{
  "video_id": "dHBEQ-Ryo24",
  "channel_slug": "machinelearningstreettalk",
  "channel_handle": "machinelearningstreettalk",
  "title": "The Dangerous Illusion of AI Coding? - Jeremy Howard",
  "duration_seconds": 5200.0,
  "url": "https://www.youtube.com/watch?v=dHBEQ-Ryo24",
  "upload_date": "",
  "transcript": "It it it it literally disgusts me.\nLike, I literally think it's it's\ninhumane.\nMy mission remains the same as it has been\nfor, like, 20 years, which is\nto stop people working like this.\nJeremy Howard, a deep learning pioneer, a\nKaggle grandmaster, he\nis a huge advocate for actually\nunderstanding what we are\nbuilding through an interactive loop, a\nnotebook, a REPL,\nthe act of poking at a problem until it\npushes back.\nHe argues this is where the real insight\nhappens.\nAnd the funny thing is they're both right.\nLLMs cosplay understanding things.\nLike, they pretend to understand things.\nNo one's actually creating 50 times more\nhigh quality software\nthan they were before.\nSo we've actually just done a study of\nthis and there's a tiny uptick,\ntiny uptick in what people are actually\nshipping.\nThe thing about AI based coding is that\nit's like a slot machine and that you you\nhave an illusion of control, you know, you\ncan get to craft your prompt, and your\nlist of mcps,\nand your skills, and whatever, and then\nbut in the end, you pull the lever.\nRight? Here's a piece of code that no 1\nunderstands.\nAnd am I going to bet\nmy company's product on it?\nAnd I the answer is I don't know.\nBecause like I I don't I don't like, I I\ndon't know what to do now\nbecause no one's, like, been in this\nsituation.\nThey're they're really bad at software\nengineering.\nAnd then I think that's possibly always\ngonna be true.\nThe idea that a human can do a lot more\nwith a computer when the human can, like,\nmanipulate\nthe objects in inside that computer in\nreal time, study them,\nmove them around, and combine them\ntogether.\nWhoever you listen to, you know, whether\nit be Feynman or whatever, like, you\nalways hear from the great scientists how\nthey build deeper\nintuition by by building mental models,\nwhich they get over time, by\ninteracting with the things that they're\nlearning about.\nA machine could kind of build an\neffective hierarchy of abstractions about\nwhat the world is and how it works\nentirely through looking at the\nstatistical correlations of a huge corpus\nof text using a deep learning model.\nThat was my premise.\nThis video is brought to you by NVIDIA\nGTC.\nIt's running March 16 until the nineteenth\nin San Jose and streaming free\nonline. The key topics this year are\nagentic AI and reasoning, high performance\ninference and training, open models, and\nphysical AI and robotics.\nI'm so excited about the DGX Spark.\nI've been on the waiting list for over a\nyear now.\nIt's a personal supercomputer that is\nabout the size of a Mac Mini.\nIt's the perfect adornment to a MacBook\nPro by the way, and you can fine tune a\n70,000,000,000 parameter language model\nwith 1 of these things and I'm giving 1\naway for free. All you have to do is sign\nup to the conference and attend 1 of the\nsessions using the link in the\ndescription.\nAs for the sessions, I'm interested in\nattending Aman Sanger's talk.\nSo, he's the co-founder of Cursor and his\nsession is code with context\nbuild an agentic IDE that truly\nunderstands your code base.\nNow, obviously Jensen's keynote is on\nMarch 16.\nHe said he's going to unveil a new chip\nthat will surprise the world.\nTheir next generation architecture Vera\nRubin is already in full production and\nthere's speculation we might even get an\nearly glimpse of their new Feynman\narchitecture. So don't forget folks, the\nlink is in the description.\nIf you're attending virtually, it's\ncompletely free.\nDon't miss it. Jeremy Howard, welcome to\nMLST.\nI mean, welcome to my home.\nThanks for coming. Yeah. Well, where are\nwe now?\nWe are in beautiful Moreton Bay in\nSoutheast Queensland.\nWe are by the sea in my backyard.\nWeather didn't disappoint.\nIt certainly didn't. It doesn't often but\nif you were here yesterday, it would have\nbeen very different. Well, don't know\nwhere to start.\nI've been I've been a huge fan probably\nsince about 02/1718.\nOf course, you had the famous ULMFiT paper\nand when I was at Microsoft I remember\ndoing a presentation about that because it\nwas actually, I mean now we take it for\ngranted that we fine tune language models\non a corpus of\ntext and then we kind of like continue to\ntrain them and and specialize them.\nBut apparently this was not received\nwisdom.\nNo. This was the first time it happened.\nYeah. Kind of the first or second.\nSo, McCann and Andrew Dai had done\nsomething a few years ago, but they had\nmissed the key point, which is the thing\nyou pre train on has to be a general\npurpose corpus. So no 1 quite realized\nthis key thing, and maybe I had a bit of\nfortune here that my background was in\nphilosophy and cognitive science, and\nso I spent some decades thinking about\nthis.\nThe technical architecture of of of\nULMFiT, just just sketch that out.\nI'm a huge fan of regularization.\nI'm a huge fan of taking a model that's\nincredibly flexible, and then making it\nmore constrained, not by decreasing the\nsize of architecture, but by adding\nregularization. So even that at the time\nwas\nextremely controversial, but that was by\nno means a unique insight of of\nours. So what Stephen Merity had done is\nhe'd taken\nthe extreme flexibility of an LSTM, a kind\nof the\nclassic state for recurrent neural net\ntowards which things are kind\nof gradually heading back towards\nnowadays, added added 5 different types of\nregularization. He added every type of\nregularization you can imagine, and then\nthat was my starting point, was to say,\nnow have a massively flexible deep\nlearning model that can be as powerful as\nI want it to be, and it can also be as\nconstrained as I need it to be, and then I\nneeded a really big corpus of text.\nFunnily enough, this is also Steven.\nHe had been at Common Crawl, and I think\nhe helped or made the\nWikipedia dataset.\nThen I realized actually the Wikipedia\ndataset made lots of assumptions.\nIt had all these like, unk for unknown\nwords, because it all assumed\nclassic NLP approaches.\nSo I redid the whole thing, created a new\nWikipedia dataset.\nNow it's my general corpus.\nAnd then I used an AWD-LSTM, trained it so\nit's actually\novernight. So for 8 hours on a gaming GPU,\nyou\nknow. Because I was at the University of\nSan Francisco, we didn't have heaps of\nresources.\nProbably like a 20 80 TI or something, I\nsuspect.\nAnd then the next morning when I woke up,\nI then it's the same 3 stage\narchitecture that we do today.\nYou know, pre training, mid training, post\ntraining.\nMhmm. So then I figured, okay, now that\nI've trained something to predict the next\nword of Wikipedia, it must know a lot\nabout the world.\nI then figured if I then fine tune it on a\ncorpus specific, so what\nwe could now call supervised fine tuning\ndataset, which in this case was the\ndataset of movie reviews.\nIt would become especially good at\npredicting the next word of those, so I'd\nlearn a lot about movies.\nDid that for like an hour, and then like a\nfew minutes of\nfine tuning the downstream classifier,\nwhich was a classic academic\ndata set. It's kind of considered the\nhardest 1, which was to take like 5,000\nword movie reviews and to say like, is\nthis a positive or negative sentiment?\nWhich today is considered easy.\nBut at that time, you know, the only\nthings that did it quite well were highly\nspecialized models that people wrote their\nwhole PhDs on, and I\nbeat all of their results, you know, 5\nminutes later when it fine tuning\nthat model.\nIt was amazing.\nAnd the other interesting thing is this\nkind of methodology around how\nyou do the fine tuning.\nYeah. So the how we do the fine tuning was\nsomething we had developed at Fast AI.\nSo this is kind of year 1 of Fast AI.\nSo this is still in our very early days.\nAnd 1 of the extremely controversial\nthings we did was we felt that\nwe should focus on fine tuning existing\nmodels, because we thought fine tuning was\nimportant. Some other folks were doing\nwork contemporaneously with that,\nso Jason Yosinski did some really great\nresearch.\nI think it's during his PhD on how to fine\ntune\nmodels, and how good they can be, and some\nother folks in the in the computer vision\nworld. We were, you know, amongst the\nfirst.\nThere's a bunch of us kind of really\ninvesting in fine tuning.\nAnd so, yeah, we we felt that using a\nsingle learning rate to fine\ntune the whole thing all at once made no\nsense, because the different layers have\ndifferent behaviors. This is 1 of the\nthings Jason Yosinski's research also\nshowed. We developed this idea of like,\nwell, it's also way faster if you just\ntrain the last layer.\nRight? Because it only has to back prop\nthe last layer.\nAnd then once that's pretty good, back\nprop the next the last 2 and then the last\n3. And then we use something called\ndiscriminative learning rates.\nSo different layers, we would give\ndifferent learning rates to.\nAnd then another critical insight that no\n1 realized for years, even though we\ntold everybody, was that you actually have\nto fine tune every batch norm.\nSo all the normalization layers, you do\nactually have to fine tune, because\nthat's moving the whole thing up and down,\nknow, changing its scale.\nSo yeah, when you do that, you can often\njust fine tune the last layer or 2.\nAnd we found that actually with ULMFiT,\nalthough we did end up unfreezing all\nthe layers, only the last 2 were really\nneeded to get the close to state of the\nart result. So it like took like seconds.\nYeah. Because the discriminative learning\nrate thing is interesting because I I\nthink the received wisdom at the time was\nwhen you fine tune a model, if the\nlearning rate is too high, you kind of\nblow out the representations.\nSo I guess the wisdom was if if you don't\nhave a really low learning rate, you'll\njust destroy the representations.\nI mean, there there was no received wisdom\nbecause nobody talked about it.\nNo 1 cared, you know.\nIt's just this\nsort of like, nearly no 1 cared.