Jonathan Blow on ChatGPT Style Things at Producing Software

This clip is taken from one of the streams of Jonathon Blow. That can be reached here:
/ 1721067935
Around two months after uploading this video, I just checked that this stream isn't available anymore. Then, know that this stream was recorded before the official release of GPT-4.
#jonathanblow #programming #ai #code #coding #chatgpt #languagemodel #software #computer #computerscience #gpt3 #gpt4 #aiassistant #gamedevelopment #gamedev #gamedeveloper #development #developer #programmer #softwareengineer #artificialintelligence #software #softwaredesign #indiegames #indiedev #dev

[@waltwhite8126]

The best use I did of ChatGPT was to learn new things when the docs were lacking and even then often the answers were partially or entirely wrong.

[@Darth_Bateman]

Apparently, hallucinations are solved.....??? and I'm just wondering : "When?!"

[@LukeI]

KZbin is sucking me into the Jonathan Blow rabbithole in the same way it sucks people into Jordan Peterson

[@quatricise]

At least this isn't a terrible rabbit hole to be in.

[@MenkoDany]

08:18 Atlassian did a study on this a loong time ago now, but back then they measured that programmers they studied (medium-large corpo environments iirc) spend ~40 minutes "at the keyboard in an IDE" a day

[@TreesPlease42]

The horse doesn't know when its job has become obsolete CGPGrey humans need not apply

[@monad_tcp]

@@TreesPlease42 corporate job is mostly bullshit

[@Darth_Bateman]

@@TreesPlease42 You have clearly never programmed in your life outside of a Udemy course if that's what you took from that...

[@CHURCHISAWESUM]

@@Darth_Batemanthe sad thing is the executives probably think like this guy. Repeating things they hear

[@evadecaptcha]

He's right. I do use chatgpt on occasion to help, but normally it's only helpful when I need to learn something new. For example, I've been a web dev for years, but recently decided to work on my own games which uses C#. ChatGPT has been very helpful learning specific syntax and how things work, and occassionally fixing a line of code if I missunderstand how that code works in C#. However, if I need to get a complex method working, it always does an annoying shotgun approach, with no consideration for the most likely causes of an issue and will still include things I just listed that I already checked in my question. In summary, it's good for pointed questions but not for getting novel emergent concepts working.

[@ryanshea5221]

Haha same boat as you. Web dev for years, returning to my c# gamedev roots. Chat gippity has been really useful for generating rough drafts of ideas or for generating utility functions. It's also great for writing powershell scripts for random tasks.

[@luckerooni1153]

ChatGPT is smart enough to be able to "google that for you," and then tell you some suggestions. It is by no means an actual subject matter expert of anything other than knowing in general what's been posted on the internet before 2020. ChatGPT is not even at the level of replacing stack overflow as much as I wish it was. I can't even get ChatGPT to tell me what good command line utilities exist in Linux and when I elaborate with a response noting the tools it missed or gave an outdated version of, the convo always goes: "Did you forget this? And why did you suggest an old version, what about this instead?" It just goes "Oh sorry you're totally right." Yeah, the chatbot that doesn't even know basic Linux tools is going to replace programmers because it observed boilerplate code reposted on the internet thousands of times and so recognizes it... my ass. If you can be replaced by ChatGPT that's because your job is something that can be replaced by googling and copying and pasting the top result.

[@Aaron-wg6ft]

Similar experience. Asked ChatGPT to translate a CRC algorithm from Matlab to C, then had to debug why it got different answers. Another time i asked it to explain some YAML syntax to me. Got that wrong too. This is supposed to be the low hanging fruit.

[@ifstatementifstatement2704]

That's been my experience too

[@karlosh9286]

spot on ! For learning a new language or module / library / package in a language , ChatGPT can save a lot of time reading docs, trawling stackoverflow etc. But it gets so many things wrong, it can't be relied upon. I've almost had arguments with chat gpt !!! As far as I can tell , It's a souped up google search that tailors the search a bit to what you asked. No more no less. It's useful , it is far from intelligent.

[@channel11121]

Gonna use this when someone tells me something useless from now on: "Everybody could have predicted with 100% accuracy what was in that message before you sent it. There are zero bits of information in that message and it can be compressed down to zero bits." 😭

[@AntiGravityC9]

now that everyone knows that that's what you'll say when you receive useless information, that information too will be deserving of the same response

[@anispinner]

@@AntiGravityC9 Everybody could have predicted with 100% accuracy what was in that message before you sent it. There are zero bits of information in that message and it can be compressed down to zero bits.

[@PhyloGenesis]

"why waste time say lot word, when few word do trick"

[@astroid-ws4py]

Great perspective on the subject, Thanks John Blow ❤

[@hasen_judi]

It doesn't save typing, it saves you looking up all the libraries and function names and signatures and how do you use this function to get this or that result. This kind of work is not uncommon, but it's kind of draining on the brain, so if you can get a bot to sort of write the rough outline of the code you can then just fix errors and then move on to the thing you actually want to program. One time I had chatgpt write the code to load a gif file and extract the last frame from it as an image and save it to a png file. All the library functions exist and are documented, so it's not particularly difficult. I just didn't want to do it at the time because my brain was busy thinking about all the other things I wanted to do that day.

