A masterclass from Jeremy Howard on why AI coding tools can be a trap -- and what 45 years of programming taught him that most vibe coders will never learn. - AI coding tools exploit gambling psychology - The difference between typing code and software engineering - Enterprise coding... AND prompt-only vibe coding are "inhumane" i.e. disconnecting humans from understanding-building - AI tools remove the "desirable difficulty" you need to build deep mental models. Out on MLST now!show more

@jeremyphoward The telling thing in these comments is how MAD people get that Jeremy is saying what he is saying. Portraying him as being anti-LLM, which he is not. Like gamblers, so many people emotionally attached to their AI right now

@jeremyphoward Every generation says the new abstraction will destroy understanding, what’s different here?

Given Grady didn't take the bait - I will answer! LLMs *break* this pattern (on average). You're not moving to a higher level of understanding imo -- you're (potentially) moving to NO understanding of the implementation. This doesn't always happen, but it seems to happen more to people who go in without domain understanding/expertise. As Jeremy eludes to in the interview, there is a strategy to mitigate this and stay mentally "tuned-in". Roughly speaking, it is about adopting a development (and usage) workflow where you have iterative, interactive, stateful feedback exposing and reinforcing the abstractions/interfaces of the application. Just so it doesn't sound like I'm chatting random bullshit, here's an actual example that I used to create the very video you watched above. It's an agentic application, so I built this timeline by just chatting with my agent. But every single iteration - I had rich visual stateful feedback using the interface of the application and its abstractions. I could iterate many times, and it was all reproducible / immutable history. So rather than doing a bunch of half-baked stuff with LLMs and having no idea what the current state is, I'm always completely tuned in and everything is validated etc. This is an AI literacy thing as we discussed in the interview, it's possible to use AI in this way and build understanding (as I do).

I think both things can be true. People who use these tools badly or lazily will end up with shallow understanding, but it still feels similar to previous abstractions. Plenty of developers wrote C software without understanding assembly, and Python without understanding C. Yet engineering still moves forward. Another relevant to you Tim is YouTube. Legacy media grumbled its low quality, production standards would collapse etc. The average quality dropped, but the creative surface exploded and those that do good work outperform the old gatekeepers. I’m a non-engineer (with domain expertise) and over Dec/Jan built an app and already have a contract with an enterprise CTO ( 🤯) Something that would’ve been unthinkable for me just six months ago. Through the development I’ve learnt so much across the stack I’d never have imagined, hence my understanding has increased significantly. These magical tools should be embraced and encouraged.

I'm as excited about this tech as you are, and I am glad you are building great things! "Plenty of developers wrote C software without understanding assembly" It didn't matter, they just understood a new representation instead at a higher level of abstraction, which could still do everything we needed to do. The whole point of abstraction is to screen off unnecessary detail, and intelligence enacted in engineering (done right) creates these new, more powerful abstractions. But we still need to be in possession of good abstractions! A good abstraction should be maximally unified and factored, without screening off too much detail. This is what David Krakauer refers to as doing "more with less" (see our interview). Quite often, when we use language models, we do "more with more". Which means we're just creating loads of fractured, discardable code that doesn't cohere well together, and complexity explodes. I'm sure the stuff you're building is great, but deep domain knowledge might prompt you to design in better abstractions early on, ask the language model important questions -- for example, reviewing the security architecture and making sure you have a really good set of interfaces (so you can independently scale different parts of the application). Right now, the knowledge is probably all in your head, and you might struggle to scale through other people without your intentions and design decisions being eroded. Not trying be negative here but you will probably hit on some of the stuff. The good thing, though, is once you develop better abstractions in your head, you can just throw the code away and start again!

@jeremyphoward Your podcast and guests give me a similar feeling. I don’t understand all of it, but I try, and I feel thrilled/excited, and in the case of David Krakauer, erect.

@jeremyphoward I have coded for 20yr. I have been CTO. I disagree. Agentic coding is a new era. There is still "desirable difficulty". There is much more. Agentic coding has make coding fun again.

@jeremyphoward I agree that there can be desirable difficulty (so does Jeremy) - it's mostly fighting against laziness and using AI in the right way. These tools are magical for experienced software engineers so you're probably already doing it the "right way" 😁

Incredibly strange take. He has a point, in that human context is far greater than we give it credit for, but for specific tasks, namely writing code, language models are a powerful accelerant. A carefully designed spec with appropriate guardrails can cut development time down to a fraction of what it once was.

@jeremyphoward ngl the gambling psychology point is so real. you get one perfect output and suddenly you're just hitting accept on everything hoping for another hit. the devs who treat AI as a collaborator instead of a slot machine are the ones actually shipping

@jeremyphoward Yes yes we get it. You dont like LLMs. Still persisting with the narrative makes the whole channel seem like Doomer content.

