Several weeks ago, I spent about a week fully reverse engineering a Stereomaker pedal. It accepts a mono signal and produces a stereo field using a 5-stage all-pass filter to mess with the phase without the use of delay (which sounds cheesy and creates a result that doesn't mix well back to mono).
I've not really worked with audio circuits previously, and I'd been intimidated to approach the domain. My journey was radically expedited by iterating through the entire process with a ChatGPT instance. I would share zoomed photos, grill it about how audio transformers work, got it to patiently explain JFET soft-switching using an inverter until the pattern was forced into my goopy brain.
Through the process of exploring every node of this circuit, I learned about configurable ground lifts, using a diode bridge to extract the desired voltage rail polarity, how to safely handle both TS and TRS cables with a transformer, that transformer outputs are 180 degrees out of phase, how to add a switch that will attenuate 10dB off a signal to switch line/instrument levels.
Eventually I transitioned from sharing PCB photos to implementing my own take on the cascade design in KiCAD, at which point I was copying and pasting chunks of netlist and reasoning about capacitor values with it.
In short, I gave myself a self-directed college-level intensive in about a week and since that's not generally a thing IRL, it's reasonable to conclude that I wouldn't have ever moved this from a "some day" to something I now understand deeply in the past tense without the ability to shamelessly interrogate an LLM at all hours of the day/night, on my schedule.
This is a phenomenal example of exactly what I am advocating.
Notice you didn't ask the AI to 'just design a stereo pedal for me.' You interrogated it, reasoned about netlists, and forced the concepts into your brain through intense friction. That is pure deep work.
Throughout the reverse engineering process, the LLM and I both were expecting each op-amp stage to use the next ladder value capacitor. We'd talked ourselves into how and why that would make sense.
At the end I was curious enough that I desoldered those five caps and realized that they were all 2.2nF except for the last stage which was 1nF.
I brought that news back to the LLM and we realigned our understanding of how the effect was achieved, ultimately coming to realize that our approach would have created notches at different frequencies instead of just shifting the phase by about 900 degrees.
It was an incredible learning experience. I try hard not to personify LLMs but this really did feel like working side by side with a friend on a problem until it was solved.
IRL, I suspect that most people who would be able to tackle that challenge with me lack both the time and patience to actually do it.
No, you touch on the aspects where you're able to use AI as an extension of your skills.
This is completely different than my colleague who isn't a software engineer, and now all of the sudden is creating PRs which I need to review and correct.
I'm a sceptic. I use it to explore the unknowns and go from there.
My reaction to this question is that it might technically be in good faith, but you're pushing it.
Let's say that LLMs didn't exist, and I learned these same skills in an oddly specific hands-on workshop, or from an oddly specific textbook, or fuck it, let's say that I hired some greybeard pedal designer to just sit beside me and answer all of my stupid questions without judgement for a few weeks at their hourly rate.
Would you feel compelled to challenge whether I had retained what I learned or inexplicably woke up this morning, tabula rasa, and realized that I'd forgotten everything I spent a week teaching myself? I honestly don't think that you would.
For the record, I could reimplement any part of the circuit on demand if I needed to. I might be tempted to look at my notes for the JFET switching because it was genuinely hard to keep in my head, but that's more of a confidence thing than a "shit, I forgot how op-amps work" thing.
I've since implemented a variation into a matrix mixer concept that I'm working on, when it detects that a TS cable has been inserted into a TRS jack.
> Would you feel compelled to challenge whether I had retained what I learned
Yes, the exact same way I would dubious when someone says they learned much from following a youtube tutorial or participating in a two week workshop or something
It's a 90 minute video that will take you a week to watch if you're doing it properly.
Seriously though... you don't learn from watching a video tutorial (which you can slow down and re-watch as many times as you need) and you apparently don't believe you can learn from an LLM which will patiently answer literally infinite questions, no matter how basic or repetitive... would you mind clarifying how you do learn?
Everyone has different learning styles so I tend to take a different strokes for different folks attitude. For example, I don't absorb highly technical stuff from books and the idea of [paying to be in a] classroom where you're forced to endure 95% what you're not interested in to get the 5% you care about (at the speed of the dumbest student in the room) gives me hives.
Yet, it kind of sounds like you might just be arguing for argument's sake. Also, you can learn A LOT in two weeks if you're motivated.
> Everyone has different learning styles so I tend to take a different strokes for different folks attitude
Okay but at the end of the day the only way to actually learn (and demonstrate that you've learned anything) is by actually doing it
And I don't really consider "I got the AI to do it" as actually doing it, which is why I'm questioning what you've actually retained.
To be clear if you feel like you've actually learned this stuff then good for you. I'm genuinely happy if that's the outcome you feel you have obtained
I'm just personally very skeptical of anyone learning fuck all from using AI to build stuff because like I said... I learn from practice. Using AI is not practice any more than copying from open source repos is.
And frankly I'm bitter because I absolutely cannot learn fuck all from using AI. It is the sort of shortcut that prevents my brain from committing anything to memory.
I guess I don't understand what about any aspect of what I explained gave you the impression that I might not have put what I learned into practice.
This is going to sound like I'm fucking with you, but I'm deadly serious: if someone taught you how to do something and you later learned that that person was actually an LLM masquerading as a human, would you forget what you had learned?
It's actually not impossible that you've hypnotized yourself, or could be experiencing a trauma response.
LLMs absolutely let you explore ideas and areas you wouldn't have otherwise...but does your new design actually _work_?
I'm curious whether the "knowledge" you gained was real or hallucinatory. I've been using LLMs this way myself, but I worry I'm contaminating my memory with false information.
I think that you're confusing what you're doing with what I'm doing.
What I'm doing is learning the circuit constructs that I need and then putting them to work in real circuits. There's usually a few breadboard steps in the middle, which you could call reinforcement learning.
To me, the telling thing about your question is the implication that I would spend a week learning how to do something and then not test it out. I know that this reply reads as salty, but I'm really struggling to contain my own "wtf" on this end.
Seriously, people that are so determined to prove that LLMs don't work despite how easy it is to test for yourself and see that they clearly do work are the ones that are hallucinating.
I have a nearly total opposite take. I can't tell you how many times I've read a book, a paper or something else and been confused by some ambiguity in the author's prose. Being able to drop the paper (or even the book!) into an LLM to dig into the precise meaning has been an unbelievable boost for me.