[@SR-ti6jj]

Without watching, let me guess: He doesn't like it.

[@ldandco]

Nop, he loves it. He plans to use LLMs from now on.

[@BigChiken44]

Why would anyone like it? Software is already crap, imagine now this avalanche of AI crap code coming😢😢

[@pythonxz]

​@@BigChiken44It's a tool like anything else where people start using it without understanding what they are using first. It's not magic, nor is it anything but a fancy database.

[@SoftBreadSoftware]

LLMs cannot self-analyse, self-correct, or use basic reasoning because of architectural self-serving bias. Use one for a few minutes and throw contradictions at it. It will agree with everything and find a reason its right no matter how nonsense and illogical it is, no matter how many times it contradicts itself. LLMs arent the tool for programming and debugging and anyone who has used one immediately knows this. LLMs are sycophants.

[@TankorSmash]

Are you talking about everything

[@zacharychristy8928]

All my uncertainty about whether chatGPT would threaten my job went away the second I tried using it to do something useful. It's basically an occasionally better alternative to googling code. If you need to see an example that doesn't exist in the form you want, ask chatGPT and it might make something that's probably right. If we're lucky, it will get as good as stack overflow answers, i.e. usually close enough to correct to keep you going.

[@nataliemreow]

ChatGPT wouldn't, that's for sure. But maybe something else.

[@zacharychristy8928]

@@nataliemreow you could have also said this 30 years ago about whatever was the cutting edge AI at the time.

[@drygordspellweaver8761]

it was helpful to find a required permission on an android app (because you literally need an encyclopedia to keep track of all the android garbage). That was the one use i got out of it so far

[@nuvotion-live]

Not really a good comparison. Googling can’t write code.

[@zacharychristy8928]

@@nuvotion-live chatGPT successfully saving me a ctrl+c and ctrl+v keystroke. What a time to be alive!

[@dmitryburlakov6920]

Well, a lot of things we do are fairly common. We don’t invent large amounts of new things every day, though I agree that good programmers are more system designers rather than typewriters. So by the end of the day it’s quite common for a programmer to write what’s Jon is calling not good code. In fact, many aspiring developers focus mostly on reinventing the things, writing yet another game engine or compiler (hi Jon). Not even combining or bringing new perspective something, just outright building own bicycle. But overall I would agree. What most programmers doing are nuanced. Those nuances are the most important. Those can’t be added by language model, as it’s the opposite of what language model does: looking for the average most likely solution. In my experience, it actually handicaps me a bit. I often started to notice what I would call GPTbugs or CoBugs: the autocomplete looks very much like what I need, I accept that and then when testing I find out that the solution was totally wrong in just one detail. The order of elements, some flag, some missing call that I would likely not miss working with the codebase as I’m aware of the context and system at large. I think I waste more time on this than save from using autocomplete. And I do not think that with the way models work right now it’s ever solvable problems for language models. It actually can very well implement code from papers, and I’m full of joy because of that, as I don’t need to solve solved problems (because most of them are solved only in Python, for example). But it always misses something important, it very much gives you a Chinese room vibes, where result makes perfect sense only if you know nothing about the world outside. Like it can do code, but it can’t solve problem in a way you need.

[@meow2646]

i think chatgpt is useful for learning. Programmers spend a lot of time learning new libraries that are common and trivial and llms can make that process easier and faster

[@dmitryburlakov6920]

@@meow2646 my friend recently asked me what kind of plugins and tools to use for C# and programming in general, as he started to learning. And I said about CoPilot, and the major reason I said that was that it actually taughts to have a good habits for non functional requirements like proper naming and program structure, even when they're writing code alone. The better your naming and comments the better CoPilot works so it's a win win for learning. Typically you would need to learn from your own mistakes, but here you have a reason to do good from the start 😁

[@Darth_Bateman]

How are you supposed to offer perspective on something if you haven't built the thing yourself. . . .?

[@SaHaRaSquad]

I am so baffled by Copilot focusing on code generation when that's the least interesting application. You know what would be an actual game-changer? An AI that can do some basic formal verification of your code, running 24/7. Or AI that has full knowledge of all your repositories, dependencies and documentation and uses that to review new commits. Or AI that can check and debug things a compiler can't, like DB queries. Or AI that automatically handles dependency conflicts. I don't need or want Copilot in its current state, if I want semi-broken copies from StackOverflow answers I can do that myself.

