How I Use "AI"

There's a running theme in here of programming problems LLMs solve where it's actually not that important that the LLM is perfectly correct. I've been using GPT4 for the past couple months to comprehend Linux kernel code; it's spooky good at it.

I'm a C programmer, so I can with some effort gradually work my way through random Linux kernel things. But what I can do now instead is take a random function, ask GPT4 what it does and what subsystem it belongs to, and then ask GPT4 to write me a dummy C program that exercises that subsystem (I've taken to asking it to rewrite kernel code in Python, just because it's more concise and easy to read).

I don't worry at all about GPT4 hallucinating stuff (I'm sure it's doing that all the time!), because I'm just using its output as Cliff's Notes for the actual kernel code; GPT4 isn't the "source of truth" in this situation.

Just today I had GPT4 implement a SwiftUI based UI for a prototype I’m working on. I was able to get it to work with minimal tweaks within 15 minutes even though I know next to nothing about SwiftUI (I’m mainly a systems person these days). I pay for this, and would, without hesitation, pay 10x for a larger model which does not require “minimal tweaks” for the bullshit tasks I have to do. Easily 80% of all programming consists of bullshit tasks that LLMs of 2024 are able to solve within seconds to minutes, whereas for me some of them would take half a day of RTFM. Worse, knowing that I’d have to RTFM I probably would avoid those tasks like the plague, limiting what can be accomplished. I’m also relieved somewhat that GPT4 cannot (yet?) help me with the non-bullshit parts of my work.

For about 20 years, chess fans would hold "centaur" tournaments. In those events, the best chess computers, who routinely trounced human grandmasters, teamed up with those same best-in-the-world humans and proceeded to wipe both humans and computers off the board. Nicholas is describing in detail how he pairs up with LLMs to get a similar result in programming and research.

Sobering thought: centaur tournaments at the top level are no more. That's because the computers got so good that the human half of the beast no longer added any meaningful value.

https://en.wikipedia.org/wiki/Advanced_chess

What’s everyone’s coding LLM setup like these days? I’m still paying for Copilot through an open source Xcode extension and truthfully it’s a lot worse than when I started using it.

Every now and then, I'll actually sort of believe an article like this. Then I go and test the current models on things like semantic search.

For instance -

The Hough transform detects patterns with certain structure in images, e.g. circles or lines.

So I'm looking for academic research papers which apply the Hough transform to audio spectra, to recognize the harmonic structure of tonal audio and thus determine the fundamental pitch. (i.e. the Hough space would be a 1D space over fundamental frequency).

I've spent probably 90 minutes, over the several times I've read an optimistic post like this, asking various LLMs (mostly GPT-4o, though my early tests predate GPT-4o, and I've also tried Gemini and Claude), prompts along the lines of

The Hough transform detects patterns with certain structure in images, e.g. circles or lines. > I'm looking for academic research papers (please link them or provide a DOI.org link at least) which apply the Hough transform to audio spectra, to identify the harmonic structure of audio and thus determine the fundamental pitch. > Make sure to provide only papers that actually exist. If you can't find anything particularly relevant, say so as a disclaimer & just provide the most relevant papers you can.

This is a reliable "fake paper generator", unfortunately - it'll just make up plausible garbage like

Here are some academic papers related to applying the Hough transform to audio spectra for identifying harmonic structures and fundamental pitch:

"An Audio Pitch Detection Algorithm Based on the Hough Transform" > Authors: Mark W. and John D. > Published In: IEEE Transactions on Audio, Speech, and Language Processing > DOI: 10.1109/TASL.2008.2000773 > Abstract: This paper proposes an audio pitch detection algorithm that utilizes the Hough transform to analyze the harmonic structure of audio spectra and determine the fundamental pitch.

This paper does not exist. Complete waste of my time. And again, this behavior persists over the >1 year period I've been trying this query.