Now I can actually get beyond conceptual misunderstanding or even ignorance and get to practice, which is how skills actually develop, in a much more streamlined way.
The key is to use the tool with discipline, by going into it with a few inviolable rules. I have a couple in my list, now: embrace Popperian falsifiability; embrace Bertrand Russell's statement: “Everything is vague to a degree you do not realize till you have tried to make it precise.”
LLMs have become excellent teachers for me as a result.
An LLM can't invent meaning in a text where there is none. It's equivalent to CSI's classic "zoom, enhance" on resolution limited photographs. You need to consider you're learning a load of rubbish from LLMs.
Right question to ask, however, good readers/professionals do have some sense for this and ability to dig further as needed. On the other hand, books and articles are often over-detailed, with the key stuff buried in the lede or even remaining tacit.
For me, LLMs have often pointed me to answers or given food for thought that even subject matter experts could not. I do not take those answers at face value, but the net result is still better than the search remaining open-ended.
By asking it to cite its sources. Whenever I use AI, I have it pull direct quotes from the text to justify its interpretation. Sometimes it's spot on, sometimes it's wrong. But skimming a paper to fact-check a few specific quotes is still vastly faster than reading a dense paper completely blind.
We actually don't disagree at all—you are perfectly illustrating my point.
Applying strict epistemic discipline (Popper, Russell) to resolve ambiguity and accelerate actual practice is the very definition of deep work. You aren't using AI as a shortcut to skip thinking; you're using it as a Socratic sparring partner to deepen it. This is exactly the paradigm shift I'm advocating for.
I’m strongly reminded of early google every time I use AI for research. I used to be able to know little about a topic, try to search on it and get shit results. But, google would give me pages of results. So I could skim a lot and eventually on page 10, I stumble across some term of art, and that term would greatly improve my search. Rinse and repeat, and I’d have a good sense about the topic I was interested in.
You can’t really do that with google anymore, and I can’t remember the last time I bothered to actually learn something that wasn’t trivial from google. ChatGPT, however, has been a game changer. I can ask a really dumb question and get some basic info about the thing I’m asking about, and while it’s often not quite what I’m looking for, it gives me clues to follow, and I can quickly zero in on what I’m looking for, often in new contexts.
As an autodidact who’s main motivation to go to college was to get access to the stacks and direct internet access, I can’t even begin to tell you how game changing LLMs seem to be for learning.
To your point though, my concern is we don’t know how to teach how to learn, and LLMs will likely seduce many into bad behavior and poor research hygiene. I treat my research the same way I attack the stacks, but take someone who’s never been to a research library and ask them to create a report on some topic, and just why? That is the basic resistance, why?, why do what an LLM is almost literally built to do. Yet that is also highly related to individual learning, to take a bunch of disperate sources and synthesize output related to the input.
I suspect we’ll learn how to use LLMs in the same way we learned how to use calculators. But I have no doubt that on average (or maybe median or mode?) calculators have made us less capable to do basic arithmetic, and I suspect LLMs will also cause a great percentage of the population to be worse at sythesizing information. I’d hope that it’s only the same people who would have otherwise only gotten their information from TV, but I do have a slight fear it will creep past that subsection of the population.
Pretend learning is absolutely the key point, for me. There is danger in shifting our reasoning from knowing "stuff", to knowing a symbolic summary of "stuff" (helpfully generated by an LLM at varying levels of accuracy).
Previously, we saw a shift with search engines where we no longer needed to learn data because we could use a search engine as a mental signpost to the data, freeing up capacity for other thought.
LLMs are shifting knowledge creation to this mental pointer model. We don't need to know real "stuff" because we know how to look it up later (never?).
Each of these summaries is a secondary source, delivered through an agent biased by whatever is in its current context window. Like a game of telephone the summaries are inherently lossy, and each one may be 95% correct and we crucially don't understand which 5% may be incorrect.
When our basis for decision making is a collection of 100s or 1000s of LLM generated "Schrodinger's facts", we risk cumulative cascading errors. We will be wrong in unpredictable, chaotic ways.
We are voluntarily capping ourselves as this childish level of thought, because it feels like we are exercising our critical judgement the same as ever. However, the integrity of the inputs has been compromised. Bad inputs always lead to bad outputs.
Anecdote: I haven’t done any web development since 2002 and I always farmed that off to someone else.
But since I started using coding agents, I have done two feature full internal web apps authenticated by Amazon Cognito. While the UI looks like something from 2002, I am good at putting myself in the shoes of the end user, I iterated often (and quickly) over the UX.
I didn’t look at a line of code and have no plans to learn web development. I might have taken the time to learn a little before AI just to help me with internal websites. Yes I know it’s secure - I validated the endpoints can’t be accessed unauthenticated and the IAM role.
Second anecdote: I know AWS (trust me on this) like the back of my hand. I also know CloudFormation. For years I’ve been putting off learning Terraform and the CDK. After AI, why bother? I can one shot either for IAC and I’m very specific about what I want.
My company is happy and my customer is happy (consulting) what else matters? Substitute “customer” for “the business” or “stakeholders”
Agreed. This is pure architectural thinking: you hold the ground truth, enforce the strict IAM boundaries, and outsource the mud-playing to the LLM. Mindless 'vibe coding' without this structural discipline is just tittytainment. The job is contract validation now, not typing.
A) The IAM role of the Lambda runtime it’s running in is least privileged and only has access read and write access to the required S3 bucket and other required AWS services and even those are tightly scoped.
B) For authentication I used Amazon Cognito and ran a curl shell script against each endpoint for authentication vs non authenticated end points
C) The database user has least privilege access
So how pray tell could insecure code overcome that?
So you've made a read only wrapper around a database? That one person needs to access? There's no tentantization? You can't access more than one person's data? So there's zero chance one user can access someone else's data?
If you answered NO to any question, refer to my previous post.
If you answered YES, you could have just hooked your DB up to power BI or tableaux or whatever. Not exactly something to start boasting about that you're doing web dev.
I see you didn’t say anything about the possible security issues if I was using AWS Cognito for auth, tightly scoped IAM privileges for the runtime environment and a tightly scoped database user.