[@mikkelens]

Unfortunately nearly all of these ideas are insanely difficult to do with current AI tech. ChatGPT is what it is, not because people wanted text generation, but because training a model to make semicoherent text is easy because nearly all text everywhere is coherent text. A sentence can be checked by a simple grammar checker and its perfect as training data, if you even need to filter at all. Code (which is also just text) in ChatGPT isn't "checked" at all, and it's not at all difficult to get snippets that won't compile. Understanding a complex codebase, even analysing a developers repositories, is WAY harder to get neat training data for in the quantities that make AI models of today actually usable, because it isn't nearly as accurately detailed. GitHub could make an "AI" or algorithm that notices when a repository is receiving less attention and gives you generic tips, but this isn't nearly as revolutionary as these suggestions would be. I definitely agree that those things would be game changers and probably also way better for society however.

[@astroid-ws4py]

To do that you first need a formal definition of C++ or Python or other languages in something like Coq, Agda, Lean or Isabelle, (BTW, Facebook is working on a project like this for their Hack languages in Coq, it’s in their GitHub page), which are proof verification systems. Then you can prove stuff but it then tends to become a complex process. This is at least hundred years in my thinking till we maybe have such a system and even then I am not sure wether it is really possible. Computers do not anything, They just calculate stuff, Proof is something else, You will never see a mathematician calculates, He will use symbols for everything with no numbers at all. That’s why human will always be > machine. Yes it can calculates tons of numbers really fast, But can it know anything about itself, About the program I wrote to it to do? Can it know that I entered into it a linear equation or a parabolic equation? No! It only executes commands, Not thinks!

[@MiaChillfox]

Those ideas strikes me as something you would have to convince a company of buying. While the current better code completion is something you can sell to individual developers who work on CRUD apps all-day.

[@mishikookropiridze5079]

Copilot X claims some analysis of your PR, but suppose AI can do all the meta work for you, there are dangers of creating false positive cases.

[@MenkoDany]

As others have said. Automatic PR comments in Copilot X sounds great

[@user-ov5nd1fb7s]

ChatGPT is helpful for learning stuff, not for programming. If there is something i am learning how to use, like a library or other stuff, ChatGPT can be used like an interactive documentation thingy.

[@leana8959]

I tried that and it didn't work so well either. It explains one library while referencing the methods and structs in another similar but distinct library.

[@Muskar2]

@@leana8959 It may often be wrong - especially with things newer than its cutoff date (Sept '21) or very niche things - but sometimes it can still can get you started. And for the times where its first answer is useless, and then a follow up answer is equally useless or even misunderstands what I'm asking, I've learned to just not use it, because it's not worth the time. But that doesn't mean that it's not useful sometimes. I think its only useful at the things where you know enough to instantly verify its output validity, but not enough to already have the answer. My most frequent usage is as a reverse dictionary or search for something common that I've forgotten the name of.

[@offensivearch]

Unfortunately it can be wrong and someone learning something is the least qualified person to tell. This is especially a problem because ChatGPT is often subtle in its wrongness. Books are better for learning the basics. Docs are better for specifics. Reading source is the best for understanding code.

[@user-ov5nd1fb7s]

@@offensivearch you need to be very experienced and have taste to use chatgpt effectively, for programming. Otherwise, it will indeed mislead you to copy paste nasty bugs.

[@romanzkv4]

i like your way of thinking and analyzing things

[@scvnthorpe__]

In general optimising aroung/worrying about literal typing of lines of code is silly, vs looking for speed of feedback or soundness of design (and yeah, perf obvs, but remember Donald Knuth). I am a neovim user just because managing things from the terminal is quite comfy and lightweight for me but I don't buy into 'using your mouse is evil maaaan' type thinking. Instead, what makes it easiest to move around and adapt your design?

[@tubbytoad]

Rather than the effectiveness of the tool itself, I think that a person that knows Vim is on the balance likely to be more experienced/passionate about software.

[@scvnthorpe__]

@@tubbytoad I suppose there is some sampling error yeah, come to think of it...

[@anthonypace5354]

It’s real value is in letting you know what it knows to be possible, as long as you go into it knowing it does not have all the answers, as it is not an oracle, and would often rather lie than look stupid. Most of my time is spent using google, wiki, looking into docs, or looking up something in a textbook, so if it can save time with that I could get a lot more done. I do have to double check what it says, but guiding me along, with regard to what to look up, has proved a bit useful. Again, I just can’t rely on it to be accurate, and thus I need to use google anyways.

[@xoutaku7600]

I treat LLMs as compilers for English as a programming language , most of the work people do is repetitive (doing special things is costly) , and yes most code people write is non-interesting code either maybe they can't think of interesting ideas or they have to write the repetitive code (boilerplate , lack of domain experience , ...whatever constraint people may face), and then even the design phase is a process that can be automated into a workflow depending on the domain , adding to that it's daunting if not outright impossible to triumph the collective work of developers that predate you , so you just develop on top of the valuable parts they provided , and if you see it from that point the LLMs may inherit that capacity of reasoning from the statistical data. TLDR non interesting code is fine , not everyone is a special bud

[@nonefvnfvnjnjnjevjenjvonej3384]

noobs lol

[@xoutaku7600]

Lmao kek top lol

[@nonefvnfvnjnjnjevjenjvonej3384]

@@xoutaku7600 what is that

[@xoutaku7600]

The way you talk

[@nonefvnfvnjnjnjevjenjvonej3384]

@@xoutaku7600 thank u uwu

[@someghosts]

Jon’s opinion on this may have shifted slightly since this video judging by his recent tweets.