\nTransfer learning was just not something\nanybody thought about.\nAnd Rachel and I felt like it matters more\nthan anything,\nyou know, because only 1 person has to\ntrain a really big model\nonce, and then the rest of us can all fine\ntune it.\nSo we thought we just should learn how to\ndo that really well.\nSo we\nspend a lot of time just trying lots of\nthings.\nBut in the end, the intuition was pretty\nstraightforward, and what intuitively\nseemed like it ought to work, basically\nalways did work.\nWhich is another big difference between\nhow people still today tend to do\nML research, is I think it's all about\nablations,\nand you can't make any assumptions or\nguesses, and it's not at all true.\nI find nearly everything that I expect to\nwork almost always works first time,\nbecause I spend a lot of time building up\nthose intuitions, that kind of\nunderstanding of how gradients behave.\nI I think there's a dichotomy though\nbetween continual learning, which is when\nwe want to keep training the thing but\nmaintain generality, versus fine tuning\na thing to do something specific.\nSo there's always been this idea that,\nyes, you can make a model specific.\nYou can bend it to your will.\nBut you lose generality, and you kind of\ndegrade the representation.\nSo tell me about that. Yeah, there's some\ntruth in that, although not as much as you\nmight think. On the whole, the big problem\nis that people don't actually look at\ntheir activations and don't actually look\nat their gradients.\nSo something we do in our software, in our\nfast AI software, is we have built into it\nthis ability to to see in a glance what\nyour entire\nnetwork looks like. And once you've done\nit a few times, it just takes a couple of\nhours to learn, you can immediately see,\noh, I I see.\nThis is over trained or under trained or\nthis layer that something went wrong.\nIt's not a mystery, you know.\nSo basically what happens is, for example,\nyou end up with with dead neurons that go\nto a point where they they've got 0\ngradient regardless of what you do with\nthem.\nThat often happens if they, you know, head\noff towards infinity.\nYou can always fix that.\nSo yeah, it's it's not as bad as people\nthink by any means.\nSomething that trains well for continuous\nlearning, when done properly,\ncan also be done well to train well for a\nparticular task,\nif you're careful. In a sense, you do want\nthe neurons to die out.\nAnd I'll explain why what I mean by this.\nLike, we want to bend the behavior of\nmodels to introduce implicit\nconstraints. Because without constraints,\nthere is no creativity, there is no\nreasoning, and and so on and so forth.\nSo so in a sense, actually want it to say,\ndon't do that.\nYou want it to do something else.\nI don't think of it that way.\nLike, to me, it's more like, I find\nthinking about humans extremely\nhelpful when it comes to thinking about\nAI.\nI find they\nbehave more similarly than differently,\nand my intuition about each tends to work\nquite well. You know, with a human, when\nyou learn something\nnew, it's not about unlearning something\nelse.\nAnd so something I always found is when I\ngot models to try to learn to\ndo 2 somewhat similar tasks, They almost\nalways got better at both of\nthem than 1 that only learned 1 of them.\nI was reminded a little bit of, you know,\nthe DINO paper from LeCun's lab.\nSo this whole kind of regime of self\nsupervised learning with with I mean, that\nthat was that was a vision model.\nBut the idea was, Okay, so we're doing pre\ntraining.\nAnd we want to maintain as much diversity\nand fidelity as possible so that when we\ndo the downstream task, we can kind of\nwe've got more things that we can latch\non. Yeah, yeah. And\nyou know, semi supervised and self\nsupervised learning was such an\nunappreciated area.\nAnd, yeah, Yann LeCun was absolutely 1 of\nthe guys who was also working\non it. Yeah. I actually did a post,\nbecause I was so annoyed at how few people\ncared about semi supervised learning, did\na whole post about it years ago.\nJan LeCun looked at it for me as well,\nand, you know, suggested a few other\npieces of work that I'd I'd missed, and\nbut I was kind of surprised at how,\nknow, how incredibly useful it is to\nbasically say, like, basically\ncome up with a pretext task, right?\nSo envision so we did this envision before\nULM fit, so it was like,\nin medical imaging, you know, take a\nhistology slide and predict,\nyou know, mask out a few squares, and\npredict what used to be there.\nSo some of my students at USF I had doing\nstuff with that.\nIt was basically entirely taking stuff\nthat we and others had already done in\nvision. Yep.\nSo like this idea of masking out squares.\nWe didn't invent it.\nMasking out words was the obvious thing,\nyou know, and this idea of\ngradually unfreezing layers, we had done\nbefore in computer vision.\nThe whole idea of starting with a pre\ntrained model that was general purpose had\nbeen in computer vision.\nThere was a really classic paper,\nactually, in computer vision in\nmight have been around 2015, was entirely\nan empirical paper saying, look what\nhappens when we take a pre trained\nImageNet model predicting what sculptor\ncreated this or predicting what\narchitecture style this is, and like in\nevery task, it got the state of the art\nresult.\nAnd it really surprised me people didn't\nlook at that and think like, I bet that\nought to work in every other area as well,\nwhether it be genome sequences, or\nlanguage, or whatever.\nBut people have a bit of a lack of\nimagination, I find.\nThey tend to assume things only work in 1\nparticular field.\nThat's really true.\nYeah. I mean, I guess there's 2 things\nthere.\nMean, of all, we were kind of hinting at\nthis notion of almost Goodhart's Law, the\nshortcut rule that you get exactly what\nyou optimize for at the cost of everything\nelse. But that doesn't seem to be the case\nbecause we can optimize for perplexity in\nthe case of language models. And as you\nsay, what seems to happen is we're getting\ninto the distributional hypothesis here a\nlittle bit.\nSo you know the word by the company it\nkeeps.\nSo when we have an incredible amount of\nassociative data, it might be masked auto\nprediction or any of these things like\nthat, the model seems to build something\nthat we might call an understanding.\nWell, I I have always thought of it as a\nhierarchy of of abstractions, You know, it\nit it needs if it's gonna predict, you\nknow, if the document\nis\nhere was the,\nyou know, opening that, you know, that\nBobby Fischer used, and has chess\nnotation to predict the next thing, it\nneeds to know something about\nchess notation, or at least openings.\nIf it's like,\nyou know, and this was vetoed by the 1956\nUS president, comma,\nyou need to know it's bit like, you don't\neven you don't just need to know who the\npresident was, but the idea that there are\npresidents.\nAnd therefore, the idea that there are\nleaders, and therefore, the idea that\nthere are groups of people who have\nhierarchies, and therefore, that there are\npeople, and therefore, that there are\nobjects, and like, you can't predict the\nnext word of a sentence\nwell without knowing all of these things.\nSo that knowing\nmy hypothesis for why I created ULMFiT was\nto say it would end to to to\ncompress that as well as possible to get\nthat knowledge, it would have to create\nthese abstractions, these hierarchies of\nabstractions somewhere deep inside its\nmodel. Otherwise, how could it possibly do\na good job of predicting the next word?\nYou know, and because\ndeep learning models are universal\nlearning machines, you know, and we had a\nuniversal way to train them, I figured if\nif we get the\ndata right and if the hardware is good\nenough, then in theory,\nwe ought to be able to build that next\nword predicting machine, which ought\nto implicitly build a hierarchical\nstructural understanding\nof the things that are being described by\nthe text that it is learning to predict?\nI think that they can know in quite a, you\nknow, they they know in quite a\nsuperficial way. So there's a myriad of\nsurface statistical\nrelationships, and they generalize\nextraordinarily well.\nIt's it's miraculous.\nIt is. But the thing is, I want to\ncontrast this with other comments you've\nmade about creativity. So I think\nknowledge is about constraints.\nAnd I think creativity is the evolution of\nknowledge, respecting those constraints.\nTherefore, AI is not creative.\nAnd you've said the same thing.\nYou've said AI isn't creative.\nSo like, on the 1 hand, how can you say\nthat they know and not think that\nthey can be created? I mean, I don't think\nI've used that exact expression.\nYou know, I know I've actually I remember\nchatting with Peter Norvig on camera, and\nboth of us said, well, actually, they kind\nof are creative, like, we just gotta be a\nbit careful about our choices of words, I\nguess. So, you know, Piotr Wozniak, who's\na guy I really really respect,\nwho kind of rediscovered spaced repetition\nlearning, built the SuperMemo system, and\nis the modern day guru of memory.\nThe entire reason he's based his life\naround remembering things is because\nhe believes that creativity comes from\nhaving a lot of stuff remembered, which\nis to say putting together stuff you've\nremembered in interesting ways\nis a great way to be creative.\nLLMs are actually quite good at that, but\nthere's a kind of creativity they're\nnot at all good at, which is, you know,\nmoving outside the\ndistribution. So which I think is where\nyou're heading with your question.\nBut I'm just kind of I'm framing it this\nway to say, you have to be so nuanced\nabout this stuff because if you say like\nthey're not creative, it gives you the can\ngive you the wrong idea because they can\ndo very creative\nseeming things.\nBut if it's like, well, can they really\nextrapolate outside the training\ndistribution? The answer is no, they\ncan't.\nBut the training distribution is so big,\nand the number of ways to interpolate\nbetween them is so vast,\nwe don't really know yet what the\nlimitations of that is.\nBut I see it every day, you know, because\nI my my work is R and D.\nI'm constantly on the edge of and outside\nthe training data.\nI'm doing things that haven't been done\nbefore.