BTW, with AWS you can also enforce DynamoDB, Postgres and Redshift (?) to only allows rows to be accessed based on the user (IAM or Cognito) so no matter what Claude did, as long as you validate your security boundary at the AWS and database level, there wouldn’t be an issue.
Why would I trust developers (or Claude) to write secure multi tenant code when I can enforce it on the database/AWS layer?
1. I did say it’s an internal admin site and I mentioned AWS and S3. I didn’t say it was a reporting site only dealing with the database.
2. It’s B2B, every company pays 5-6 figures annually and they each have their own AWS account. No company can access any other company’s data because they each have their AWS account, user pool and database.
3. How am I “boasting” about doing “web dev” (poorly paid commodity work) when I specifically said I hadn’t done web development “since 2002” and talked about the UI was something from 2002?
4. I said it was an “Admin site”. I didn’tg say it was a reporting dashboard.
Productivity can’t be faked. Productivity also mostly aligns with incentives in my opinion. You want someone to be productive then give them a reason to be.
You are here. You know what you want. You know what are the correct rules and constraints. You know what the correct path looks like. You are able to spot drifting. You are able to objectively review the outcome. AI is the tool to go from here to where you want to arrive. There.
If not. AI is the tool to get "here" faster. And then go from here to there.
All you need is to take a little time to learn how to use this new tool: AI.
Take time to learn.
Have you taken the time to document what mathematicians were able to do with AI? What researchers were able to do with AI? They took the time to learn the AI tool. Then they used it with great results.
What are you waiting for? Learning is something you should also do. Go do it.
I can't see the issue of being lazy, and laziness becoming more productive than liability in these times we have AIs, most lazy people find better shortcuts and better ways to make things that other cannot find due to overly being filled with the need to be more productive than lazy and your brain capacity will get overloaded. Sometimes AI also can be an assistive tool for disabled people, and make even them Productive people, so calling it "lazy", over exaggerated.
I agree with your point but i think there are many cases in which its good. like if you're building something and get slowed down by topics you’re not familiar with, a summery of a topic is enough to remove the obstacle of not knowing how to continue. but yeah you can't call this "real" learning.
The issue, in my experience, is that there is a lot of productive work that does not look productive at first glance. Long term work may not look productive for years until it suddenly is tremendously productive. And there is a lot of quiet and often thankless maintenance work that goes on largely unnoticed that helps others do their jobs well. Both have value despite superficially looking unproductive at times. I'd argue that both look productive at long time scales but unproductive at short time scales.
Maybe for you reading a paper deeply is the most constructive way that you have to absorb information.
For me, it is having a document and interrogating it. Maybe having many sets of documents about a whole category of information. Getting the bullet points. getting the high level and then interrogating and digging down and being able to get bubbled up information as I need it.
That is the learning style that matches how I learn.
I have never been able to skim, so reading a large document WILL teach me that topic, but getting through that doc is tough.
I can dump a very large set of docs in a reader that lets me interrogate the whole data set and I can fly through looking for what is interesting to me, and what I may need, and along the way I will likely dive into other parts too. Asking questions keeps my hyperfocus active.
I think it is just a different style. I have synesthesia and a hard time not working on three to five things at once. I am use to knowing I learn differently than others.
And it doesn't matter. To each their own. Take one example: cooking. Some may choose to be a gourmet chef, whether professionally or just on their own time. Some will just regularly cook their food. Some will cook only when they have to. And some will avoid cooking no matter what, leaving it to family or going out to buy food, etc.
Now apply to every task and endeavor that one may be involved in. It doesn't matter if any particular thing sticks or not. Some may care and dive deeply, and some may prefer a hands-off approach. Nothing changes either way; life goes on.
The primary reason to get into anything deeply before was because it contributed directly to survival, eg studying and building a career to provide a product/service others needed. Things had to stick because living depended on it. Now with AI, well it just doesn't matter anymore with the essentials and everything beyond increasingly being automated away.
I have some algorithms I absolutely must know. So I’m hand coding them and asking the agent to critique me.
I do a very similar thing in writing - I need feedback, don’t rewrite this!
In both cases I need the struggle of editing / failing to arrive at a deeper understanding.
The future dev will need to know when to hand code vs when to not waste your time. And the advantage will still go to the person willing to experience struggle to understand what they need to.
This was the issue with some the ads Apple was running when launching the iPhone 16. It showed the worst worker using Apple Intelligence to impress the boss and get promotions, which being generally lazy and terrible. I felt it was the wrong message to send. [0]
I don’t think AI is all bad for summaries though. I used to add stuff to a reading list with good intentions, but things went there to die. Hundreds of articles added, but with so much new content each day, I would never actually read any of it. Now, I use AI summaries to get more context on what the article is. If it sounds interesting and I want more info, I can read the whole thing in the moment. If I’m satisfied with the summary alone, I can move on with my life. No more pushing it off to a reading list that only generates guilt. I actually end up reading more articles due to this, not less.
But to actually answer the question:
I’ve been putting research paper pdfs in notebook llm , and turning them into ~40 minute podcasts which I listen to on my walks.
Yes it’s shallow learning, and it might have some hallucinations in there but I wouldn’t have read some of those otherwise.
I think the risk is this; when non-technical users who've never shipped software in their life can dictate to a machine and get "instant results" it going to bring back managers not understanding that you don't just ship code. Especially these days where one bad dependency can mean downtime or worse.
Getting your directions from Google Maps might make you seem more knowledgeable about a city's geography than you actually are.
However, what does it mean to say that's deceptive? It means you care more about social signalling than you do about arriving at the right destination on time. Showing that you're not the sort of person who gets lost isn't really the primary reason people use Google Maps. When it's not a test of your navigation skills, it's not cheating.
Similarly, doing Google searches before posting might be "deceptive" in that it makes you seem more knowledgeable than you are, but on the whole I would prefer more knowledgeable posts, so the social signalling seems like a secondary consideration.
Similarly for using AI. Sometimes it's just a way to get more information.
A side-track and a possibly controversial opinion:
It seems to me that Agile methodology did a similar thing. The idea of Agile is not to skip understanding requirements, design, upfront reasoning and due diligence, as seen in seen in waterfall methods. It however sometimes turned into laziness looking like faster incremental progress.