[@jakubsebek]

link or summary? thanks

[@y01cu_yt]

Here's his latest tweet related to ChatGPT (28.09.2023) : twitter.com/Jonathan_Blow/status/1707403767275196457

[@y01cu_yt]

​@@jakubsebek Maybe you want to take a look at his latest tweet also: twitter.com/Jonathan_Blow/status/1707403767275196457

[@mscottveach]

Blow isn't wrong that often. But when he is, he's always wrong with so much certitude. I understand where he's coming from. This take was pretty much the right take on things from 1990 to 2019. It's the take I had as well. But he clearly has not been paying enough attention to what happened when we combined transformers with huge compute and massive data. He dismisses the potential of LLMs based on the idea that they probabilistically string words together. but never stops to even consider the fact that maybe that's exactly what our brains are doing too. I believe that the undeniable success of LLMs will not lead us to realize we have unlocked some deep, complex understanding of intelligence but rather will cause an existential crisis as we realize that this is, perhaps, all that there is to intelligence. His idea that "if everyone can predict with 100% accuracy then it's not interesting is true" but not everyuone can predict is the entire point. When Dirac says to Feynman "A method of diagramming quantum fields that would be innovative and useful is...." The fact that Feynman completed that prompt with Feynman Diagrams doesn't make the answer uninteresting. It won him a Nobel Prize, ffs. He won't hold this position for long. It's absurd and fragile as the proof is in the pudding. Anyone who' has used these systems knows that the question of whether they'll take over a massive amount of programmer was answered 5 months ago. This is not even a question anymore To dismiss the improbable success of LLMs because you think they're just Markov Chains is be the equivalent of having dinner with Gates and Jobs in 1980 and refusing to invest because, "they're just calculators, guys, and they don't even fit in my pocket!"

[@fizzlefritz9782]

I highly agree with Jonathan Blow's opinion on many things. With that said I agree here; when hes wrong it is spectacularly so. Interestingly enough this certitude in his wrongness is the exact kind you will find in a LLM. Someone who has talked to multiple of such models aside from the ChatGPT and Bard variants would not be willing to claim such things. I speculate these type of takes come down to three things: A lack of engagement with the subject matter, denial that intelligence is a spectrum and the courage to confront the possibility that it's quite possible consciousness is an aggregate present in all things. I would highly recommend anyone who hasn't done so to read Eliezer Yudkowsky's 1996-1999 article Staring Into the Singularity and compare to where we now are, not even 20 years later, and really consider the consequences of this being true.

[@kevon217]

I appreciate his take, but it seems to me there are many straw men living in his argument. A huge component of this all is how humans (coders) will integrate this, tweak it, and expand upon it. Of course out of the box it can be suboptimal, but with clever integrations like agents, tools, etc., there’s a whole new paradigm he doesn’t acknowledge in his argument. Also, the “most likely” answer or even interesting/novel ideas may not literally exist verbatim in the training data OR it could be subtly spread out across many sources, but it could in principle be converged upon/discovered with the right set of prompting, feedback, and synthesis. That, for now, still requires human domain knowledge/understanding, creativity, and perspiration… To my estimation, LLMs provides a rich foundation, who’s information content hasn’t fully been tapped or appreciated i.e., there are infinite number of routes that seem to exist (if you use them interactively and iteratively) and we are just beginning to explore these paths.

[@nowayjosedaniel]

He is right here and you are just wrong. Want proof? Wait a year. Then 2. Then 5. Then 10. Eventually you will stop believing imaginary things and will stop attributing wishful thinking to incredibly dumb "A.I.".

[@nowayjosedaniel]

Maybe your brain is like a LLM. For the remaining humans with cognitive functioning beyond room temperature IQ, this isn't true.

[@Closecall30]

He is absolutely right, but he underestimates the usefulness of e.g. Github Copilot being context-aware and not having to do dredge work. Example: For my game engine, I was missing a cross product function of vectors (because it is 2D), but I have the scalar product. Copilot generated it EXACTLY in my style with the same parameter names. Also, I prompted it to write a function to take a screenshot, and it correctly found my local render texture and used functions of SFML to save it to a PNG. I wasn't even aware of that functionality in SFML. But that's still not actual interesting code, I'd say...

[@astroid-ws4py]

ChatGPT is a documentation accessibility utility or a knowledge accessibility system which can help to retrieve information by asking it using plain English, And this is how we should treat it. AI = statistical and probabilistic model, No intelligence is there in it. All it is returning is based on tons and tons of human data fed to it.

[@mwcz5190]

Very true. IMO, Copilot is at it's best at generating code in a semi-familiar area. If you the programmer are very familiar with the code you're editing, you'll do much better than copilot. If you have no familiarity, then you won't be able to verify copilot's suggestion. Semi-familiar is the sweet spot, since it saves a costly documentation search.