\nAnd there's this weird thing, I don't know\nif you've ever seen it before, I see it\nbut I see it multiple times every day,\nwhere the LM goes from being incredibly\nclever\nto like worse than stupid, like like not\nunderstanding the most basic fundamental\npremises about how the world works.\nYeah. Yeah. And it's like, oh, whoops, I\nfell outside the training data\ndistribution. It's gone dumb.\nAnd then, like, there's no point\nhaving that discussion any further because\nYes.\nYou know, you've lost it at that point.\nYes. I mean, I love, you know, Margaret\nBoden.\nShe had this kind of hierarchy of\ncreativity.\nSo there's like combinatorial, exploratory\nand transformative.\nAnd the models can certainly do\ncombinatorial creativity.\nBut for me, it's all about constraints.\nThat I mean, this is what Boden said.\nAnd even Leonardo da Vinci, he said that\ncreativity is all about constraints.\nAnd you've spoken about we'll talk about\nthis dialogue engineering.\nBut what happens is when we talk with\nlanguage models, it's a specification\nacquisition problem. So we go back and\nforth.\nAnd actually, when we think, the process\nof intelligence is about building this\nimaginary Lego block in our mind and\nrespecting various constraints.\nAnd when you respect those constraints and\nyou just continue to evolve, then those\nthings are said to be creative.\nSo language models, when you add\nconstraints to them, so this could be via\nsupervision, via critics, via verifiers,\nThen they are creative.\nAnd AlphaEvolve, we've seen many examples\nof this.\nBut the illusion is, on their own, sans\nconstraints obviously, they have this\nbehavioral shaping stuff that we're\ntalking about.\nThey don't have hard constraints.\nAnd that's why they can't go outside their\ndistribution.\nI mean, I think they can't go outside\ntheir distribution because it's just\nsomething that that type of mathematical\nmodel can't do.\nYou know, I mean, it can do it, but it\nwon't do it well.\nYou know, when you look at the kind of 2 d\ncase of fitting a curve to data,\nonce you go outside the area that the data\ncovers, the curves disappear\noff into space in wild directions, you\nknow.\nAnd that's all we're doing, but we're\ndoing it in multiple dimensions.\nYep. I think Boden might be pretty shocked\nat\nhow far compositional creativity can go\nwhen you can compose\nthe entirety of the human knowledge\ncorpus.\nAnd I think this is where people often get\nconfused, because it's like So\nfor example, I was talking to Chris Latner\nyesterday about\nhow Claude Anthropic, you know, had had\ngot\nClaude to write the C compiler.\nAnd they were like, oh, this is a clean\nroom C compiler.\nYou can tell it's clean room because it\nwas created in Rust, you know,\nand so Chris created the kind of,\nyou know, I guess it's probably the top\nmost widely used c c plus plus compiler\nnowadays playing on top of LLVM, which is\nthe most widely used kind of foundation\nfor compilers. They're like, Chris didn't\nuse rust.\nThis is, you know, and we didn't give it\naccess to any\ncompiler source code.\nSo it's a clean room implementation.\nBut that misunderstands how LLMs work.\nRight? Which is all of Chris's work was in\nthe training data.\nMany many times LLVM is used widely and\nlots and lots of things are built on it,\nincluding lots of c and c plus plus\ncompilers.\nConverting it converting it to Rust is an\ninterpolation\nbetween\nparts of the training data, you know.\nIt's a style transfer problem.\nSo it's definitely compositional\ncreativity at most, if you can call it\ncreative at all. And you actually see it\nwhen you look at the the\nrepo that it created.\nIt's copied\nparts of the LLVM code, which today Chris\nsays like, oh, I\nmade a mistake. I shouldn't have done it\nthat way.\nNobody else does it that way, You know?\nOh, wow. Look. They're the only other 1\nthat did it that way.\nThat doesn't happen accidentally.\nThat happens because you're not actually\nbeing creative.\nYou're actually just finding the kind of\nnonlinear average point in your\ntraining data between, like, Rust things\nand building compiler things?\nAll of that is true. I mean, first of all,\nI I think we shouldn't underestimate the\nsize of how big this combinatorial\ncreativity is.\nSo all of that is true. So the code is on\nthe Internet, but also, they had a whole\nbunch of tests which were scaffolded,\nwhich meant that every single time some\ncode was committed, they could run the\ntest and they basically had a critic.\nAnd they could then do this autonomous\nfeedback loop.\nSo in a sense, it's very similar to the\nrecent research by OpenAI and Gemini,\nwhere you're trying to solve a problem in\nmath and you already\nhave an evaluation function, the same on\nthe ARC prize, right?\nYou have an evaluation function.\nAnd what people discount is even knowledge\nof what the evaluation function is,\nis partial knowledge of the problem.\nSo you can then brute force search.\nYou can use the statistical pattern\nmatching.\nUse the verifier as a constraint, and you\ncan actually they don't even need to do\nthat. Right? Like, they literally already\nknow how to pass those tests because\nthere's lots of software that already does\nit.\nRight. So it just uses that and translates\nthem to\nRust. Like, that's that's all it did,\nwhich is impressive.\nYeah. And if you I'm much less familiar\nwith math than I am computer science, but\nfrom talking to mathematicians,\nthey tell me that that's also what's\nhappening with like, Erdős's problems and\nstuff. It's some of them are newly solved.\nYeah.\nBut they are not\nsparks of insight.\nYou know, they're solving ones that you\ncan solve by meshing up together\nvery closely related things that humans\nhave already figured out.\nSo on the subject of Claude code.\nNow I know you've spoken extensively about\nvibe coding.\nActually, Rachel had some interesting\nwriteup.\nI mean, she she quoted the the METR study,\nwhich showed that productivity actually\nwent down when people were vibe coding.\nBut I think And they thought that they\nwent up, which is the most interesting And\nthen also there was the Anthropic study.\nMean, know, maybe we should rewind a\nlittle bit.\nI mean, Dario had this essay out the other\nday.\nI think it's called the Adolescence of\nTechnology or something like that.\nAnd then he was basically saying, look,\nyou know, we have all of these amazing\nsoftware engineers at Anthropic.\nAnd they are just so productive.\nAnd he was extrapolating to the average\nsoftware engineer.\nSo there's gonna be mass unemployment\nbecause soon, we're gonna be able to\nautomate all of this with AI.\nI mean, it it doesn't make any sense.\nElon Musk said something a bit similar a\nfew days ago, saying like, oh, LLMs will\njust spit out the machine code directly.\nWe won't need libraries, programming\nlanguages.\nYeah.\nYeah. Look, the thing is none of these\nguys have\nhave been software engineers recently.\nI'm not sure Dario's ever been a software\nengineer at all.\nSoftware engineering is a unusual\ndiscipline, and a lot of people mistake it\nfor being the same as typing code into an\nIDE.\nCoding is another 1 of these style\ntransfer problems.\nYou take a specification of the problem to\nsolve and you can use your\ncompositional creativity to find the parts\nof the training data which interpolated\nbetween them solve that problem, and\ninterpolate that with\nsyntax of the target language, and you get\ncode.\nThere's a very famous essay by Fred Brooks\nwritten many decades ago,\nno silver bullet, and which it almost\nsounded like he was talking about today.\nIt it he was specifically saying something\nhe was just pointing something very\nsimilar, which is in those days it was all\nlike, oh, what about all these new fourth\ngeneration languages and stuff like that,\nyou know.\nWe're not gonna need any coders anymore,\nany software engineers anymore, because\nsoftware is now so easy to write, anybody\ncan write it.\nAnd he said, well,\nhe guessed that you could get at maximum a\n30% improvement.\nHe specifically said a 30% improvement in\nthe next decade, but I don't think he\nneeded to limit it that much.\nBecause the vast majority of work in\nsoftware engineering isn't typing in the\ncode. Yep.\nSo in some sense, parts of what Dario said\nwere right, just\nlike for quite a few people now, most of\ntheir\ncode is being typed by a language\nmodel.\nThat's true for me.\nSay, like, maybe 90%.\nBut it hasn't made me that much more\nproductive,\nbecause that was never the slow bit.\nIt's also helped me with kind of the\nresearch a lot and figuring out, you know,\nwhich files are gonna be touched.\nBut anytime I've made any attempt to\ngetting an LLM to like design\na solution to something that hasn't been\ndesigned lots of times before,\nit's it's horrible.\nBecause what it actually, every time,\ngives me is the design of something that\nlooks on its surface a bit similar.\nAnd often that's gonna be an absolute\ndisaster, because things that look on the\nsurface a bit similar and like I'm\nliterally trying to create something new\nto get away from the similar thing.\nIt's very misleading.\nFirst of all, I'm I'm exasperated by what\nI see as the tech bro\npredilection to misunderstand cognitive\nscience and philosophy and and what not.\nBecause we've we've spoken to so many\nreally interesting people on MLST like for\nexample César Hidalgo, he wrote this book,\nThe Laws of Knowledge.\nAnd and even Mazviita Chirimuuta, she's a\na philosopher of neuroscience and she was\ntalking all about, you know, like flipping\nbasically that knowledge is protean.\nSo yeah, I think that knowledge is\nperspectival.\nI don't think that knowledge can be this\nabstract perspective free thing that can\nexist on Wikipedia.\nAnd I also think that knowledge is\nembodied and it's alive.\nIt's something that exists in us.\nAnd the purpose of an organization is to\npreserve and evolve knowledge.\nSo when you start delegating cognitive\ntasks to language models, you actually\nhave this weird paradoxical effect that\nyou erode the knowledge inside the\norganization. Well, that's true.\nAnd that's terrifying. There's often these\nthese arguments online between\npeople who are like, LLMs don't understand\nanything.