I think quality of software has become worse over the time, with "unknown error occurred try again later" becoming more common, and I wonder if the root causes of it includes jumping to building things without properly thinking through about the customer problem, requirements and/or design.
I may easily be wrong, would like to hear corrective thoughts.
Agreed. LLMs have helped me achieve much deeper reading, _when directed to do so_. Asking an LLM to “Teach me Socratically about this paper/code. One question at a time”, usually allows me to get a much deeper reading of the material than I would otherwise.
That's a different take than I've been considering AI to be genuinely useful. I try to not use it for deep work, infact I try to use it minimally but frequently for short checks on my own understanding.
Using your research paper reading example, I would read the research paper, but then ask an AI tool specific questions about the work, frequently in new chats. Then at the end I might ask it to implement my description of the paper. I guess it's your 'debate with me' conclusion, the only difference is I would try to have multiple short conversations.
Ironically, what you described is exactly using AI to help with deep work. You do the heavy lifting (reading), and use AI strictly for stateless verification and testing your mental model. That is the ideal synergy.
Valid points, with which I agree and share the concern. I can't compete with colleagues, whom do things fast, if I want to learn. On the other hand, I no longer have to toil to work through things that I never truly learned, like tasks that require to be done a few times a year. Mastery is never acievable bc I forget, side quests become much less derailing. However, I am deprived of going through the motions and researching.
Thank you for this helpful differentiation. I agree - and if it‘s undermining our trust into ‚effort‘ (we start to be suspicious about how much some piece of work is really ‚worth‘), it undermines also our relationships.
I'm sick of hearing about AI, but I'm significantly more sick of anyone who knows how to write English prose at a level higher than "typical rural American" being accused of using AI to write.
It doesn't have to be. Comments such as yours add nothing to the conversation. It's an ad hominem attack. In the absence of explaining why you believe it "looks like AI", it's a baseless accusation
Implicit in "this post looks like AI"—at least the vast majority of the time—is that it is a wordy ramble with no real value because it says nothing novel or substantive—so I would not call it an hominem attack, but rather an honest criticism of the actual (lack of) content.
It has the typical patterns: em dashes, "it's not A, it's B" constructions. Also relatively new, low karma account, and its other comments are similarly LLM-ish.
What’s important? That bridges get built and stay up, or that they’re built only after toiling X amounts of hours. AI will change the nature of work, it’s going to make a lot of people uncomfortable. But more importantly, it’s going to let people who understand things faster get the info they need to be productive.
I have a feeling we would all be terrified if we knew how much AI had a role in building bridges at the moment.
TBD if they stay up, I suppose.
The stories I hear from various white collar professions not related to tech are... interesting, to say the least. There is a whole lot of unsanctioned shadow IT going on regardless of policy.
Stop using Vector RAG for coding agents. It's the wrong math
My Argument: Vector similarities don't understand that auth.py requires config.py to compile. Instead of fuzzy semantic search, we solve the exact Knapsack optimization problem with graph constraints
https://github.com/juyterman1000/entroly
I find value in learning some things deeply but not all things.
The ability to be more selective about where I attend deeply, while leveraging fast shallow learning to complete other tasks... That seems like a potential benefit and a nice choice to have in the toolbox.
I don't think it's all that bad. There's definitely vibe coding that is "copy paste / throw away" programming on ultra steroids. But after vibe coding two products and then finding them essentially impossible to then actually get to a quality bar I considered ready to launch, I've been working on a more measured approach that leverages AI but in a way that simply speeds up traditional programming. I use it to save tons of time on "why is pylance mad about X" "X works from the docs example but my slightly modified X gives error Y" "how do I make a toggle switch in css and html" "how am I supposed to do Python context managers in 2026 (I didn't know about the generator wrapper thing)" all that bullshit that constantly slows you down but needs to be right . AI is great at helping you kickstart and then keeping you unblocked.
I've been using Gemini chat for this, and specifically only giving it my code via copy paste. This sounds Luddite but actually it's been pretty interesting. I can show it my couple "core" library files and then ask it to do the next thing. I can inspect the output and retool it to my satisfaction, then slot it in to my program, or use it as an example to then hand code it.
This very intentional "me being the bridge" between AI and the code has helped so much in getting speed out of AI but then not letting it go insane and write a ton of slop.
And not to toot my own horn too much, but I think AI accelerates people more the wider their expertise is even if it's not incredibly deep. Eg I know enough CSS to spot slop and correct mistakes and verify the output. But I HATE writing CSS. So the AI and I pair really well there and my UIs look way better than they ever have.
Pure 'vibe coding' is essentially technical 'tittytainment'. Using AI for the horizontal spread while you enforce vertical architectural depth is true deep work.
Several weeks ago, I spent about a week fully reverse engineering a Stereomaker pedal. It accepts a mono signal and produces a stereo field using a 5-stage all-pass filter to mess with the phase without the use of delay (which sounds cheesy and creates a result that doesn't mix well back to mono).
I've not really worked with audio circuits previously, and I'd been intimidated to approach the domain. My journey was radically expedited by iterating through the entire process with a ChatGPT instance. I would share zoomed photos, grill it about how audio transformers work, got it to patiently explain JFET soft-switching using an inverter until the pattern was forced into my goopy brain.
Through the process of exploring every node of this circuit, I learned about configurable ground lifts, using a diode bridge to extract the desired voltage rail polarity, how to safely handle both TS and TRS cables with a transformer, that transformer outputs are 180 degrees out of phase, how to add a switch that will attenuate 10dB off a signal to switch line/instrument levels.
Eventually I transitioned from sharing PCB photos to implementing my own take on the cascade design in KiCAD, at which point I was copying and pasting chunks of netlist and reasoning about capacitor values with it.
In short, I gave myself a self-directed college-level intensive in about a week and since that's not generally a thing IRL, it's reasonable to conclude that I wouldn't have ever moved this from a "some day" to something I now understand deeply in the past tense without the ability to shamelessly interrogate an LLM at all hours of the day/night, on my schedule.
If you're lazy, perhaps you're just... lazy?
Anyhow, I highly recommend the Surfy Industries Stereomaker. It's amazing at what it does. https://www.surfyindustries.com/stereomaker
This is a phenomenal example of exactly what I am advocating.
Notice you didn't ask the AI to 'just design a stereo pedal for me.' You interrogated it, reasoned about netlists, and forced the concepts into your brain through intense friction. That is pure deep work.