[@AlexDubois]

He is already wrong because we can now have browser plugins.

[@xhivo97]

At this rate it 100% will outperform most humans.

[@AlexDubois]

@@xhivo97 Not hard to predict, there is no other way. However at this stage it is far too early to say if it will plato like self driving car AI did... Until we find a new way to train these models, possibly use better substrat to increase connectivity between nodes. FORWARD-FORWARD algo may be a step closer.

[@44fippe]

Has he tried GPT-4 or just the free version?

[@muscleheadmeatball]

“It refuses to answer your question because it’s programmed by blue haired people in San Francisco” I’m dying 😂

[@martmcmahon]

lol. Yeah, what a nerd

[@ozordiprince9405]

Republican shill

[@Muskar2]

It's far too easy to criticize a small team dealing with going from an academic project to an international audience in a short timespan. Remarks like that one is very narrowminded - but understandable, given that it's not at all his field of expertise. Making indie games is very different and I'd bet virtually no international company or government is interested in the outcome, no matter how technically impressive and cutting edge it is - unless perhaps it's a popular AAA game that is politically controversial or demonstrates some major insights that an enterprise might learn something from, which doesn't interest Jon to produce (as far as I can tell). I think he loves being niche - so he's the wrong person to ask.

[@jupiterapollo4985]

@@Muskar2 Pretty sure Jon wasn't criticizing their work ethic, but criticizing their militant political views and the fact that Chatgpt has so many political-esq filters. That's what he means by "refuses to answer your question because it’s programmed by blue haired people in San Francisco", as people with highlights in their hair stereotypically lean left.

[@luckerooni1153]

@@Muskar2 Who cares about any of the random nonsense you're talking about? Jon loves making games but he still knows a lot more than you, and governments and rich scammers disguised as "international companies" are worthless groups to pander your wisdom to as they only seek to exploit others in the first place, which is why they love the idea of something like ChatGPT, and if you think they aren't full of shit and dyeing their hair blue in San Fran you just have your eyes closed.

[@tuxlu1761]

The point of ChatGPT coding is not creating original research work, but help you learn technologies you don't know yet by compiling documentation , blogs and stack overflow posts, and presenting you in a way that match your context. For example I asked how to iterate a list in a multithreaded way in Rust, or playing an animation on collision in Unity3D, and it gave me valid answers instantly, without having to read too complex documentation, a long blog post, or having to go to the 3rd answer in stack overflow. JB being a 20 year expert on his specialized domain may fail to understand this usage tough ^^ And same , when he says "if the code already exists it's low value", it really matches his views in modern tech stacks and frameworks: of course today when creating a backend in node, doing data crunching in python or other usages you will need to use basic features to achieve your complex goals, and that's what ChatGPT can speed you up. I mean if we go in this line of reasoning, if recreating existing features is low value, why did you reinvent the wheel by creating your own 3D engine ? 😛

[@anderdrache8504]

I think you don't understand what he values. It's about understanding what you're working with instead of using it to generate code you can copy-paste. Sure, it can act as a better google but using stack overflow to solve basic problems is already the wrong mindset in many cases since understanding and learning from your code is more important than getting the function done as quickly as possible. ChatGPT helps you "skip" the learning step, which means beginners can be temporarily more effective but if they don't write simple code themselves, they will always stay beginners.

[@parzival-3141]

I don't think what he said is at odds with your use case, though he'd probably still disagree with you (I agree with @Anderdrache). Also, it doesn't seem fair to compare using a language model to generate boilerplate or trivial code to building a custom game engine. Like he said that's something that would require a sophisticated general AI.

[@tuxlu1761]

I don't know, but when you try chatGPT you can ask him to explain code you don't understand, and it works pretty well, often for ASM even! Obviously GPT output may be inconsistent and wrong, but truth is it's already pretty reliable. Then I did use GPT on technologies I already knew and understood, and double checked GPT's output. Of course copy/pasting without understanding is a sin ^^ And for the game engine it was meant to be thought provoking, of course it won't create a full engine if I ask, but I can ask him how quaternions work, how to do backface culling, use specific Direct3D features... (I have no more examples, I'm not an engine developer!)

[@tuxlu1761]

Just view GPT as a tool to accelerate your workflow: if you're such a good developer that you know your technologies so well, you don't need GPT more than you need syntax highlighting and code completion then 😛

[@nightmaregam3r704]

@@tuxlu1761 Accelerate? More like bugs waiting to happen.

[@denisblack9897]

a chatroom with 100x GPT's doing async calls can generate a frankenstein project for you to start from in a few minutes i do this abusing coffee and drugs cloning dozens of projects similar to what i'm tasked to create and manually gather perfect snippets out of them, then i get rest and start fresh in that monster of a project already containing 90% of what im gonna need async ai agent could save me a lot of health

[@eljapel]

Love it, and I feel exactly the same

[@JianJiaHe]

I think ChatGPT is good at helping the small things like naming a function or variable, it could also serve as a fact database. But it’s not good at checking complex stuff like evaluating the architecture of the entire project, doing a very specific thing like writing a long sql command that does a very specific thing, it’s not going to give the right answer.