And it's not just search-like tasks. I've tried asking for code and gotten stuff that's outright dangerous (try asking for code to do safe overflow-checked addition on int64_t in C- you have about an 80% chance of getting code that triggers UB in one way or another). I've asked for floating-point calling conventions on RISC-V for 32-bit vs 64-bit (would have been faster than going through the extension docs), and been told that RV64 has 64 floating-point registers (hey, it's got a 64 in the name!). I've asked if Satya Nadella ever had COVID-19 and been told- after GPT-4o "searched the web"- that he got it in March of 2023.

As far as I can tell, LLMs might conceivably be useful when all of the following conditions are true:

1. You don't really need the output to be good or correct, and 2. You don't have confidentiality concerns (sending data off to a cloud service), and, 3. You don't, yourself, want to learn anything or get hands-on - you want it done for you, and 4. You don't need the output to be in "your voice" (this is mostly for prose writing, for code this doesn't really matter); you're okay with the "LLM dialect" (it's crucial to delve!), and 5. The concerns about environmental impact and the ethics of the training set aren't a blocker for you.

For me, pretty much everything I do professionally fails condition number 1 and 2, and anything I do for fun fails number 3. And so, despite a fair bit of effort on my part trying to make these tools work for me, they just haven't found a place in my toolset- before I even get to 4 or 5. Local LLMs, if you're able to get a beefy enough GPU to run them at usable speed, solve 2 but make 1 even worse...

Burn down the rain forests so researchers can save time writing code

I think the author does a decent job laying out good ways of using the LLMs. If you’re gonna use them, this is probably the way.

But he acknowledges the ethical social issues (also misses the environmental issues https://disconnect.blog/generative-ai-is-a-climate-disaster/) and then continues to use them anyway. For me the ickiness factor is too much, the benefit isn’t worth it.

This mostly matches my experience but with one important caveat around using them to learn new subjects.

When I'm diving into a wholly new subject for the first time, in a field totally unrelated to my field (similar to the author, C programming and security) for example biochemistry or philosophy or any field where I don't have even a basic grounding, I still worry about having subtly-wrong ideas about fundamentals being planted early-on in my learning.

As a programmer I can immediately spot "is this code doing what I asked it to do" but there's no equivalent way to ask "is this introductory framing of an entire field / problem space the way an actual expert would frame it for a beginner" etc.

At the end of the day we've just made the reddit hivemind more eloquent. There's clearly tons of value there but IMHO we still need to be cognizant of the places where bad info can be subtly damaging.

This just does not match my experience with these tools. I've been on board with the big idea expressed in the article at various points and tried to get into that work flow, but with each new generation of models they just do not do well enough, consistently enough, on serious tasks to be a time or effort saver. I don't know what world these apparently high output people live in where their days consist of porting Conway's Game of Life and writing shell scripts that only 'mostly' need to work, but I hope one day I can join them.

If I knew why something is [flagged] I could probably learn something from it.

I just want to emphasise two things, which are both mentioned in the article, but I still want to emphasise them as they are core to what I take from the article as someone who has been a fan boy of Nicholas for years now

1. Nicholas really does know how badly machine learning models can be made to screw up. Like, he really does. [0]

2. This is how Nicholas -- an academic researcher in the field of security of machine learning -- uses LLMs to be more efficient.

I don't know whether Nicolas works on globally scaled production systems with have specific security/data/whatever controls that need to be adhered to, or whether he even touches any proprietary code. But seeing as he heavily emphasised the "i'm a researcher doing research things" in the article -- I'd take a heavy bet that he does not. And academic / research / proof-of-concept coding has different limitations/context/needs than other areas.

I think this is a really great write up, even as someone on the anti-LLM side of the argument. I really appreciate the attempt to do a "middle of the road" post which is absolutely what the conversation needs right now (pay close attention to how this was written LLM hypers).

I don't share his experience, I still value and take enjoyment from the "digging for information" process -- it is how I learn new things. Having something give me the answer doesn't help me learn, and writing new software is a learning process for me.

I did take a pause and digested the food for thought here. I still won't be using an LLM tomorrow. I am looking forward to his next post, which sounds very interesting.