Throughout the reverse engineering process, the LLM and I both were expecting each op-amp stage to use the next ladder value capacitor. We'd talked ourselves into how and why that would make sense.
At the end I was curious enough that I desoldered those five caps and realized that they were all 2.2nF except for the last stage which was 1nF.
I brought that news back to the LLM and we realigned our understanding of how the effect was achieved, ultimately coming to realize that our approach would have created notches at different frequencies instead of just shifting the phase by about 900 degrees.
It was an incredible learning experience. I try hard not to personify LLMs but this really did feel like working side by side with a friend on a problem until it was solved.
IRL, I suspect that most people who would be able to tackle that challenge with me lack both the time and patience to actually do it.
No, you touch on the aspects where you're able to use AI as an extension of your skills.
This is completely different than my colleague who isn't a software engineer, and now all of the sudden is creating PRs which I need to review and correct.
I'm a sceptic. I use it to explore the unknowns and go from there.
With as much kindness as I can manage, it doesn't sound like you're exploring very hard.
All of this stuff is remarkably easy to self-verify if you aren't, well, lazy.
Okay, it's been several weeks.
How much of what you did have you retained? Could you do all of, some of, a small fraction of, or none of the work again today if you had to?
My reaction to this question is that it might technically be in good faith, but you're pushing it.
Let's say that LLMs didn't exist, and I learned these same skills in an oddly specific hands-on workshop, or from an oddly specific textbook, or fuck it, let's say that I hired some greybeard pedal designer to just sit beside me and answer all of my stupid questions without judgement for a few weeks at their hourly rate.
Would you feel compelled to challenge whether I had retained what I learned or inexplicably woke up this morning, tabula rasa, and realized that I'd forgotten everything I spent a week teaching myself? I honestly don't think that you would.
For the record, I could reimplement any part of the circuit on demand if I needed to. I might be tempted to look at my notes for the JFET switching because it was genuinely hard to keep in my head, but that's more of a confidence thing than a "shit, I forgot how op-amps work" thing.
I've since implemented a variation into a matrix mixer concept that I'm working on, when it detects that a TS cable has been inserted into a TRS jack.
> Would you feel compelled to challenge whether I had retained what I learned
Yes, the exact same way I would dubious when someone says they learned much from following a youtube tutorial or participating in a two week workshop or something
That's funny... I taught myself Fusion in about a week following this excellent tutorial:
https://www.youtube.com/watch?v=mK60ROb2RKI
It's a 90 minute video that will take you a week to watch if you're doing it properly.
Seriously though... you don't learn from watching a video tutorial (which you can slow down and re-watch as many times as you need) and you apparently don't believe you can learn from an LLM which will patiently answer literally infinite questions, no matter how basic or repetitive... would you mind clarifying how you do learn?
Everyone has different learning styles so I tend to take a different strokes for different folks attitude. For example, I don't absorb highly technical stuff from books and the idea of [paying to be in a] classroom where you're forced to endure 95% what you're not interested in to get the 5% you care about (at the speed of the dumbest student in the room) gives me hives.
Yet, it kind of sounds like you might just be arguing for argument's sake. Also, you can learn A LOT in two weeks if you're motivated.
> would you mind clarifying how you do learn?
Practice
> Everyone has different learning styles so I tend to take a different strokes for different folks attitude
Okay but at the end of the day the only way to actually learn (and demonstrate that you've learned anything) is by actually doing it
And I don't really consider "I got the AI to do it" as actually doing it, which is why I'm questioning what you've actually retained.
To be clear if you feel like you've actually learned this stuff then good for you. I'm genuinely happy if that's the outcome you feel you have obtained
I'm just personally very skeptical of anyone learning fuck all from using AI to build stuff because like I said... I learn from practice. Using AI is not practice any more than copying from open source repos is.
And frankly I'm bitter because I absolutely cannot learn fuck all from using AI. It is the sort of shortcut that prevents my brain from committing anything to memory.
I guess I don't understand what about any aspect of what I explained gave you the impression that I might not have put what I learned into practice.
This is going to sound like I'm fucking with you, but I'm deadly serious: if someone taught you how to do something and you later learned that that person was actually an LLM masquerading as a human, would you forget what you had learned?
It's actually not impossible that you've hypnotized yourself, or could be experiencing a trauma response.
LLMs absolutely let you explore ideas and areas you wouldn't have otherwise...but does your new design actually _work_?
I'm curious whether the "knowledge" you gained was real or hallucinatory. I've been using LLMs this way myself, but I worry I'm contaminating my memory with false information.
At some point this existential doubt about your own work and others seems pretty weird.
Go ahead and figure out ways to interrogate on your work with technical means, that's a critical part of the process with an LLM or not.
Thank you.
I think that you're confusing what you're doing with what I'm doing.
What I'm doing is learning the circuit constructs that I need and then putting them to work in real circuits. There's usually a few breadboard steps in the middle, which you could call reinforcement learning.
To me, the telling thing about your question is the implication that I would spend a week learning how to do something and then not test it out. I know that this reply reads as salty, but I'm really struggling to contain my own "wtf" on this end.
Seriously, people that are so determined to prove that LLMs don't work despite how easy it is to test for yourself and see that they clearly do work are the ones that are hallucinating.
This will always happen as long as people are led by their egos. Also they probably aren't autodidacts and don't understand learning for fun.
I have a nearly total opposite take. I can't tell you how many times I've read a book, a paper or something else and been confused by some ambiguity in the author's prose. Being able to drop the paper (or even the book!) into an LLM to dig into the precise meaning has been an unbelievable boost for me.
Now I can actually get beyond conceptual misunderstanding or even ignorance and get to practice, which is how skills actually develop, in a much more streamlined way.
The key is to use the tool with discipline, by going into it with a few inviolable rules. I have a couple in my list, now: embrace Popperian falsifiability; embrace Bertrand Russell's statement: “Everything is vague to a degree you do not realize till you have tried to make it precise.”
LLMs have become excellent teachers for me as a result.
An LLM can't invent meaning in a text where there is none. It's equivalent to CSI's classic "zoom, enhance" on resolution limited photographs. You need to consider you're learning a load of rubbish from LLMs.
If you're not sure what something is saying, how can you be sure that the AI had picked the correct interpretation?