\nThey're just pretending to understand.\nMhmm. And then other people are like,\ndon't be ridiculous.\nLook what this LLM just did for me.\nRight? And the funny thing is they're both\nright.\nLLMs cosplay understanding things.\nLike, they pretend to understand things.\nAnd this is the interesting thing about\nthe early kind of work with, like,\ncognitive science work with, like, Daniel\nDennett.\nThat's basically what the Chinese room\nexperiment is.\nRight? It is you've got a guy in a room\nwho can't speak Chinese\nat all, but he sure looks like he does\nbecause you can feed in questions and he\ngives you back answers, but all he's\nactually doing is looking up things in a\nhuge array of books or machines or\nwhatever.\nThe difference between pretending to be\nintelligent and actually being intelligent\nis entirely unimportant, as long as you're\nin the region in which the\npretense is actually effective, you know.\nSo so it's actually fine for a great many\ntasks that\nLLMs only pretend to be intelligent,\nbecause for all intents and\npurposes, it it it just doesn't matter\nuntil you get to the\npoint where it can't pretend anymore.\nAnd then you realize, like, oh my god.\nThis thing's so stupid.\nI'm a fan of Searle, by the way.\nSo, know, he said that understanding is\ncausally reducible but ontologically\nirreducible. And he was saying there was a\nphenomenal component to understanding.\nYou don't even need to go there. Like the\ninteresting thing about knowledge being\nprotean is this idea that, you know, it's\nbasically this Kantian idea.\nThe world is a complex place.\nNone of us understand it. It's like the\nblind men and the elephant.\nWe all have different perspectives. It's a\nvery complex thing.\nAnd so we all we all do this kind of\nmodeling.\nBut the the interesting thing is that the\nlanguage model, sometimes they seem to\nunderstand. And they understand because\nthe supervisor places them in a frame.\nSo inside that frame, so when you have\nthat perspective of the elephants, they're\nactually surprisingly coherent.\nBut we discount the supervisor placing the\nmodels in that frame.\nYeah. Yeah. So that so Searle versus\nDennett, or is it versus Searle and\nDennett, was what everybody was talking\nabout back when I was doing my undergrad\nin philosophy, you know, so and\nConsciousness Explained came out about\nthen, probably Chinese Room a little bit\nbefore.\nThat's interesting because the discussions\nwere the same discussions we're having\nnow, but they've gone from being abstract\ndiscussions to being real discussions.\nIt's helpful if people go back to the\nabstract discussions, because it helps you\nget out of your\nyou know, it's very distracting at the\nmoment to look at something that's\ncosplaying intelligence so well, and go\nback to the\nfundamental question.\nSo anyway, I just wanted to mention that's\nkind of it's it's this interesting\nsituation we're now in, where it's very\neasy\nto\nreally get the wrong idea about what AI\ncan do.\nParticularly when you don't understand the\ndifference between coding and software\nengineering. Yeah. Which then takes me to\nyour point or your question about\nthe implications of that\nfor organizations.\nYep. You know, a lot of organizations are\nbasically betting their\nfutures on a speculative premise,\nwhich is that AI is gonna be able to\ndo everything better than humans,\nor at least everything in coding better\nthan humans.\nI I worry about this a lot, both for the\norganizations and for the humans, you\nknow. For the humans, when you're not\nactively using your\ndesign and engineering and coding muscles,\nyou don't grow.\nYou might even wither.\nBut you at least don't grow.\nAnd, you know, speaking of the CEO of an r\nand d startup, you know, if if\nmy staff aren't growing, then we're gonna\nfail.\nYou know? We that we can't let that\nhappen.\nAnd getting better at the particular\nprompting skills, whatever details of the\ncurrent generation of\nAI, CLI frameworks isn't growing.\nYou know, that's that's like that's as\nhelpful as learning about the\ndetails of some AWS API when you don't\nactually understand how the\nInternet works, you know. It's not it's\nnot reusable\nknowledge. It's ephemeral knowledge.\nSo like, if you wanted to, you can\nactually use it as a\nlearning superpower.\nBut also, it can do the opposite.\nYou know, the natural thing it's gonna do\nis\nremove your confidence over time.\nI agree that that's the natural thing.\nAnd this this is especially pertinent for\nyou because your your career has been\naround basically educating people to to\nget, you know, technology and AI literacy.\nSo the default behavior is very similar to\na self driving car.\nBut there's this tipping point where at\nsome point, you're not engaged anymore.\nYou're not paying attention. And you get\nthis delegation of competence.\nAnd you get understanding debt.\nThat's the default thing. So this study\nfrom Anthropic a couple of weeks ago, it\ncontradicted Dario completely because it\neven said that, yeah, there were a few\npeople in the study that were asking\nconceptual questions that are actually\nkind of keeping on top of things.\nAnd they had a gradient of learning, but\nmost people didn't.\nAnd my hypothesis about that is that the\nideal situation for Gen AI\ncoding is that like us, we've been writing\nsoftware for decades.\nWe already have this abstract\nunderstanding.\nWe're using it in domains that we know\nwell.\nAnd we can specify, we can remove loads of\nambiguity, we can track and we can go back\nand forth and we can we can stay in touch\nwith the process.\nBut what happens is that the the default\nattractor is for people to just go into\nthis autopilot mode and they've got no\nidea what's happening and it's actually\nmaking them dumber. I I created a the the\nfirst deep learning for\nmedicine company called Enlitic back in,\nwhat was that, like, '14.\nAnd our initial focus was on radiology,\nand a lot of people\nwere worried\nthat this would cause radiologists to\nbecome less effective at radiology.\nAnd I strongly felt the opposite, which is\nand I did quite a bit of research into\nthis, of like what happens when there's\nlike fly by wire in airplanes or anti lock\nbrakes in cars or whatever.\nIf you can successfully automate parts of\na task that really are\nautomatable, you can allow the expert to\nfocus on the things that they need to\nfocus on.\nAnd we saw this happen. So in radiology we\nfound if we could automate\nidentifying the possible nodules in a lung\nCT\nscan, we were actually good at it, which\nwe were, and\nthen we've the radiologist then can focus\non looking at the nodules and trying to\ndecide if they're malignant or what to do\nabout it.\nSo again, it's 1 of these subtle things.\nSo if there's things which you can fully\nautomate effectively in a way that you can\nremove that cognitive burden from a human,\nso that they can focus on things that they\nneed to focus on,\nthat can be good, you know.\nI don't know where we sit in software\ndevelopment because, you know,\nI've been coding for\n40 ish years.\nSo I've written a lot of code and I can\nglance at a screen of\ncode and then, you know, unless it's\nsomething quite weird or sophisticated, I\ncan immediately tell you what it does and\nwhether it works and whatever.\nI can kind of see intuitively things that\ncould be improved, you know, possible\nthings to be careful of.\nI'm not sure I could have got to that\npoint if I hadn't have written\na lot of code. So the people I'm finding\nwho can really\nbenefit from AI right now are either\nreally junior people who\ncan't code at all, who can now write some\napps that they have in their head.\nAnd as long as they work reasonably\nquickly\nwith the current AI capabilities, then\nthey're happy.\nAnd then really experienced people like\nlike me or like Chris Latner, because\nwe can basically have it do some of our\ntyping for us, you know, and some of our\nresearch for us. People in the middle,\nwhich is most people, most of the time, it\nreally worries me because how do you get\nfrom 0.1 to point b Yeah.\nWithout typing code?\nIt might be possible, but we don't have a\nwe have no experience of that.\nWe don't is is it possible?\nHow would you do it? Like, is it kind of\nlike going back to school where at primary\nschool we don't let kids use calculators\nso that they develop their\nnumber muscle. Do we need to do that for\nlike first 5 years as a\ndeveloper? You have to write all the code\nyourself.\nI I don't know.\nBut if I wasn't between, like, 2 and 20\nyears of experienced developer, I would be\nasking that question of myself a\nlot.\nBecause otherwise, you might be in the\nprocess of making yourself\nobsolete. Yeah.\nWell, this is another thing about\nknowledge that this Cesar Hidalgo guy\nsaid. So he said that knowledge is non\nfungible, which means it can't be\nexchanged. So what he means by that is the\nprocess of learning is in some important\nsense not reducible.\nSo you have to have the experience.\nAnd the experience has to have friction.\nAnd when we build models of the world, we\nactually learn like there's this phrase,\nreality pushes back. So we make lots of\nmistakes, and we update our models, and\nwe're just placing these coherence\nconstraints in our model.\nAnd that's how we come to learn. So you\nuse called code, and there's so little\nfriction in the process. That's exactly\nwhat this study from Anthropics said.\nIt said there was so little friction, they\ndidn't learn anything.\nRight. Yeah. No.\nExactly.\nDesirable difficulty is the concept that\nkinda comes up in education.\nBut even going back to the work of\nEbbinghaus, who was the original\nrepetitive space learning guy in the\nnineteenth century, And then Piotr Wozniak\nmore recently, we find the same like, we\nwe we\nknow that memories don't get formed unless\nit is hard\nwork to form them.\nSo, you know, that's where you kind of get\nthis somewhat surprising result that\nsays revising too often is a bad idea,\nbecause it comes to mind too quickly.\nAnd so with repetitive spaced learning,\nwith stuff like Anki and SuperMemo, the\nalgorithm tries to schedule the flashcards\nat a just\nbefore the moment you're about to forget.\nSo then it's hard work.