[0]: https://nicholas.carlini.com/papers

i appreciate the article and the full examples. But I have to say this all looks like a nightmare to me. Going back and forth in English with a slightly dumb computer that needs to be pestered constantly and hand-held through a process? This sounds really really painful.

Not to mention that the author is not really learning the underlying tech in a useful way. They may learn how to prompt to correct the mistakes the LLM makes, but if it was a nightmare to go through this process once, then dealing with repeating the same laborious walkthrough each time you want to do something with Docker or build a trivial REST API sounds like living in hell to me.

Glad this works for some folks. But this is not the way I want to interact with computers and build software.

It is overhyped. If you don't know much about what you're trying to do, then you're not going to know how bad or suboptimal the the LLM's output is. Some people will say it doesn't matter as long as it gets the job done. Then they end up paying a lot extra for me to come in and fix it when it's going haywire in prod.

I think all of these can be summarized into three items

1. Search engine - Words like "teach" or "learn" used to be slapped on Google once upon a time. One real great thing about LLMs here is that they do save time. The internet these days is unbelievably crappy and choppy. It often takes more time to click through the first item in the Google result and read it than to simply ask an LLM and wait for its slowish answer.

2. Pattern matching and analysis - LLMs are probably the most advanced technology for recognizing well-known facts and patterns from text, but they do make quite some errors especially with numbers. I believe that a properly fine-tuned small LLMs would easily beat gigantic models for this purpose.

3. Interleaving knowledge - this is the biggest punch that LLMs have, and also the main source of all the over-hype (which does still exist). It can produce something valuable by synthesizing multiple facts, like writing complex answers and programs. But this is where hallucination happens most frequently, so it's critical that you review the output carefully.

The biggest AI skeptics i know are devops/infrastructure engineers.

At this point i believe most of them can not be convinced that LLMs are valuable or useful by any sort of evidence, but if anything could do it, this article could. Well done.

Good article and it matches my own experience in the last year. I use it to my advantage both on hobby projects and professionally and it's a huge timesaver.

LLMs are far from flawless of course and I often get stuck with non working code. Or it is taking annoying shortcuts in giving a detailed answer, or it just wastes a lot of time repeating the same things over and over again. But that's often still useful. And you can sometimes trick them into doing better. Once it goes down the wrong track, it's usually best to just start a new conversation.

There are a few neat tricks that I've learned over the last year that others might like:

- you can ask chat gpt to generate some files and make them available as a zip file. This is super useful. Don't wait for it to painfully slowly fill some text block with data or code. Just ask it for a file and wait for the link to become available. Doesn't always seem to work but when it does it is nice. Great for starting new projects.

- chat gpt has a huge context window so you can copy paste large source files in it. But why stop there? I wrote a little script (with a little help of course) that dumps the source tree of a git repository into a single text file which I can then copy into the context. Works great for small repositories. Then you can ask questions like "add a function to this class that does X", "write some unit tests for foo", "analyze the code and point out things I've overlooked", etc.

- LLMs are great for the boring stuff. Like writing exhaustive unit tests that you can't be bothered with or generating test data. And if you are doing test data, you might as well have some fun and ask it to inject some movie quotes, litter it with hitchhiker's guide to the galaxy stuff, etc.

The recent context window increase to 128K with chat gpt 4o and other models was a game changer. I'm looking forward to that getting even larger. The first few publicly available LLMs had the memory of a gold fish. Not any more. Makes them much more useful already. Right now most small projects easily fit into its context already.

No need for programmers anymore

The problem I have with LLMs is that one can never be sure that it will give you the best possible solution. In fact in coding very often it will give you a working but also outdated solution. And this is futile. Because in coding even the best possible solution nowadays gets old very quickly. But if you use LLMs your code will be outdated from the start. That is nothing I would pay for.

I also use LLMs similarly. As a professional programmer, LLMs save me a lot of time. They are especially efficient when I don't understand a flow or need to transform values from one format to another. However, I don't currently use them to produce code that goes into production. I believe that in the coming years, LLMs will evolve to analyze complete requirements, architecture, and workflow and produce high-quality code. For now, using LLMs to write production-ready applications in real-time scenarios will take longer.