[@karlosh9286]

I always plug my laptop into an external keyboard , mouse and monitor . Rarely does it get used on the laptop keyboard, track pad and screen !

[@pierreollivier1]

the only completion needed, is variables, and function completion, I like to be descriptive and to me completion and lsp is only useful for preventing typo in code. Would never use copilot.

[@pythonxz]

The most senior programmer at my job was baffled when I told him that I wasn't really using Chat GPT. He uses it regularly, and I just use it to "rubber duck" sometimes. This is the guy who likes to tell me that I have "magical" thinking when I am solving a problem. Maybe he's projecting.

[@seriouscat2231]

Maybe someone should gift him a real rubber duck.

[@ifstatementifstatement2704]

"One of the things that chatGPT does is it refuses to answer your questions because it's programmed by blue-haired people in san francisco" LOL. This video gets a like just for that statement lol.

[@yash1152]

7:27 when u r programming, u dont type that much yeah, n/vi/m proponents and notepad++ author john doe (or whoever) also said that he's very slow typer.

[@davidjohnston4240]

Taken as given, that in my field I know my shit. That field is random number generators. I wrote the book. I probably designed the RNG in your computer. The internet is full of very wrong opinions on the theory of RNGs. Quora, stack overflow, anywhere else. It's mostly wrong. So when I ask questions of ChatGPT concerning these matters - it gets it 100% wrong, because it's been trained on wrong data. ChatGPT in its current form cannot advance knowledge, because it cannot derive new things from things it knows. It can only make things that reflect the things that are out there on the internet. I asked Google's equivalent about myself and it said I was an expert and leader in my field (I am) then threw in that I am a fellow of the ACM and IEEE. I am neither of those things. But I bet a lot of leaders in their technical fields are fellows of technical organizations, so it looks like the kind of thing that would be said in that context so it said it. Try asking tricky questions about something you know deeply. You will not be impressed with the result.

[@ldandco]

Quick question, do I go for a lava lamp ? 😉

[@davidjohnston4240]

@@ldandco If you like. Although when you take in image of the lava lamp and hash it down to a random number there will be more entropy in the sensor noise than in the lava lamp image. You would be better off putting a lens cap on the camera and turning the gain up.

[@davidjohnston4240]

@@ldandco If you like. But if you point a digital camera at the lava lamp and hash the resulting image down to a random number, there will be more entropy in the sensor noise than in the lava lamp image. So you could dispense with the lava lamps, but the lens cap on and wind up the gain. Then you will have even more entropy and no one else can see the source of entropy whereas with a lava lamp, anyone can see the state of the lava lamp unless you hide it.

[@ldandco]

@@davidjohnston4240 Thank you ! I think it actually makes sense, and by your description, I think I started to make sense of "entropy" too ! I would've imagined high res sensors these days would project a very stable image, although thinking about it, you can't control external light, therefore the input of the lens will be random either way no matter the lens. Conclusion: lava lamps make for an overcomplicated and fancy way to generate RNG

[@ldandco]

@@davidjohnston4240 Always have found the subject of RNG quite interesting. There are some theories about it also being interlinked with quantum mechanics which used to take me into a rabbit hole. Forgive me if I am just talking nonsense, you are either way and obviously the expert here.

[@u9vata]

I made chatGPT write a sort algorithm that did NOT exist before - I mean it existed on paper on my desk, but I was at the moment too lazy to type it in. I had to force chatgpt a few times to not do the bad thing - but in the end it was also a faster-than-ska-sort algorithm in the end. It "felt" like conversation with someone who is junior and took some time, but end result was totally found nowhere alg. Its not faster than my magyarsort, but uses less memory. Still: it did not find the algorithm on its own, but it could be faster to do this way than me typing in into vim - this was first time I realized this thing can indeed help be more 10x these days. Again: Things very close to it does not exist - also this was not "totally" designed on paper, just had a "I had to sketch this down on paper" kind of little handwritten picture with few words ;-)

[@SpaceApe020]

But you have to keep in mind that most people don't spend their work days writing algorithms in the first place. ChatGPT doesn't understand the product and solution that you're working on, it can't, so you can't rely on it to make your day to day decisions.

[@u9vata]

@@SpaceApe020 To be honest in he not algorithm part, product development code its just even more useful than for algs... The reason why I talked about alg develolpment is because "its also good for that somewhat" - I should point out that not by itself, but as a human + chatgpt it is a gain over just human alone. Like having a not so smart junior on your side you can instruct to do stuff... Yes it will be sometimes wrong or bad, but usually worth it after you overseeing the results. Not always - sometimes you can be faster just alone (which is again true if you would have a not so smart junior on your side).