@jeremyphoward I love LLMs, and I'm maxing out my Claude Max sub. For some reason, my Claude Code stats reset in January, but it's been like this continuously 😀

@jeremyphoward This was very good. Jeremy one of the few sane voices on this unfortunately.

the desirable difficulty thing is so real: seen tons of devs completely freeze when copilot can't figure something out because they never actually learned how things work under the hood.... and yeah the gambling psychology hits hard, you just keep hitting regenerate thinking "this next one will be it" :))

GTC is coming, the premier AI conference, great opportunity to learn about AI. NVIDIA and partners will showcase breakthroughs in physical AI, AI factories, agentic AI, and inference, exploring the next wave of AI innovation for developers and researchers. Register for virtual GTC for free, using my link and win NVIDIA DGX Spark (

@jeremyphoward Jeremy is a legend, but he is at heart an academic and teacher, not great at business. so when he draws the parallel to pulling the gambling lever with these tools and betting the product on these tools... well that's exactly what business is. Risk, and if it's not gtfo

@jeremyphoward Howard's right that understanding matters more than speed. But I've used Claude Code daily for a year and I still read every line it writes. That review is where I learn the most. "Desirable difficulty" and better tools aren't mutually exclusive.

@jeremyphoward People are mad in the comments because a person with over 20 years of experience in programming and machine learning said what they think about LLMs. Those “people” are probably bots, and if they’re human, they weren’t even born when Jeremy first started coding, respectfully.

@jeremyphoward

@jeremyphoward Cope harder

@jeremyphoward weird take from someone who cofounded which abstracted away the difficulty and requirement for low-level understanding from working with NNs. This is just another layer up. AIs coding won't kill true curiosity, nor will forced coding drudgery elicit it

@jeremyphoward 20 years to be good at coding? Really?

@jeremyphoward 45 years of programming = knowing where the bodies are buried. the trap isn't AI. it's skipping the part where you understand why the code works.

@jeremyphoward Totally agree with this. But I cannot resist the temptation of vibe coding in hours what I took days. Well I know I know that itself is akin to gambling :(

That bit about typing code vs software engineering is spot on. I spent 8 months building Drop with Claude and yeah, the prompting part was easy. The hard part was understanding why the geofencing broke at scale, how to architect real-time data pipelines, debugging production issues at 2am. AI was the tool but I had to learn what good architecture actually looks like. Drop's a free iOS app now:

Love Jeremy persona, realy sparks a good discussion. From the positive notes on the matter - I feel a bit more unstucked with AI sometimes, and def can iterate faster on solutions (not incrementing). What scares me is the fact that we are producing so much code that it doesn't leave neural marks in our heads thus we don't even remember commiting certains lines of code. Engineers were most often domain experts and now if they don't ask AI with the right context incjected they don't even know what's in the codebase, and can hardly answer domain specific questions. Also learning process for junior devs suffers - meaning you can prompt your AI or spend 2 months perfecting a craft - choice is simple from our dopamine seeking brains.

@jeremyphoward Is the full interview available somewhere?

@jeremyphoward @MLStreetTalk The gambling psychology angle is spot on. AI coding gives that first-run dopamine hit, then the debugging spiral hits hard.

@jeremyphoward That line about typing code versus software engineering is where ROI shows up, not in prompt hype

@jeremyphoward vibe coding works until u hit a bug u have no framework to debug. then u realize u skipped all the parts that actually matter.

@jeremyphoward Need this counterpoint to define the drawbacks one feels after some time. It’s not about deep models imo but alignment between programmer intent and where tool takes you instead and difficulty bridging the two vectors.

@jeremyphoward @IncrementsPod @VadenMasrani @BennyChugg did you guys listen to this? Jeremy's got some interesting takes on AI, creativity, problem solving etc, right up your alleys!

the craft is collaboration after working in the ‘field’ for several years can attest that it is challenging AI is of course based on probabilities resemblance to an advanced game of chance is apropos suggesting that AI is an obstacle to developing a deep mental model is part of the friction that makes the collaboration challenging it motivates you to indulge in shortcuts discipline and structure is part of that collaboration

@jeremyphoward This gentleman is not realizing he’s exploiting a mirrored psychological pattern and is also gambling.. 1: He doesn’t understand his own tacit knowledge. 2: He doesn’t understand how LLMs work. 3: He doesn’t understand the gap is the difference.

@jeremyphoward This one was epic. Really liked it.

@jeremyphoward In large language model defense, most people in the corporate world pretend to understand the subject. Few actually have a clue about the topics they claim expertise in. Most projects are run by hustlers. Very little of what gets shipped reflects any deep thinking whatsoever.

I would say there is a range of nuance that causes people to agree or disagree based on their own experiences. Some use it like a slot machine yes, the pure “I don’t look at anything because why should I, I don’t know any of it any way” people. The “Now I don’t have to learn anything at all” narrative is so off putting. Because to them everything including even watching it is “a chore.” But this would be like doctors offloading 100% to AI because “it’s all a chore.” But the other side of this, that I like to talk about is all the potential learning. I have and continue to learn a lot because I choose to use the coding agents to discover and be curious. For example I’ve learned a ton about swift in one of my projects. I learned how to use some tools even better. I learned how to be more clear in my asks. I’ve learned to rethink/refine how I approach things. I think there’s a win on that side people don’t talk about. Because the loudest voices are the slot players and the “finally those SWE thought they were so special now they’re min wagers” people.

@jeremyphoward clueless

@jeremyphoward I actually built something on this line of thought

@jeremyphoward If you have a greenfield project with unlimited budget, no time constraints, and zero turnover, then hand crafting it is clearly the way to go. If your competitor is using AI and you're not it's survival of the shippest.