My biggest use for LLMs - situations where I use them heavily:

- CLI commands and switches I don't care to or easily remember

- taking an idea and exploring it in various ways

- making Slack messages that are more engaging

Using GPTs has a cost of breaking my concentration/flow, so it's not part of my core workflows.

I really need to start weaving it into the programming aspects of my workday.

Trimming down large codebases to significantly simplify the project.

I was a bit excited at something being able to do that, but this apparently means simplifying a single file, based on their example.

I suspect they're having an unusually positive experience with these tools due to working on a lot of new, short, programs.

"I understand this better than you do" twice in about 30 lines. Okay then.

I mean, sure, you do, but there's less off-putting ways to display your credentials...

I had a brain haemorrhage so I can not form sentences as well as I could. I nudge my AI to more accurately explain what I mean.

It is strange, I can read okay - by formations of words is challenging.

So coding

What is very useful for me is when I conduct research outside of my field of expertise, I do not even know what keywords to look for. An LLM can help you with this.

If you use it as an intern, as a creative partner, as a rubber-duck-plus, in an iterative fashion, give it all the context you have and your constraints and what you want... it's fantastic. Often I'll take pieces from it, if it's simple enough I can just use it's output.

I'm also a security researcher. My day-to-day job for nearly the last decade now has been to show all of the ways in which AI models fail spectacularly when confronted with any kind of environment they were not trained to handle.

... And yet, here I am, saying that I think current large language models have provided the single largest improvement to my productivity since the internet was created.

In the same way I wouldn't write off humans as being utterly useless because we can't divide 64 bit integers in our head---a task completely trivial for a computer---I don't think it makes sense to write off LLMs because you can construct a task they can't solve. Obviously that's easy---the question is can you find tasks where they provide value?

If I know technology which I am using llm for then llm helps me to do it faster. If I am not familiar with technology then llm helps me to learn it faster by showing me win the code that it generates which part of technology is important and how it works in real examples. But I do not think it is helpful and I would say it may be dangerous depending on task you do if you do not know technology and also do not what to learn it and understand how generated code works.

100%

Sounds like the author is trying really hard to find an edge use case for an LLM. Meanwhile on YouTube... "I Made 100 Videos In One Hour With Ai - To Make Money Online"

I also make use of LLMs to help me with certain programming problems. But this author simply glides over a very important issue: how do you use LLMs responsibly? What does it mean to be responsible in your work?

If all of this is just a hobby for you, then it doesn't matter. But it matters a lot when you are serving other people; it matters when you must account for your work.

You could make the case that all testing is a waste of time, because "I can do this, and this, and this. See? It appears to work. Why bother testing it?" We test things because it's irresponsible not to. Because things fail fairly often.

I'm looking through the author's examples. It appears that he knows a lot about technology in general, so that he can be specific about what he wants. He also appears to be able to adjust and evaluate the results he gets. What if someone is bad at that? The LLM can't prompt itself or supervise itself.

I come to everything with the mindset of a tester. LLMs have most definitely been overhyped. That doesn't mean they are useless, but any article about what they are able to do which doesn't also cover how they fail and how you be ready for them to fail is a disservice to the industry.

What if we had something that could fill the gaps in docs for devlopers using a library? It doesn't actually write the docs, but simply hints at what a function could do. Would be pretty useful for beginner devs

pretty much exactly how I settled on using LLMs after a month of making an effort to.

I use pretty much as an "better google". I formulate questions and try to be as especific as possible, and I have good results in fixing some code troubles I had.

Is pretty much a better indexed search engine.

This fully matches my experience using Chat GPT for the past 12 months. You just have to allow yourself to ask it questions like you might ask a very smart co-worker and it just keeps delivering. In many ways it has delivered as a co-CTO on one rather complicated project I've been working on.