\nSo I I studied Chinese for 10 years in\norder to try\nto learn about learning myself.\nAnd I really noticed this that I used\nAnki, and because it was always\nscheduling my cards just before I was\nabout to forget them,\nit was always incredibly hard work.\nYeah. You know, to do reviews, because\nalmost all the cards were once I was on\nthe verge of forgetting. That was\nabsolutely exhausting.\nBut my God, it worked well. Here I am, I\ndon't really haven't done any study for 15\nplus years, and I still remember my\nChinese.\nWell, I mean, also, coming back to your\nradiology example,\n1 example people give is call centers.\nSo we have this notion that in an\norganization, we have high intelligence\nroles and low intelligence roles.\nAnd for me, intelligence is just the\nadaptive acquisition and synthesis of\nknowledge. So we assume that the low\nintelligence roles doing the call center\nstuff, it doesn't adapt, which means we\ncan there are certain things that an\norganization does that do not change.\nSo we could automate them and we don't\nneed to update our knowledge.\nAnd I think that discounts actually maybe\nwith the radiology example that having\nthis holistic knowledge.\nLike, you know, in a call center, there\nare so many weird edge cases that come in.\nSo many weird things happen and that\nfilters up in the organization and we\nadapt over time. So when you start to\nautomate things, and you actually lose the\ncompetence to create the process which\ncreated the thing in the first place, and\nyou lose the evolvability of that\nknowledge in the organization, you're\nactually kind of cutting your legs off.\nYeah, absolutely.\nAnd so, you know, all I know is in my\ncompany,\nI just I tell our staff all the time,\nalmost the only thing I care about is how\nmuch\nyour your personal human capabilities are\ngrowing.\nYou know, I I don't actually care how many\nPRs you're doing,\nhow many features you're doing.\nLike,\nthere's that nice, you know, John\nOusterhout, the Tcl guy,\nrecently released some of his Stanford\nFriday takeaway\nlectures, and he has this nice 1 called a\nlittle bit of slope makes\nup for a lot of intercept.\nJust basically the idea that that, you\nknow, in your life, if you can focus on\ndoing things that cause you to grow faster\nYeah.\nIt's way better than\nfocusing on focusing on the things that\nyou're already good at, you know, that has\nthat high intercept. So the only thing I\nreally care about, and I think is the only\nthing that matters for my company, is that\nmy team are focusing\non their slope. Yeah.\nIf you focus on just driving out results\nat the limit\nof whatever AI can do right now, you're\nonly caring about the intercept, you know.\nSo I think it's basically a path to\nobsolescence through both the company\nand the people who are in it.\nAnd so I'm really surprised how many\nexecutives of big companies are pushing\nthis now, because it feels like if they're\nwrong, which they probably\nare, and they have no way to tell if they\nare because this is an area they're not at\nall familiar with and never learned it in\ntheir MBAs.\nThey're basically setting up their\ncompanies to be destroyed.\nYeah. And really surprised that,\nyou know, shareholders would let them\ndo that. It'll set up such an incredibly\nspeculative action.\nYeah. Here we are. It feels like a lot of\ncompanies are gonna fail as a result of\nthe amassed tech debt that causes them to\nnot be able to maintain or build their\nproducts anymore. There are loads of folks\nout there like François Chollet.\nLike, he he he really gets it.\nHe he understands this.\nYeah. And, you know, so he's always said\nthat it it's it's about this kind of\nmimetic sharing of cognitive models about\nthe domain and how we refine it together.\nOn the sharing thing, this is another big\nscaling problem with Gen AI coding, right?\nSo the ideal case, I've done this.\nI know a domain really well and I can\nspecify it with exquisite detail and I\ntell Claude code, go and do this thing and\nthe models in my mind doesn't matter.\nAnd then you go into an organization and\nnow I need to share like my knowledge\nwith all of the other people, right? And\nI'm sure you have this in your company as\nwell. You need to that that this knowledge\nacquisition bottleneck is a real serious\nproblem in in organizations.\nSo when it's just me, I I think I'm\nprobably about 50 times more productive\nusing Claude Code. It's absolutely magic.\nAnd I can see why people are so excited\nabout it.\nBut people don't seem to understand the\nbottleneck and and how that doesn't really\ntranslate to many real world\norganizations.\nNo one's actually creating 50 times more\nhigh quality\nsoftware than they were before.\nSo we've actually just done a study of\nthis, and there's a tiny uptick,\ntiny uptick in what people are actually\nshipping.\nThat's the facts. Obviously, I'm an\nenthusiast of AI and what it can do.\nBut also, my wife Rachel recently pointed\nout in an article,\nall of the pieces that make gambling\naddictive are present in Oh, yeah.\nDark flow. Yeah. I was gonna bring that\nup.\nYeah. It's this really awkward situation\nwhere it's\nvery almost everybody I know who got very\nenthusiastic about AI\npowered coding in recent months have\ntotally changed their\nmind about it when they finally went back\nand looked at, like, how much stuff that I\nbuilt during those days of great\nenthusiasm am I using\ntoday? Are my customers using today?\nAm I making money from today?\nAlmost all the money is being made by\ninfluencers, you know, or by\nthe companies that produce the tokens.\nThe thing about AI based coding is that\nit's like a slot machine, and that you you\nhave an illusion of control, you know, you\ncan get to craft your prompt, and your\nlist of MCPs,\nand your skills, and whatever, and then in\nthe end, you pull the lever.\nRight? You put in the prompt, and\nsomething comes back,\nand it's like cherry, cherry, it's like,\noh, next time\nI'll change my prompt a bit, I'll add a\nbit more context, pull the lever again,\npull the lever again. It's the stochastic\nthing.\nYou get the occasional win.\nIt's like, oh, I won.\nI got a feature.\nSo it's got it's got all these hallmarks\nof like, loss disguised as a win,\nsomewhat stochastic, feeling of control,\nall the stuff\nthat gaming companies try to engineer into\ntheir gaming\nrooms. Now, none of that means that AI is\nnot useful,\nbut gosh, it's hard to tell.\nI know. And and Rachel, just just to be\nclear as well, she she also said that 1 of\nthe hallmarks of gambling is that you kind\nof delude yourself that you have some\nawareness of what's going on, but but\nactually you don't.\nBut let let's do the bull case a little\nbit though.\nSo I do I do think in restricted cases, it\nit is it is very useful.\nAnd these are cases where we understand\nand we and we can place constraints and\nspecification. But even in those cases,\ncould argue on the 1 hand that we're\nnot, you know, we're not gonna be\nunemployed anytime soon because you just\ndo more work. On the addiction thing, I've\nnoticed that.\nSo I've had 14 hour Claude code marathon\nsessions and and I actually feel addicted\nto it. It's like a slot machine.\nYou know, it it really is.\nAnd there too. Absolutely.\nYeah. Know. It's and it just I've never\nfelt more drained writing code.\nI actually need to take a rest afterwards,\nlike a few days rest because it completely\ncrap, you know. Yeah.\nDefinitely. I've had some successes.\nRight? And so, in fact, we've spent the\nlast couple of years building\na whole product based around where we know\nthe successes are gonna be, which is when\nyou're working on reasonably small pieces\nthat you can fully\nunderstand, and that you can design, and\nyou can build up your own layers of\nabstraction to create things that are\nbigger than the parts that you're building\nout of. Had a very interesting situation\nrecently where I just it's kind of an\nexperiment basically, which is we we rely\nvery heavily on something called IPy\nkernel, which is the thing that powers\nJupyter notebooks.\nAnd there had been a major version release\nof IPykernel from '6 to '7,\nand it stopped working. And it stopped\nworking in both of the products that we\nwere trying to use it with. 1 was was\ncalled nbclassic, which is the original\nJupyter notebook. And then our own product\ncalled solve it.\nThey would just randomly crash.\nAnd Ipy kernels over 5,000 lines of code.\nIt's very complex code, multiple threads,\nevents,\nlocks, interfaces with IPython, you know,\nwith ZMQ, you know, all kinds of different\npieces,\nDebugPy, and I I couldn't get my head\naround it, and I couldn't\nsee why it was crashing.\nThe tests are all passing.\nI wonder if AI can solve this.\nYou know, it's like I'm always interested\nin the question of like how big a chunk\ncan AI handle on its own right now.\nThe answer turned out to be yes.\nI think it can just.\nIt was like so I spent a couple of weeks,\nI\ndidn't develop a lot of understanding\nabout how IPykernel really worked in the\nprocess, but I did spend quite a bit time\nkind of pulling out separate comp like, so\nthe answer was in 2 hours, codecs 5 point\nI think it was\n5.2 at that time or maybe 3 had just come\nout, couldn't do it.\nThen if I got the $200 a month\nGPT 5.3 pro to fix the\nproblems it could.\nAnd so by rolling back between those 2\npieces of software and those 2\nmodels, I could get things working\nover a couple of weeks period.\nAnd like you say, it wasn't at all fun.\nIt was very tiring and it felt stressful\nbecause I wasn't really in control.\nBut the interesting thing is I now am in a\nsituation where I have\nthe only implementation of an of a Python\nJupyter kernel that\nactually works correctly, as far as I can\ntell, with these new version\n7 protocol improvements.\nAnd now I'm like, well, this is\nfascinating because we don't have a kind\nof a software engineering theory of what\nto do now.\nIt's like, here's a piece of code that no\n1 understands.\nYeah.\nAm I going to bet my company's product on\nit?\nAnd I the answer is I don't know.\nBecause like I I don't I don't like I\ndon't know what to do now,\nbecause no one's like being in this\nsituation, and like, it\ndoes it have memory leaks?\nWill it still work in a year's time if\nthere's some minor change to the protocol?