladder up behind me. We aren't hiring grads. We aren't hiring apprentices, we aren't hiring juniors. I don't know who is supposed to replace me when I go. Weird to see my generation (elder millennial) gap those below them just as hard as the ones they were complaining about 10 years prior.
EDIT: on discovering the mechanisms via which LLMs work, I agree there is progress there, I'm just not that bullish on what the possibilities are, and certainly don't think that we will bind our usage of them by our understanding of the underlying mechanics.
I think you might be looking a this slightly askew though. The way you've phrased it here implies that the person who's learning must also understand how to be an effective teacher, and furthermore must also self-educate themselves on how to teach with an LLM specifically.
I work in higher education (as staff), and I've seen some of the ways faculty are beginning to experiment with LLMs in a teaching context. Some of them are kinda ham-fisted and naieve, but some have interesting potential. I think the educational value of LLMs is that they can engage with students in a way no other technology can possibly do. They are the closest we have ever come to being able to give every single student a 1:1 tutor that specializes in every single subject they have.
There are at least a dozen examples that I'm aware of off the top of my head of faculty attempting in various ways to build a customized LLM that will effectively teach students specific concepts. They are refining and specializing the technology to be an effective teaching tool. Where exactly this ends up and how effective it will be is still a wide open question, but I do agree with @kacey that there is enormous potential.
The reason I am phrasing it this way is because no matter what educators produce, the commercial/freemium models and harnesses are not going away. This is like sending alcoholics to hangout at the bar. It will take tremendous force of will to restrict oneself to making life "harder" for themselves now for a payoff later. I cite the entirety of the brain-rotting internet as my evidence.
I think it's great though that schools are looking at building better tools though!
Oh, yeah, agreed that that's not where the money is (although Khanmingo does exist as a counterexample). I don't think it'd take a hundred million dollars to get this right, though, and what we've seen from open weight model development/the recent leaps in frontier models seems to indicate that better harnesses + training on smaller models is the key to improved benchmark performance anyhow? As it'd ultimately only take one model and harness to nail this UX, then minor updates over time to shove updated info into the weights (either with retraining from early checkpoints or something like ROME to edit memories).
(sidebar: while I still had throughput to think about this, it seemed like distributed training of a ~30b equivalent model might be possible. EleutherAI has been trying to do this for a while, too. Widespread training of small, cheap models might adjust the prior for this prediction, too)
Hah, fair. You're definitely ahead of me by nearly a full generation, but those're many of the reasons I bailed on the industry. I was always pushing to hire on more interns and expand the breadth of experiences we'd accept into the company, but management hated it.
I wouldn't worry about it. It feels selfish in the moment, but putting your own mask on first seems like generally good advice. I was beating myself up this morning about how I spent a large chunk of my life's earnings, since if only I'd made some different choices, I might've had a chance to build some at-cost housing for friends who dearly need it. But I don't have a crystal ball, and neither do you -- maybe investing in yourself now will present you with an even greater opportunity in the future to contribute, that otherwise you'd be unable to access.
Mmhm, understood. Honestly I'm just glad that I don't need to talk someone down from blindly following the Anthropic's gospel on circuits š
It's funny, as someone from a third world country, I worry about the exact opposite. LLMs will only exacerbate the current global inequality. Someone who can comfortably spend 300 USD on tokens will always have an advantage over someone for whom 20 USD is a tenth or even a twentieth of their income, even if their level of understanding is the same.
Edit: I realized that I did the math wrong and didnāt mean to use such low numbers. I recently met a developer who makes 500 USD a month and is subscribed to the 100 USD ChatGPT tier. Somehow, in my mind, āspending a substantial amount of your incomeā automatically translated to the cheapest subscriptions. Regardless, my original point still stands.
That's been on my mind a lot too. I'm pursuing study in pure math, which as a field has always prided itself on being uniquely accessible in the sense that all you need to do math research is pen and paper. It's seeming like lately you also may need a GPT Pro subscription...
Totally valid, I've spent a lot of time looking at how to increase "value per token" ratio. While it's a way to approach it, most the effort seems to be on "accuracy per request" because accuracy needs to be higher most the time to have actual vs perceived value since hallucination rates are still high and can be deceptive.