Right question to ask, however, good readers/professionals do have some sense for this and ability to dig further as needed. On the other hand, books and articles are often over-detailed, with the key stuff buried in the lede or even remaining tacit.
For me, LLMs have often pointed me to answers or given food for thought that even subject matter experts could not. I do not take those answers at face value, but the net result is still better than the search remaining open-ended.
Critical thinking.
In good faith, how do you tell yourself have good critical thinking?
I believe you're talking to an LLM, just look at the comment history
By asking it to cite its sources. Whenever I use AI, I have it pull direct quotes from the text to justify its interpretation. Sometimes it's spot on, sometimes it's wrong. But skimming a paper to fact-check a few specific quotes is still vastly faster than reading a dense paper completely blind.
We actually don't disagree at all—you are perfectly illustrating my point.
Applying strict epistemic discipline (Popper, Russell) to resolve ambiguity and accelerate actual practice is the very definition of deep work. You aren't using AI as a shortcut to skip thinking; you're using it as a Socratic sparring partner to deepen it. This is exactly the paradigm shift I'm advocating for.
I’m strongly reminded of early google every time I use AI for research. I used to be able to know little about a topic, try to search on it and get shit results. But, google would give me pages of results. So I could skim a lot and eventually on page 10, I stumble across some term of art, and that term would greatly improve my search. Rinse and repeat, and I’d have a good sense about the topic I was interested in.
You can’t really do that with google anymore, and I can’t remember the last time I bothered to actually learn something that wasn’t trivial from google. ChatGPT, however, has been a game changer. I can ask a really dumb question and get some basic info about the thing I’m asking about, and while it’s often not quite what I’m looking for, it gives me clues to follow, and I can quickly zero in on what I’m looking for, often in new contexts.
As an autodidact who’s main motivation to go to college was to get access to the stacks and direct internet access, I can’t even begin to tell you how game changing LLMs seem to be for learning.
To your point though, my concern is we don’t know how to teach how to learn, and LLMs will likely seduce many into bad behavior and poor research hygiene. I treat my research the same way I attack the stacks, but take someone who’s never been to a research library and ask them to create a report on some topic, and just why? That is the basic resistance, why?, why do what an LLM is almost literally built to do. Yet that is also highly related to individual learning, to take a bunch of disperate sources and synthesize output related to the input.
I suspect we’ll learn how to use LLMs in the same way we learned how to use calculators. But I have no doubt that on average (or maybe median or mode?) calculators have made us less capable to do basic arithmetic, and I suspect LLMs will also cause a great percentage of the population to be worse at sythesizing information. I’d hope that it’s only the same people who would have otherwise only gotten their information from TV, but I do have a slight fear it will creep past that subsection of the population.
Pretend learning is absolutely the key point, for me. There is danger in shifting our reasoning from knowing "stuff", to knowing a symbolic summary of "stuff" (helpfully generated by an LLM at varying levels of accuracy).
Previously, we saw a shift with search engines where we no longer needed to learn data because we could use a search engine as a mental signpost to the data, freeing up capacity for other thought.
LLMs are shifting knowledge creation to this mental pointer model. We don't need to know real "stuff" because we know how to look it up later (never?).
Each of these summaries is a secondary source, delivered through an agent biased by whatever is in its current context window. Like a game of telephone the summaries are inherently lossy, and each one may be 95% correct and we crucially don't understand which 5% may be incorrect.
When our basis for decision making is a collection of 100s or 1000s of LLM generated "Schrodinger's facts", we risk cumulative cascading errors. We will be wrong in unpredictable, chaotic ways.
We are voluntarily capping ourselves as this childish level of thought, because it feels like we are exercising our critical judgement the same as ever. However, the integrity of the inputs has been compromised. Bad inputs always lead to bad outputs.
Anecdote: I haven’t done any web development since 2002 and I always farmed that off to someone else.
But since I started using coding agents, I have done two feature full internal web apps authenticated by Amazon Cognito. While the UI looks like something from 2002, I am good at putting myself in the shoes of the end user, I iterated often (and quickly) over the UX.
I didn’t look at a line of code and have no plans to learn web development. I might have taken the time to learn a little before AI just to help me with internal websites. Yes I know it’s secure - I validated the endpoints can’t be accessed unauthenticated and the IAM role.
Second anecdote: I know AWS (trust me on this) like the back of my hand. I also know CloudFormation. For years I’ve been putting off learning Terraform and the CDK. After AI, why bother? I can one shot either for IAC and I’m very specific about what I want.
My company is happy and my customer is happy (consulting) what else matters? Substitute “customer” for “the business” or “stakeholders”
Agreed. This is pure architectural thinking: you hold the ground truth, enforce the strict IAM boundaries, and outsource the mud-playing to the LLM. Mindless 'vibe coding' without this structural discipline is just tittytainment. The job is contract validation now, not typing.
I didn’t look at a line of code... I know it’s secure - I validated the endpoints can’t be accessed unauthenticated and the IAM role
Oh god, this made me laugh so hard.
Best 'we gonna get hacked' comment of the day.
Please tell me how is going to be “hacked”?
A) The IAM role of the Lambda runtime it’s running in is least privileged and only has access read and write access to the required S3 bucket and other required AWS services and even those are tightly scoped.
B) For authentication I used Amazon Cognito and ran a curl shell script against each endpoint for authentication vs non authenticated end points
C) The database user has least privilege access
So how pray tell could insecure code overcome that?
So you've made a read only wrapper around a database? That one person needs to access? There's no tentantization? You can't access more than one person's data? So there's zero chance one user can access someone else's data?
If you answered NO to any question, refer to my previous post.
If you answered YES, you could have just hooked your DB up to power BI or tableaux or whatever. Not exactly something to start boasting about that you're doing web dev.
I see you didn’t say anything about the possible security issues if I was using AWS Cognito for auth, tightly scoped IAM privileges for the runtime environment and a tightly scoped database user.
BTW, with AWS you can also enforce DynamoDB, Postgres and Redshift (?) to only allows rows to be accessed based on the user (IAM or Cognito) so no matter what Claude did, as long as you validate your security boundary at the AWS and database level, there wouldn’t be an issue.