\nIs there some weird edge case that's gonna\ndestroy everything?\nNo 1 knows because no 1 understands his\ncode.\nIt's a really curious situation.\nI mean, first of all, we should\nacknowledge the pernicious erosion of\ncontrol. So at the very beginning, you\nhave 10% AI generated code, and then you\ncan just see how it creeps up and up.\nAnd then at some 0.6 months down the line\nthe PR comes in and now you know 60% of\nthe code is AI generated and do you see\nwhat happens?\nYou just you slowly become disconnected\nbut the bull case for this is you know in\nAI there's this idea called functionalism\nthat you know we don't care what the\nintelligent thing is made out of as long\nas it does all of the right things, then\nwe know, you know, we would say it's AI.\nAnd it's the same thing with software.\nSo the bull case is, I I understand the\ndomain.\nI don't need to write.\nI don't need to know how to write the\nquick sort algorithm.\nI just need to understand it.\nRight? And then and then, you know, so I\njust need to have all of these tests.\nAnd it needs to go into deployment.\nAnd these things need to happen.\nAnd at that point, you know what?\nI don't actually care. And I and I could\nalso And I And to be clear, I quite like\nthat framing.\nBut you know, what that actually does is\nit says, well, software engineering sure\nis important then. Because software\nengineering is all about finding what\nthose pieces are, and how they should\nbehave, and then how you can put them\ntogether to create a bigger piece, and\nthen how you can put them together to\ncreate a bigger piece. And if we do that\nwell, then in 10 years time,\nwe could have software that is far more\ncapable than anything we could even\nimagine today.\nBut\nyou're only gonna get that with really\ngreat software engineering.\nYeah. You wanna be careful.\nI think in the end like IPykernel, I'm\nfinding for example, it's just\ntoo big a piece, right?\nBecause in the end, the the team that made\nthe original IPykernel were not able to\ncreate a set of tests\nthat correctly exercised it, and therefore\nreal world\ndownstream projects, including the\noriginal nb classic, you know, which is\nwhat IPykernel was extracted from, didn't\nwork anymore.\nSo this is this is kind of where our focus\nis on now on the development side at\nAnswer dot ai, is finding the right sized\npieces\nand making sure they're the right pieces.\nKnowing how to recognize what those pieces\nare, and how to design them, and how\nto put them together is actually something\nthat normally requires\nsome decades of experience before you're\nreally good at it.\nCertainly, it's true for me.\nI reckon I got pretty good at it after\nmaybe 20 years of experience.\nYeah. It's a big question.\nIt's like how do you build these software\nengineering chops which are now\neven more important than they've ever been\nbefore, they're the difference between\nsomebody who's good at writing computer\nsoftware and somebody who's not.\nThat feels like a challenging question.\nI know. And there's also this notion that\nthere are so many different ways to\nabstract and represent something.\nKnow, the world is a very complex place.\nAnd maybe the way we've been abstracting\nand representing software is mostly\na reflection of our own cognitive\nlimitations, right?\nAnd even in the sciences and in physics,\nyou tend to have a lot of quite reductive\nmethods of modeling the world.\nAnd then you've got complexity science, is\njust embracing the constructive,\ndissipative, gnarly nature of things.\nAnd I think a lot of software today, we\ndon't understand.\nRight? So for example, there are many\nglobally distributed software applications\nthat use the actor pattern.\nAnd this is just this ins it's basically\nlike a complex system.\nRight? And the only way we can understand\nit is by doing simulations and tests\nbecause no 1 actually knows how all of\nthese things fit together.\nSo you could argue, I guess, as a bull\ncase that maybe we already are doing this\nat the top of software engineering, and\nthat is what we want to do eventually\nanyway. Yeah.\nI'd say probably not.\nYou see companies like Instagram and\nWhatsApp dominate their sectors whilst\nhaving\n10 staff, and beating companies like\nGoogle and Microsoft in the process.\nI would argue this way of building\nsoftware in very large\ncompanies is actually failing.\nAnd I think we're seeing a lot of these\nvery large companies becoming, you know,\nincreasingly desperate.\nAnd, you know, for example, the quality of\nMicrosoft Windows and Mac OS has very\nobviously\ndeteriorated greatly in the last 5 to 10\nyears.\nYou know, back when Dave Cutler was\nlooking at every line\nof the NT kernel and making sure it was\nbeautiful,\nit was a elegant and marvelous piece of\nsoftware, you know.\nAnd this I don't think there's anybody in\nthe world who's gonna say that Windows 11\nis an elegant and marvelous piece of\nsoftware.\nSo I actually think we do need to find\nthese smaller components that we do fully\nunderstand, and that we need to build them\nup.\nAnd here's the problem.\nAI is no good at that.\nSo and and so I say that empirically.\nThey're really bad at software\nengineering.\nAnd then I think that's\npossibly always gonna be true, because,\nyou know, we're\nwe're asking them to\noften move outside of their training data,\nyou know, if we're trying to build\nsomething that literally hasn't been built\nbefore and do it in a better way than has\nbeen done before, we're saying, like,\ndon't just copy what was in the training\ndata.\nSo\nand again, this is a confusing point for a\nlot of people, because they see\nAI being very good at coding.\nAnd then you think like, that's software\nengineering.\nYou know, it's like, it must be good at\nsoftware engineering.\nBut it's they're different tasks.\nThere's not a huge amount of overlap\nbetween them.\nAnd there's no current empirical data to\nsuggest that LLMs are gaining any\ncompetency at software engineering.\nEvery time you look at a piece of software\nengineering they've done, like the\nbrowser, for example, which Cursor\ncreated, or the C\ncompiler, which Anthropic compared\ncreated.\nLike I've read the source code of those\nthings quite a bit.\nChris Latner is much more familiar with\nthe compiler example than me.\nBut they're they're very very obvious\ncopies of things that already exist.\nSo\nthat's the challenge, you know, is if you\nwant to build something that's not just a\ncopy, then you can't outsource that to an\nLLM.\nThere's no theoretical reason to believe\nthat you'll ever be able to,\nAnd there's no empirical data to suggest\nthat you'll ever be able to.\nYes. I think the punch line of this\nconversation is, and I'm sure you would\nagree of this, that we need to have the\ncombination of AI and humans working\ntogether. Right? Because Right. The humans\nprovide the understanding and all of the\nstuff we were saying about knowledge.\nBut we can still use AIs as a tool.\nBut we to design operating models or ways\nof working\nthat make that we say we don't want to\ndiminish our competence\nand understanding. Right.\nSo it's very it's a very fine line.\nThat's that's been our focus, and we both\nfocus on that for teaching and for our\nown internal development.\nThe stuff I've been working on for 20\nyears has turned out\nto be the thing that makes this all work.\nStephen Wolfram should get credit for\nthis.\nHe was the guy that created the notebook\ninterface.\nAlthough also lots of ideas kind of go\nback to Smalltalk and Lisp and\nAPL. But basically, the idea that a human\ncan do a lot\nmore with a computer when the human can,\nlike,\nmanipulate the objects in inside that\ncomputer in real\ntime, study them, and move them around,\nand combine them together.\nYeah. That's what small talk was all\nabout, you know, with objects, and APL was\nthe same with arrays.\nMathematica basically is a super powered\nLisp, which then also added on this very\nelegant notebook interface that allowed\nyou to construct kind of a living document\nout of all this. So I built this thing\ncalled nbdev a few years ago, which is a\nway of creating production software inside\nthese\nnotebook interfaces, inside these rich\ndynamic environments.\nAnd I found that made me dramatically more\nproductive as a programmer.\nAnd like today, even though I've\nnever been a full time programmer as my\njob, when you look at my kind of GitHub\nrepo output, I think GitHub produced some\nstatistics about it, and I was like, just\nabout the most productive programmer in in\nAustralia.\nYou know, like it it's working.\nAnd a lot of the stuff I build has lots\nand lots of people use it,\nbecause it's such a rich, powerful way to\nbuild things.\nAnd so it turns out, we've now discovered\nthat if you put AI in the same\nenvironment with a human, again, in a in a\nrich, interactive\nenvironment,\nAI is much better as well, which perhaps\nisn't shocking to hear.\nBut the normal, like, you use Claude code,\nwhich I know you do and it's a very\ngood piece of software, but the\nenvironment we give Claude code is very\nsimilar to the environment that people had\n40 years ago, you\nknow. It's a it's a line based terminal\ninterface.\nYou know, it can use MCP or whatever.\nMost of the times, it just nowadays uses\nbash tools, which again, very powerful.\nI love bash tools. I use them all the, you\nknow, the CLI tools all the time.\nBut it's still just it's using text files,\nyou know, as its as its interface to the\nworld. It's it's it's really meager.\nSo so we put the human and the\nAI inside a Python interpreter.\nAnd now suddenly you've got the full power\npower of a very elegant\nexpressive programming language that the\nhuman can use to talk to the AI.\nThe AI can talk to the computer.\nThe human can talk to the computer.\nThe computer can talk to the AI.\nLike, you have this really rich thing,\nthen we let the human and\nthe AI in real time build tools that each\nother can\nuse. And that's what it's about to me.\nRight? It's about, like, creating an\nenvironment where humans can\ngrow and engage and share.\nIt's like for me, when I use Solveit, it's\nthe opposite of that experience you\ndescribed with Claude Code.\nAfter a couple of hours, I feel energized\nand\nhappy and fulfilled.