[@ShotgunRaider88]

chatGPT is very good at solving high level programming problems. it can very easily fix snippets of code. Low level stuff is still to be tested because there is less context for the AI to use

[@xxxanonxxx]

Simple Example: I want to open a data sheet from a json or csv file and turn the header information into a pdf. Imagine having no idea how to that and searching how to do that on google or youtube. You'll be clueless for hours on doing something so simple. ChatGPT will explain exactly how to do and can spit that code out faster than anybody ever trying to show me. Simple code that is impossible to find for beginners because everything is behind a paywall or adblocked. Why do I need to click a million links for such simple code that is "everywhere"?

[@leslieviljoen]

GPT4 absolutely does understand the code - it can explain it, analyse it for errors or security issues and modify it. Copilot saves me a little bit of typing, it's probably not much. GPT4 is much more impressive from what I've seen.

[@nowayjosedaniel]

It is impossible for ChatGPT4 to understand the code. You dont even know how ChatGPT works, despite the information being freely available and overwhelmingly simple. So why should anyone trust anything you say? Predictive text can never develop intelligence. It's impossible because all it is and can do is predict the next likely token in a language.

[@bobby_tablez]

I agree with some of the points here, but saving time typing is only one aspect of code completion. The other part of code completion not talked about here is remembering the complete name of identifiers. We're all used to it, there's an identifier in some library that we remember part of, then code completion steps in and reminds us of what we wanted. And that's great. Though I don't use it at the moment, ML powered code completion could certainly do a decent job at this, and isn't a sign that you're writing meaningless code.

[@doltBmB]

if you can't remember it is a sign that it's poorly designed

[@bobby_tablez]

@@doltBmB no developer ever is going to remember every identifier name they need to use always.

[@doltBmB]

@@bobby_tablez they used to, you're just so immersed in the culture of overcomplication that you can't see it

[@bobby_tablez]

@@doltBmB autocompletion is overcomplication??

[@seriouscat2231]

The best thing is when it autocompletes your variable names from someone else's codebase.

[@carlsmith8593]

I would happily carry a 1" thick laptop, if it included a mechanical keyboard.

[@LJ-to8in]

'It's just a super markov chain where it's putting words together probabilisticly'. For all we know that could be how humans do language/programming too. There might just be a probability gap in accuracy that AI is rapidly closing. For sure it's not a suitable substitute for programmers yet, but time and time again, even the most wise and skilled people like Jon who say that 'this new technology can't do X' have to eat their words faster than anyone expected. I swear that is my real opinion and I'm not just simping for Roko's Basilisk.

[@xhivo97]

Maybe, he's definitely wrong though GPT4 proves this, and I don't mean the public chat version. Here's a comparison, the raw version can actually generates a full working 3d game compare that to ChatGPT-4 which cannot do that, it won't even try it'll refuse the prompt if it's something like "build a 3d game with x mechanic" it also passed most or all of the programming interviews, and outperformed the average human on leetcode. Also it solved some really difficult math problems that only the top humans can do, people study for months for those competitions.

[@gianni50725]

I think saying "that's how humans do it" is pretty dumb, because it's obvious its intelligence is different from ours. It does extremely well on some tasks most humans would struggle with at least a little (e.g. writing a poem that rhymes) but extremely poorly on others that are trivial (the transparent Monty Hall problem before it was RLHF'd)

[@LJ-to8in]

​@@gianni50725 That would very dumb if someone said "that's how humans do it". Almost as dumb as saying "It's obvious that humans don't do it that way".

[@gianni50725]

@@LJ-to8in It's good that I didn't say the latter then, right? I said its intelligence (from what we can directly observe) very clearly operates differently from our own. If you want to point to "for all we know..." -- well, for all we know, entropy breaks somewhere and the heat death of the universe won't happen. But that's totally useless to say when the evidence contradicts that speculation. You're just appealing to ignorance. Nothing in neuroscience implies that we are predicting words from a huge corpus of works we've digested. You didn't even need to bring up anything relating to how humans work, you don't need to operate at all like a human in order to totally out-compete us in a variety of tasks (like a calculator). Who cares if it thinks like us? It doesn't need to. It just distracts from your broader point and loses focus.

[@LJ-to8in]

​@@gianni50725 This is a much more fair criticism of my comment. You're right that It isn't useful to appeal to ignorance, (It is fun though). What evidence contradicts the speculation that humans predict words from a huge corpus of works we've digested? I genuinely want to know. I think Jon would care if an AI could think like him or better. He at least wouldn't be so dismissive of the usefulness of AI as a programming tool.

[@aerocodes]

I agree with this take looking at his perspective of how things are at the moment. Of course, if you ask chatgpt to print out a tiktaktoe it will do it in the most simple way. But GPT-4 is actually really powerfull when it has the data or the direction that you are aiming. So yes I agree that a general LLM will be general, but what most developers don't see is that these things will be very soon modelled towards the user/programmer

[@xzaz2]

This guy when color TV's came out: All it does it putting color pixels on a screen.