Why would I trust developers (or Claude) to write secure multi tenant code when I can enforce it on the database/AWS layer?
https://aws.amazon.com/blogs/database/multi-tenant-data-isol...
https://docs.aws.amazon.com/IAM/latest/UserGuide/reference_p...
1. I did say it’s an internal admin site and I mentioned AWS and S3. I didn’t say it was a reporting site only dealing with the database.
2. It’s B2B, every company pays 5-6 figures annually and they each have their own AWS account. No company can access any other company’s data because they each have their AWS account, user pool and database.
3. How am I “boasting” about doing “web dev” (poorly paid commodity work) when I specifically said I hadn’t done web development “since 2002” and talked about the UI was something from 2002?
4. I said it was an “Admin site”. I didn’tg say it was a reporting dashboard.
Productivity can’t be faked. Productivity also mostly aligns with incentives in my opinion. You want someone to be productive then give them a reason to be.
You are here. You know what you want. You know what are the correct rules and constraints. You know what the correct path looks like. You are able to spot drifting. You are able to objectively review the outcome. AI is the tool to go from here to where you want to arrive. There.
If not. AI is the tool to get "here" faster. And then go from here to there.
All you need is to take a little time to learn how to use this new tool: AI.
Take time to learn.
Have you taken the time to document what mathematicians were able to do with AI? What researchers were able to do with AI? They took the time to learn the AI tool. Then they used it with great results.
What are you waiting for? Learning is something you should also do. Go do it.
No mystery here.
I can't see the issue of being lazy, and laziness becoming more productive than liability in these times we have AIs, most lazy people find better shortcuts and better ways to make things that other cannot find due to overly being filled with the need to be more productive than lazy and your brain capacity will get overloaded. Sometimes AI also can be an assistive tool for disabled people, and make even them Productive people, so calling it "lazy", over exaggerated.
I agree with your point but i think there are many cases in which its good. like if you're building something and get slowed down by topics you’re not familiar with, a summery of a topic is enough to remove the obstacle of not knowing how to continue. but yeah you can't call this "real" learning.
I'm convinced that at some point looking like being productive and being productive becomes the same thing.
The issue, in my experience, is that there is a lot of productive work that does not look productive at first glance. Long term work may not look productive for years until it suddenly is tremendously productive. And there is a lot of quiet and often thankless maintenance work that goes on largely unnoticed that helps others do their jobs well. Both have value despite superficially looking unproductive at times. I'd argue that both look productive at long time scales but unproductive at short time scales.
There's a point where they meet, but "faking it until you make it" doesn't work for productivity in the same way it doesn't work for getting rich.
But there's a secret: just buy my $399 masterclass and I'll teach you 17 simple productivity hacks to 100x your income.
So what is important is not that 10 or 20 times the work can be done, but that you are stressed out and exhausted while doing your work?
It is difficult to have any bandwidth to protest about the state of everything if you're burned out.
Maybe for you reading a paper deeply is the most constructive way that you have to absorb information.
For me, it is having a document and interrogating it. Maybe having many sets of documents about a whole category of information. Getting the bullet points. getting the high level and then interrogating and digging down and being able to get bubbled up information as I need it.
That is the learning style that matches how I learn.
I have never been able to skim, so reading a large document WILL teach me that topic, but getting through that doc is tough.
I can dump a very large set of docs in a reader that lets me interrogate the whole data set and I can fly through looking for what is interesting to me, and what I may need, and along the way I will likely dive into other parts too. Asking questions keeps my hyperfocus active.
I think it is just a different style. I have synesthesia and a hard time not working on three to five things at once. I am use to knowing I learn differently than others.
> nothing actually sticks.
And it doesn't matter. To each their own. Take one example: cooking. Some may choose to be a gourmet chef, whether professionally or just on their own time. Some will just regularly cook their food. Some will cook only when they have to. And some will avoid cooking no matter what, leaving it to family or going out to buy food, etc.
Now apply to every task and endeavor that one may be involved in. It doesn't matter if any particular thing sticks or not. Some may care and dive deeply, and some may prefer a hands-off approach. Nothing changes either way; life goes on.
The primary reason to get into anything deeply before was because it contributed directly to survival, eg studying and building a career to provide a product/service others needed. Things had to stick because living depended on it. Now with AI, well it just doesn't matter anymore with the essentials and everything beyond increasingly being automated away.
You still have to check output as it could be wrong
With things like Tiktok I've learned that we need to break up bigger works into smaller digestible pieces
Another issue is there is too much content for people to read or consume already (a problem independent of AI)
Yeah, it's about "effective use of AI as a tool"
I have some algorithms I absolutely must know. So I’m hand coding them and asking the agent to critique me.
I do a very similar thing in writing - I need feedback, don’t rewrite this!
In both cases I need the struggle of editing / failing to arrive at a deeper understanding.
The future dev will need to know when to hand code vs when to not waste your time. And the advantage will still go to the person willing to experience struggle to understand what they need to.
This was the issue with some the ads Apple was running when launching the iPhone 16. It showed the worst worker using Apple Intelligence to impress the boss and get promotions, which being generally lazy and terrible. I felt it was the wrong message to send. [0]
I don’t think AI is all bad for summaries though. I used to add stuff to a reading list with good intentions, but things went there to die. Hundreds of articles added, but with so much new content each day, I would never actually read any of it. Now, I use AI summaries to get more context on what the article is. If it sounds interesting and I want more info, I can read the whole thing in the moment. If I’m satisfied with the summary alone, I can move on with my life. No more pushing it off to a reading list that only generates guilt. I actually end up reading more articles due to this, not less.
[0] https://youtu.be/YP-ukrBVDH8 (this is sadly the best copy I can find)
Does this post feel AI generated to anyone else ?
But to actually answer the question: I’ve been putting research paper pdfs in notebook llm , and turning them into ~40 minute podcasts which I listen to on my walks. Yes it’s shallow learning, and it might have some hallucinations in there but I wouldn’t have read some of those otherwise.
I think the risk is this; when non-technical users who've never shipped software in their life can dictate to a machine and get "instant results" it going to bring back managers not understanding that you don't just ship code. Especially these days where one bad dependency can mean downtime or worse.
Getting your directions from Google Maps might make you seem more knowledgeable about a city's geography than you actually are.