"And that's where language models come in. Because most new-to-me frameworks/tools like Docker, or Flexbox, or React, aren't new to other people. There are probably tens to hundreds of thousands of people in the world who understand each of these things thoroughly. And so current language models do to. " Apparently not using it to proof-read or it would end with "too. "

I would love to know which plugin or custom code the author uses to automate workflows in emacs as well as the shell

I've been getting a similar feeling lately, in that if a thing has been done before and knowledge is publicly available, asking the "AI" (the LLMs) first about it is the best place to start. It looks like that's how things are going to be from now on and it's only going to amplify.

But as the AI gets increasingly competent at showing us how to do things, knowing what task is worth doing and what not is still a task for the one who asks, not the AI.

Edit: One type of question I've found interesting is to make it speculate, that is asking questions that it doesn't know the answer to, but still is able to speculate because they involve combining things it does know about in novel (though not necessarily valid) ways.

Something weird is going on with this web page on Chrome in Ubuntu. The table of contents is obscuring the page text.

I wonder what the author thinks it openinterpreter/similar, which is a higher level of indirection, so you ask the computer to do it for it and it just does it for you. the first section is the kind of thing I'd use it for.

"make me a docker container that does foo." "now have it do bar."

though the author uses emacs, so maybe they get the same level of not-having-to-copy-and-paste.

I've had the most current publicly available models fail to give me even simple correct boilerplate code, but the guy is like: ...we have to be nuanced but, "Converting dozens of programs to C or Rust to improve performance 10-100x."? Seriously?

I also recently asked openai's gpt 4 which number is bigger, 3.8 oder 3.11, and it was pretty damn sure that it's 3.11, because 11 bigger 8, obviously. Another time I asked Meta Llama 3.1 70B and gpt 4 multiple times using a variation of prompts to suggest a simple code example for a feature in a project I was working on. They confidently provided code that was total nonsense, configuration that did nothing and even offered a dependency that didn't exist but somewhat sounded like the prompt itself.

I cannot predict the future. Maybe all of this will lead to something useful. Currently though? Massively overhyped. I've talked to CS colleagues and friends who also work as software developers that are all way more competent than me about their experiences, and some were exited about the prospects, none could provide a current use case for their work. The only instances I know in which people talk positively about these models are in online articles like this or in comment sections adjacent to them. Never in real live among actual developers in person.

I thought author meant how they use the two-letter sequence “AI” and I just came here to say, Allen Iverson.

I work with multiple programming languages and it's a godsend. Having something that gives you mostly correct instructions on how to do a generic thing without having to wade through today's garbage web search experience is fantastic.

Must be nice to work on stuff that doesn’t compete with “intelligence as a service.” I feel that’s an empty set, and everyone using these services actively rationalizes selling out the human race by *paying to get brain raped.*

“Open” AI - customer noncompete Copilot - customer noncompete Anthropic - customer noncompete Gemini - customer noncompete (api only, wow)

Just imagine millions of people know about the imitation game and still pay someone to fuck them over like that.

Surely, our descendants will thank us for such great contributions to the memory banks of the monopolies of the boring dystopia ..

This perfectly echoes my experience with AI.

It’s not perfect, but AI for working with code has been an absolute game changer for me.

I use it for boilerplate, totally uninteresting and boring code. Stuff like Bash parameters validation (it's only validation, so the damage when it hallucinates is quite limited and usually quickly shows up) and Google spreadsheets formula generation, stuff like: "extract ticker name from the OCC symbol in the previous column, write '-' instead if it's empty". It's really boring stuff to do manually and it's actually faster have GPT 4o generate it from me from the sentence than write it myself.

Typically there is fixing needed (e.g. it shall fuck up things as trivial as parentheses placement) but it's still useful.

Lots of french/english translation too: it's actually good at that.

LLM is the best human to computer interface I have ever seen.

Together with voice to text through whisper for example we broke the UI barrier.

It takes a little bit of time to rebuild our ecosystem but llms are game changer already.

I'm waiting for a finetuned none fact knowing small LLM which knows everything it needs to know for this specific task.

And I'm waiting until everything critical is rewritten so I can use one ai agent to control my bank, calendar, emails and stuff.

Perhaps through banking read only account permissions or whatnot.

I mean it’s good but all the answers seem to be coding which seems like is the main use for large language models.