Trying Deepseek v4 Pro with a token-efficient harness called Reasonix (built by the same company) is great because you can go really, really far on no money, but it creeps up often enough that it might be cutting corners and not giving accurate information and that requires MORE tokens to verify claims (go slow: measure twice, cut once) and so you lose token efficiency for confidence in accuracy. It cost me time and pre-paid usage, so there's a higher cost at the end of the day.
Between this and smaller local models, I can only hope more accessibility finds its way into non-frontier labs as a focus so edge devices (phones, personal machines) can achieve value that matters to those individuals.
I think we will be seeing more price competition soon. Open weights models are getting pretty good, and so are cheaper proprietary models. I sometimes use Luna instead of Terra and rarely try Sol or Opus.
In a bizarre and disturbing twist, it is convincingly argued by https://boxobark.ing/3mj42airv3s2o that this entire blog post may itself be written by AI. Given that I thought I could pretty reliably detect AI writing, and that I thought the blog post in question was generally cogent and interesting, I'm not really sure how to react. At the very least, it would be remiss of me not to include this additional context for any readers later on in this increasingly macabre world.
I spend an embarrassing amount of time talking to Claude and the whole essay screamed Claude to me. I didn't want to say anything because it felt like turning the discussion into a bun-flight about authorship was less interesting that the discussion about the underlying arguments themselves.
I've seen the original post floating around but wasn't really interested. The idea that LLMs can contribute to skill athrophy, or not building skills in the first place, is low hanging fruit that's been discussed from every angle over the last couple years.
But after I saw your comment I gave it a skim to see if it was LLM generated and, yes, it definitely is. The author gave it at least one human editing pass after generation which makes it a bit harder to detect. And they did a better job than many do of editing, but as always happens the human editing gets progressively lazier as you go.
If you wonder if something is AI written, ignore the first two paragraphs. They'll often be the only human written part, or the most carefully post edited. Instead start halfway through.
Anyway the tells in the post are so numerous and frequent that anyone who's used frontier agents in any volume will recognize the tropes.
I imagine the "author" convinced themselves they were being profoundly ironic by using AI to write the post. But I suspect that really they're just lazy and cynical.
In the context of the point their LLM is trying to make for them: The prompter robbing themselves of the opportunity to develop skills and deep mental models around writing. They're also missing out on the rewarding dynamic created by respecting your audience.
Its that obfuscation part that disappoints me the most, and its pretty damning to me. Mainly because it goes against the entire premise of the author's intent, a very Bob move. So does the author not really believe what they write, or do they make exceptions for themselves because its just that easy, and they don't want the backlash?
In any case you know they will update claude's memory to not ever touch quotes going forward.
Thank you for including it, I have had a quick skim, and I'm torn about whether I think it is AI generated text or not. I don't think I am that good at detecting AI text (or at least I'm not according to the the wikipedia article on it), I will see if the author writes a rebuttal (maybe they already have, but I don't have time to search for it atm).
Detecting LLMisms is something you learn from practice and you get better at it. There are patterns that are invisible at first, but you start seeing them after repeated exposure. And I wonder if there are more patterns that I donāt see yet?
I enjoyed this read and I think it sums up my thoughts about AI: if you donāt do the grunt work you canāt understand what you are doing. If you canāt understand it, then you canāt check it, and you lose half the reason of doing it in the first place (that deep understanding).
I tend to avoid LLMs in the same way I avoid using th car for trips I can walk- I donāt want those muscles to atrophy, and learning to struggle sometimes makes you strong and competent.
Thanks for sharing. This is exactly my biggest issue with AI. I'm grumpy about all sorts of other issues that it brings, but this is the one that can't really be "fixed" in my mind. I'm in an interesting position that I'm somewhat new to my field but not new to learning technical skills. The balance of doing grunt work to strengthen my chops and just getting things done or avoiding annoying work is ongoing. I don't have answers, but more and more I've been doing the grunt work. One, because I've found some flagship models less useful lately, not sure if they've been detuned or just tuned away from how I like to us them. And two because I've found that using AI is more friction over the long term than doing the work to memorize or master the boring or annoying thing. There's joy for me in knowing things and as the article points out: you can't nod along with an answer and call it knowledge gained.