\nI'll give you my take.\nI think that the thing that you're\npointing to here is there's something\nmagic about having an interactive,\nstateful environment that gives\nyou feedback. And that is because our\nbrains kind of they can do\na certain unit of work.\nSo we actually think through refining and\ntesting with reality.\nAnd that's why, I mean, during my PhD, I\nused Mathematica and MATLAB.\nAnd I agree. So we've got this REPL\nenvironment and, you know, here's the\nmatrix, let's do an image plot, you know,\ndo a change.\nThis is what it looks like now.\nAnd it's actually a wonderful way to kind\nof just just refine my mental\nmodel about something. But Claude code\ndoes a lot of this stuff.\nI think it's mostly a skill issue.\nI think the people that use Claude codes\neffectively do this.\nI've written a content management It's\npossible.\nIt is possible. It is possible.\nYeah. So I've written a content management\nsystem called Rescript.\nAnd when I'm putting together documentary\nvideo, it can pull transcripts.\nAnd then I can verify the claims.\nAnd part of AI literacy is just\nunderstanding the asymmetry of language\nmodels, right? So when you give them a\nsort of discriminative task, they're\nactually quite good. So if I tell it in a\nsub agent to go and verify every\nindividual claim, it's much more accurate\nthan if I was in generation mode and I was\ngenerating a bunch of claims.\nAnd the stateful feedback thing, again, I\ncan have some kind of schematized XML\ndump. And I can have like, an application\nhere on the side, which is visualizing,\nand it's like a feedback loop.\nAnd for me, this is an AI literacy thing.\nLike, the the good people at AI are\nalready doing this.\nYeah. So I don't fully agree with you.\nI agree you can do it in Claude code, and\nI agree it is a AI\nliteracy thing as to whether you can.\nBut also Claude code was not designed to\ndo this.\nIt's not very good at it, and it doesn't\nmake it the natural way of working with\nit. I don't wanna say it's an AI literacy\nproblem, because that's like saying like,\noh, it's a you problem.\nTo me, if a tool is not making it the\nnatural way\nfor a human to become more knowledgeable,\nmore\nhappy, more connected, with a deeper\nunderstanding and\na deeper connection to what they're\nworking on, that's a tool problem.\nThat that should be how tools are designed\nto work.\nSo so many models and tools expressly are\nbeing evaluated\non can I give it a complete piece of work\nand have it go away and do the whole\nthing, which feels like a huge mistake to\nme,\nversus have you evaluated whether a human\ncomes out the other end with a\ndeep understanding of a topic, you know,\nso that they can really easily\nbuild things in the future? I agree with\nall of that.\nBut then there's the other interesting\nangle, which is that there was a famous\ntalk by Joel Grus, and we'll talk about\nthis.\nAnd and he said that notebooks are\nterrible.\nThey're they're really bad from a software\nengineering point of view.\nAnd and at the time, and maybe still now\nto a certain extent, I I agree with him\nbecause,\nyou know, I've I've I've done ML DevOps.\nI've worked in large organizations, you\nknow, like trying to figure out how do we\nbridge like data science and software\nengineering.\nAnd Claude code is already more towards\nthe software engineering side and what\nthat means is it creates idempotent,\nstateless, repeatable\nartifacts, right? So as you say, from a\npedagogical point of view, it's really\ngood having this stateful feedback because\nI can understand what's going on.\nBut then I need to translate that into\nsomething which is deployable.\nAnd Can you tell us the story of you you\nyou responded to Joel Grus, didn't\nyou? And and it was a bit of a fiasco,\nwasn't it?\nBut what just just tell us about that\nstory.\nHe did a really good video called I don't\nlike notebooks.\nIt was hilarious.\nIt was really well done.\nAnd, yeah, I was totally wrong.\nAnd all the things he said notebooks can't\ndo, they can.\nAnd all the things he said you can't do\nwith notebooks, I do with notebooks all\nthe time. So it was a very good, amusing\nincorrect talk.\nSo then I did a kind of a parody of it\ncalled I like notebooks,\nin which I basically copied with credit\nmost of his slides and\nshowed how every 1 of them was totally\nincorrect.\nBut, like, I actually think your comment\nabout it\ndoes come down to the heart of it, which\nis this this this difference between,\nlike,\nhow software engineering is normally done\nversus how\nscientific research and similar things is\nnormally done.\nAnd I think and I agree there is a\ndichotomy there.\nAnd I think that dichotomy is a real shame\nbecause I think software development is\nbeing done wrong. It's being done in this\nway, which is, yeah, all about\nreproducibility and these like dead these\ndead\npieces, you know.\nIt's all dead code, dead files.\nI will never be able to express this 1\nmillionth as clearly as Brett\nVictor has in his work, so I'd encourage\npeople who haven't watched Brett\nVictor to to watch him.\nBut, you know, he he shows again and again\nhow a\ndirect connection, you know, a direct\nvisceral connection with\nthe thing you're doing is is all that\nmatters, you know.\nAnd that's his mission, is to make sure\npeople have that connection.\nAnd that's basically my mission as well.\nSo for me, traditional software\nengineering is as far from that as it is\npossible to get. I think it's I think it's\ngross.\nLike, I I I find it disgusting.\nAnd I find it sad that people are being\nforced to work like that.\nIt's like, I think it's inhumane.\nAnd I just don't think it works very well.\nI mean, empirically, it doesn't work very\nwell.\nAnd it's much less good for for AI as\nwell, as it's much less good for humans.\nIt hasn't always been that way.\nLike, you know, with with Alan Kay and\nSmalltalk and\nIverson and APL, you know, Lisp\nWolfram with Mathematica.\nTo me, these were the golden days when\nwhen\npeople were focused on the question of how\ndo we get the human\ninto the computer to work as closely with\nit as possible.\nYou know, that's where the the mouse came\nfrom, for example.\nI've got to like click and drag and\nvisualize entities in your computer as\nthings you can move around.\nSo I feel like we've lost that.\nI think it's really sad.\nYeah. With Claude code and stuff, the the\ndefault way of working with them\nis to go super deep into it.\nIt's like, okay, there's a whole folder\nfull of files, you never even look at\nthem, your entire interaction with it is\nthrough a prompt.\nYeah. I it it literally disgusts me.\nLike, I literally think it's it's\ninhumane, and it's my\nmission remains the same as it has been\nfor, like, 20 years, which is to stop\npeople working like this.\nI I know. But the so casting my mind back.\nI used to work with data scientists.\nThey were using Jupyter Notebooks. And\nwhat I found was typically, I mean, back\nthen, you couldn't if you check them into\nGit, it wouldn't look very good.\nMost of these data scientists didn't know\nhow to use Git.\nThey would run the cells out of order,\nwhich means it wouldn't be reproducible.\nThere are all sorts of things like that.\nThe thing is, I agree with you that you\nyou can use them in this in this workflow.\nBut it comes back to what I was saying\nbefore about, you know, we we were talking\nabout the call center and it being like a\nlow intelligence job.\nYou know, the data scientists, the reason\nwhy they they are doing intelligent work\nis they are creating something that\ndoesn't exist.\nThey are figuring out the contours of a\nproblem.\nThey're actually working in a domain that\nis poorly understood.\nBut you could argue now the bull case is\nwhen the data scientists can succinctly\ndescribe the contours of the problem,\nmaybe we could go to Claude code, and we\ncould implement it properly. But how do we\nbridge between those 2 worlds?\nI think that would be a terrible, terrible\nidea.\nLike,\nyou don't wanna remove people from their\nexploratory environment,\nyou know?\nResearch and science is developed by\npeople building insight, you know?\nWhoever you listen to, you know, whether\nit be Feynman or whatever, like, you\nalways hear from the great scientists how\nthey build deeper\nintuition by by building mental models,\nwhich they get over time, by\ninteracting with the things that they're\nlearning about.\nAnd like in Feynman's case, because it was\ntheoretical physics, he couldn't actually\npick up spinning quark, but he did\nliterally study spinning\nplates, you know.\nYou gotta find ways to to deeply interact\nwith with what you're working with.\nLike, so so many times I've seen data\nscience teams because you're right.\nData science teams aren't very familiar\nwith Git and aren't very\nfamiliar with things that they do need to\nunderstand.\nAnd so often, I've seen a software\nengineer will become their manager, and\ntheir fix to this will be to tell them all\nto stop using Jupyter notebooks.\nAnd now they have to use all these\nreproducible blah blah virtual, you know,\nvirtual end blah blah. They destroy these\nteams over and over again.\nI've seen this keep happening.\nBecause the solution is not create more\ndiscipline and bureaucracy.\nIt's solve the actual problem.\nSo for example, we we built a thing called\nan n\nb merge driver,\nwhich so a lot of people don't realize\nthis, but actually notebooks are extremely\ngit friendly. It's just that git doesn't\nship with a merge driver for them.\nSo git only ships with a merge driver for\nline based text files.\nBut it's fully pluggable.\nAnd so you can easily plug in 1 for JSON\nfiles instead.\nAnd so we wrote 1.\nSo now when you diff, you know, when you\nget a git diff with our\nmerge driver, you see cell level diffs.\nIf you get a merge conflict, get search\nlevel cell level merge conflicts.\nThe notebook is always openable in\nJupyter.\nNBDime did the same thing, so 2\nindependent implementations of\nthis. So yeah, there were problems to\nsolve, you know, but the solution\nto it was not\nthrow away Brett Victor's ideas and make\npeople further away from from\ntheir exploratory tools.