However, what does it mean to say that's deceptive? It means you care more about social signalling than you do about arriving at the right destination on time. Showing that you're not the sort of person who gets lost isn't really the primary reason people use Google Maps. When it's not a test of your navigation skills, it's not cheating.
Similarly, doing Google searches before posting might be "deceptive" in that it makes you seem more knowledgeable than you are, but on the whole I would prefer more knowledgeable posts, so the social signalling seems like a secondary consideration.
Similarly for using AI. Sometimes it's just a way to get more information.
A side-track and a possibly controversial opinion:
It seems to me that Agile methodology did a similar thing. The idea of Agile is not to skip understanding requirements, design, upfront reasoning and due diligence, as seen in seen in waterfall methods. It however sometimes turned into laziness looking like faster incremental progress.
I think quality of software has become worse over the time, with "unknown error occurred try again later" becoming more common, and I wonder if the root causes of it includes jumping to building things without properly thinking through about the customer problem, requirements and/or design.
I may easily be wrong, would like to hear corrective thoughts.
Agreed. LLMs have helped me achieve much deeper reading, _when directed to do so_. Asking an LLM to “Teach me Socratically about this paper/code. One question at a time”, usually allows me to get a much deeper reading of the material than I would otherwise.
That's a different take than I've been considering AI to be genuinely useful. I try to not use it for deep work, infact I try to use it minimally but frequently for short checks on my own understanding.
Using your research paper reading example, I would read the research paper, but then ask an AI tool specific questions about the work, frequently in new chats. Then at the end I might ask it to implement my description of the paper. I guess it's your 'debate with me' conclusion, the only difference is I would try to have multiple short conversations.
Ironically, what you described is exactly using AI to help with deep work. You do the heavy lifting (reading), and use AI strictly for stateless verification and testing your mental model. That is the ideal synergy.
Valid points, with which I agree and share the concern. I can't compete with colleagues, whom do things fast, if I want to learn. On the other hand, I no longer have to toil to work through things that I never truly learned, like tasks that require to be done a few times a year. Mastery is never acievable bc I forget, side quests become much less derailing. However, I am deprived of going through the motions and researching.
Thank you for this helpful differentiation. I agree - and if it‘s undermining our trust into ‚effort‘ (we start to be suspicious about how much some piece of work is really ‚worth‘), it undermines also our relationships.
A good example is ‚birthday wishes‘:
https://m.youtube.com/watch?v=2IYqhdJuRfU&t=5m47s
(AutoCorrect, AutoComplete - generate? AutoCongratulate? How much is ‚okay‘?)
Weird post given it looks like an LLM wrote it.
I'm sick of hearing about AI, but I'm significantly more sick of anyone who knows how to write English prose at a level higher than "typical rural American" being accused of using AI to write.
Agreed. Feels like I see this accusation made on every other post on this site nowadays. And maybe it's even true.
But if it's not, it's insulting to the poster, and if it is, then who cares if people are engaging with the post.
Well that's the world we live in now.
It doesn't have to be. Comments such as yours add nothing to the conversation. It's an ad hominem attack. In the absence of explaining why you believe it "looks like AI", it's a baseless accusation
Implicit in "this post looks like AI"—at least the vast majority of the time—is that it is a wordy ramble with no real value because it says nothing novel or substantive—so I would not call it an hominem attack, but rather an honest criticism of the actual (lack of) content.
It has the typical patterns: em dashes, "it's not A, it's B" constructions. Also relatively new, low karma account, and its other comments are similarly LLM-ish.
Em dashes are—hear me out—easy for anyone who knows how to type a hyphen twice.
What’s important? That bridges get built and stay up, or that they’re built only after toiling X amounts of hours. AI will change the nature of work, it’s going to make a lot of people uncomfortable. But more importantly, it’s going to let people who understand things faster get the info they need to be productive.
AI does not currently build bridges that stay up
I have a feeling we would all be terrified if we knew how much AI had a role in building bridges at the moment.
TBD if they stay up, I suppose.
The stories I hear from various white collar professions not related to tech are... interesting, to say the least. There is a whole lot of unsanctioned shadow IT going on regardless of policy.
Not really, there’s a lot it does right. But any automated tool or calculator will be as good as its operator.
Stop using Vector RAG for coding agents. It's the wrong math My Argument: Vector similarities don't understand that auth.py requires config.py to compile. Instead of fuzzy semantic search, we solve the exact Knapsack optimization problem with graph constraints https://github.com/juyterman1000/entroly
I find value in learning some things deeply but not all things.
The ability to be more selective about where I attend deeply, while leveraging fast shallow learning to complete other tasks... That seems like a potential benefit and a nice choice to have in the toolbox.
trick is maintaining enough domain expertise... so we can actually audit those shallow outputs.
If the baseline knowledge drops too low we cannot tell when the AI is being lazy or wrong
I don't think it's all that bad. There's definitely vibe coding that is "copy paste / throw away" programming on ultra steroids. But after vibe coding two products and then finding them essentially impossible to then actually get to a quality bar I considered ready to launch, I've been working on a more measured approach that leverages AI but in a way that simply speeds up traditional programming. I use it to save tons of time on "why is pylance mad about X" "X works from the docs example but my slightly modified X gives error Y" "how do I make a toggle switch in css and html" "how am I supposed to do Python context managers in 2026 (I didn't know about the generator wrapper thing)" all that bullshit that constantly slows you down but needs to be right . AI is great at helping you kickstart and then keeping you unblocked.
I've been using Gemini chat for this, and specifically only giving it my code via copy paste. This sounds Luddite but actually it's been pretty interesting. I can show it my couple "core" library files and then ask it to do the next thing. I can inspect the output and retool it to my satisfaction, then slot it in to my program, or use it as an example to then hand code it.
This very intentional "me being the bridge" between AI and the code has helped so much in getting speed out of AI but then not letting it go insane and write a ton of slop.
And not to toot my own horn too much, but I think AI accelerates people more the wider their expertise is even if it's not incredibly deep. Eg I know enough CSS to spot slop and correct mistakes and verify the output. But I HATE writing CSS. So the AI and I pair really well there and my UIs look way better than they ever have.
Pure 'vibe coding' is essentially technical 'tittytainment'. Using AI for the horizontal spread while you enforce vertical architectural depth is true deep work.
We need to allocate some % of our AI use to tackle this problem. Help us learn and find better abstractions and methods.