\nBut to fix the exploratory tools, and I\nthink all software developers\nshould be using exploratory based\nprogramming to deepen their\nunderstanding of what they're working\nwith.\nSo that they end up with a really strong\nmental model of the system that they're\nbuilding and they're working with.\nAnd then they can come up with better\nsolutions, more incrementally, better\ntested. I basically never have to use a\ndebugger, because I basically never have\nbugs. And it's not because I'm a\nparticularly good programmer, it's because\nI build things up small little steps, and\neach step works and I can see\nit working and I can interact with it.\nSo there's no room for bugs, you know?\nYou know, I'm so torn on this because I\nagree with you.\nAnd I'm also skeptical of people who say\nthat organizations,\nthey they they converge onto ways of doing\nthings, they no longer need to evolve.\nThey no longer need to adapt. Innovation\nis adaptivity, right?\nAnd we should increase the surface area of\nadaptivity as much as we possibly\ncan. So we need people that are constantly\ntesting new ideas, finding these\nconstraints. But by the same token, we\nneed to use the cloud, we need to use\nCICD. We need to get this stuff into\nproduction.\nYeah. So do you do you do but, like,\nthere's absolutely no like, so\nNB dev ships with out of the box CI\nintegration,\nand that's like the tests are literally\nthere,\nlike, because the source is a notebook,\nthe entire exploration of like, how does\nthis API work, you know, what does it look\nlike when you call it, The implementation\nof the functions, the examples of them,\nthe documentation of them, the tests of\nthem are in 1 place.\nSo it's much easier to be a good\nsoftware engineer in this environment.\nSo, yeah, like, do do both, you know.\nSo do you remember there was there was\nthat existential risk should be an urgent\npriority, and it was signed by folks like\nHinton and and Demis\nHassabis. And you responded, basically\nwith a rebuttal, and that was with Arvind,\nyou know, the the snake oil guy.\nTell me about that. Do you think we should\nbe worried about AI existential risk?\nI mean, that was a certain time, wasn't\nit?\nAnd I feel like things have changed a bit.\nThank God. I feel like we we, not just\nbeing Aravind,\nbut broadly speaking, the community of\nwhich we're a part, kind of probably\nwon that. Now we have other problems to\nworry about.\nBut you know, basically, at that point,\nthe prevailing narrative was\nAI is about to become autonomous.\nIt could become autonomous at any moment,\nand could destroy the world.\nSo it very much comes from,\nyou know, Alizia Yukowski's Yep.\nWork, which\nI think clearly has been shown to be wrong\nat many levels at this\npoint. That they would refute that,\nobviously.\nOf course, they would. Yeah. It's 1 of\nthose things that they can always refute\njust like any doomsday cult unless you\ngive it a date and the date passes.\nWell, I've I've updated a little bit in\nthe sense that I I now think I I would\nnow say that these models can be said to\nbe intelligent in restricted domains.\nThe ARC challenge showed that.\nSo if you place constraints into the\nproblem, you you can you can go faster\ntowards a known goal. Even agency, you you\ncan put a planner on that and you can go\nif you if you know where you're going, you\ncan go there faster.\nBut that doesn't help you. Like, you can\nhave all the intelligence and agency in\nthe world. But if you don't have the\nknowledge and the constraints, then you're\ngoing in the wrong direction faster.\nAnd I think they don't seem to appreciate\nthat these models don't actually know the\nworld. Like, none of that was even\nrelevant to Ivan and my point,\nwhich was and is, that it's\nmisunderstanding\nwhere the actual danger is.\nYep. Which is that when you have a\ndramatically more powerful\ntechnology\nentering the world,\nthat can make some people dramatically\nmore powerful,\npeople who\nare in love with power will seek to\nmonopolize that technology.\nAnd the more powerful it is, the more\nstrong that\nurge from those power hungry people will\nbe.\nSo to ignore people so here's the problem.\nIf you're like, I don't care about any of\nthat.\nAll I care about is autonomous AI taking\noff, you know,\nsingularity paper clip, nano goo,\nwhatever.\nThe obvious solution to that is, oh, let's\ncentralize power.\nAnd this is which is what we kept seeing,\nparticularly at that time.\nLet's give either very rich technology\ncompanies, or\nthe government, or both, all of this\npower, and make sure\nnobody else has it.\nIn in my threat model, that's the worst\npossible thing you can do, because you've\ncentralized the ability to control in 1\nplace.\nAnd therefore, these people who are\ndesperate for power just have to take over\nthat thing. Could could we distinguish,\nthough, what you mean by power?\nBecause we've we've just spent some of\nthis conversation talking about how it's\nnot actually as powerful as people think\nit is.\nBut I'm I'm not even that's what but mine\nis an even if thing.\nRight? So, like, I I I just say, even if\nit turns out\nto be incredibly powerful.\nRight, like, don't I don't even wanna\nargue about whether it's gonna be powerful\nbecause that's speculative.\nEven if it's gonna be incredibly powerful,\nyou still shouldn't centralize all of that\npower in the hands of 1 company or the\ngovernment.\nYeah. Because if you do, all of that power\nis going to be\nmonopolized by power hungry people, and\nused\nto\ndestroy civilization, basically.\nYou'll end up with a case where all of\nthat wealth and power will\nbe centralized with the kinds of people\nwho who want it centralized.\nSo, like,\nsociety for hundreds of years have faced\nthis again and again and again, you know.\nSo when it's like, you know, writing used\nto be\nsomething that only the most\nexclusive people had access to knowing\nabout writing.\nAnd the same arguments were made.\nIf you let everybody write, they're gonna\nuse it to write things that we don't want\nthem to write, and it's gonna be really\nbad, you know.\nDitto with printing, ditto with the vote,\nlike\nlike and again and again, society has to\nfight against this\nnatural predilection of the people that\nhave the status quo power to be like,\nno, this is a threat.\nSo when we're saying like, okay, what if\nAI turned out to be\nincredibly powerful?\nWould it be better for society to be that\nto be kept in the hands of a few\nor spread out across society?\nMy argument was the latter.\nNow, there's also an argument which is\nlike, hey, don't worry about it.\nIt's not gonna be that powerful anyway.\nI I just didn't wanna go there, because\nit's not\nan argument that's easy to win, because\nyou can't really say what's gonna happen.\nWe're all just guessing.\nBut I can very clearly say, like, well, if\nit happens, would it be a really good\nidea to only let Elon Musk have it?\nOr would it be a good idea to only let\nDonald Trump have it?\nDan Hendricks spoke about this offense\ndefense asymmetry.\nUh-huh. So it's actually very important\nfor us to have to have countervailing, you\nknow, defenses. But let let's just take\nthat as a given for a minute.\nBecause obviously, when we look at\nsomething like Meta and Facebook, it's\nit's quite clear what the power imbalance\nis, you know, they they they control all\nof our data. They they know what we're\ndoing.\nWith with something like OpenAI and\nClaude, so it's not as good as we thought\nit was because actually humans still need\nto be involved.\nBut for example, they have all of our\ndata, right?\nAnd you might be working on some new\ninnovative technology and you're using\nClaude and you're sending all of your\ninformation up there and they can now copy\nyou. Mean, what what kind of risks are you\ntalking about to be more concrete?\nYeah. No. I mean, so I was not talking\nabout any of those things, right?\nSo at the time, I was talking about this\nspeculative question of what if AI gets\nincredibly powerful.\nWell, I mean, like like now, for example,\nthey they say that this is the new means\nof production. And that's that seems\ncompletely hyperbolic to me.\nBut, like, in your best estimation now, if\nthere are risks, what are they?\nIf there are risks with the current state\nof technology, I mean, I think some of\nthem are the ones we've discussed, which\nis\npeople enfeebling themselves by\nbasically losing their ability to be to\nbecome more competent over time.\nThat's that's that's the big risk I worry\nabout the most.\nThe privacy risk, it's there, but I'm not\nsure it's much\nmore there than it was for Google and\nMicrosoft before.\nLike, you know, you used to work at\nMicrosoft.\nYou know how much data they have about the\naverage\nOutlook, Office, etcetera, user.\nDitto for Google, you know, the average\nGoogle Workspace or Gmail user.\nThose privacy issues are real.\nAlthough I think there are bigger privacy\nissues around these companies which\nthe government can outsource data\ncollection to.\nSo back in the day, it used to be\ncompanies like ChoicePoint and Acxiom.\nNowadays, it's probably more companies\nlike Palantir.\nThe US government is actually prohibited\nfrom\nbuilding large databases about US\ncitizens, for example.\nBut it's not prohibited companies are not\nprohibited from doing so, and the\ngovernment's not prohibited from\ncontracting things to those companies.\nSo, I mean, that's a huge worry, but I\ndon't think it's 1 that AI is\nuniquely creating. It certainly so, like,\nyou're in The UK.\nAs you know, in The UK, surveillance has\nbeen universal for quite a while\nnow. It certainly makes it easier to use\nthat surveillance.\nBut a sufficiently well resourced\norganization could\njust throw 1000 bodies at the problem.\nSo, yeah. I'm not sure these are due\nprivacy problems. There's maybe more\ncommon ones than they used to be.\nYeah. Jeremy, I've just noticed the time.\nI need to get to the airport. Alright.\nThis has been amazing. Thank you, sir.\nThank you for coming. Yeah. Hope you had a\nnice trip.\nThank you so much.",
  "transcript_chars": 79831,
  "ingested_at": "2026-05-12T00:40:22.519880+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 155989,
    "like_count": 4272,
    "channel_id": "UCMLtBahI5DMrt0NPvDSoIRQ",
    "categories": [
      "Science & Technology"
    ],
    "tags": []
  }
}