[@seriouscat2231]

If that's all you care about, then fine. The point is in the meaning of the words. Or pixels. And no AI or LLM can guarantee there's any meaning to what it outputs. Except to the extent that has been put into it during its training. It's like a well you need to carry all the water into before you can take anything out.

[@user-eh5wo8re3d]

Yes, LLMs predict probable sentence continuations. The whole thing with Neural Networks is that they approximate high dimensional functions. Who is to say a good approach to sentence continuation is not actually approximating some sort of "understanding" of topics like coding. Sentence completion is the interface, the surface, not the whole thing that implements that interface

[@seriouscat2231]

It actually is the whole thing. A few university courses on AI will tell that to anyone.

[@wilfridtaylor]

Chat GPT is really good at Microsoft code documentation. Since the actual docs from Microsoft is so bad even GPT with hallucinations is more useful lol.

[@kubre]

Nah they have seen what happened to older ms chatbot from Twitter, last thing they want a consumer product to spew stuff would get banned

[@SuperLazyCat]

What about when it's added with copilot? I would like to see a update on what he thinks now.

[@k6l2t]

I used copilot earnestly for several weeks of real work, just to see if it actually helps me develop things "better" by any metric. I was not impressed. Copilot is great a generating lots of crappy code, and you are constantly spending crazy amounts of time "code-reviewing" it, which completely defeats the purpose.

[@Zoronoa01]

"programmed by blue haired people" xDDDDD

[@Georgur]

So the point is 'this just helps automate routine stuff' and supposedly that means that it will not replace anybody? Automation of routine stuff it's like summarization of any technical progress that had been done in human history. And as the history tells us there are cases when it in fact killed some professions

[@monad_tcp]

"this stream was recorded before the official release of GPT-4" Nothing changed much, all his statements still are true. GPT4 is just more convincing at producing something that looks right at the first glance, but its completely bullshit. Its just that GPT-3 wasn't that good and you could tell even if you're a bad/beginner programmer. The only goal of optimization of a large-language-model isn't understanding, just passing the test of creating something that looks real so you can create hype and marketing, so you can pad up your pocket with money.

[@jeffwells641]

Stuff like ChatGPT can help with boilerplate, and it is going to better the more common the problem is. The harder your problem the less useful it's going to be. Case in point in another field: A lawyer recently used ChatGPT to help write some motions, and there's a half decent chance he's going to lose his license for falsifying the legal cases the motions were based on. All the cases were either non-existent, or did not say what ChatGPT said they did. He's in deep doodoo right now.

[@lucasjames8281]

His point about laptop keyboards causing injuries arent true really. Yes a flat and unergonomic keyboard can give you repetative strain injury, but ALL laptop keyboards are flat and un ergonomic, from the shitty 2016 macbook butterfly switches, to the thiccboi 2000s thinkpads, your wrist and hand position is basically the same with all of them. There is no laptop with an ergonomic keyboard that will force your hands to maintian correct posture, doesnt work on a flat plane😂

[@timstevens3361]

so microsoft bought github which is full of programmer software projects. did they do that and also build free vscode text editor that runs on windows linux n macs out of the goodness of their hearts, or did they want EVERY piece of software being written to be input for GPT4 and chaGPT A.I. models so they would have this programmer god model for themselves ?

[@_megla]

you can set the video speed to 1.5x and it just sounds like a normal person talking

[@AbuAl7sn1]

so true , even my english is shitty

[@carlinhos10002]

It means he is thinking before saying

[@_megla]

​@@carlinhos10002 no

[@carlinhos10002]

@@_megla yes

[@0xCAFEF00D]

I don't think people who say things like this have any contact with real humans.

[@haskellelephant]

I think this description is changing massively every day. Now, with the browser plugin, I can ask it to look for bugs in open source software by directing it to github. I have asked it to look for and solve issues on github repositories and it does outline the correct action. Going from that to action is essentially what autogpt is trying to solve.

[@mac2105]

I've not been working as a developer for as long as Jonathan has, but I'm still absolutely certain that he is wrong about this.

[@ldandco]

can you elaborate ? why do you think he is wrong about this ?

[@mac2105]

@@ldandco AI is absolutely coming after programmer jobs, just work with chat GPT and you'll clearly get the sense that a large amount of work done by humans today can be automated using similar technology. Obviously it's not gonna replace eccentric, opinionated star indie game developers first, so I understand why Jon doesn't feel threatened ;)

[@ldandco]

​@@mac2105 I am a software engineer, (20+ years experience) I use chatgpt on a daily basis. The code I work with is somehow complicated, multithreaded, high speed algos. (fintech) Chatgpt has helped me, but I would say it has only been an improvement of 10% to 20% on my speed. can it understand highly complex environments and work on top of them ? well, if you explain to it, yes, but it will make 1000s of assumptions because the size of my prompt is limited, therefore, i will spend so much time explaining to it so much, that i rather work alongside with it, and not delegate the entire work (you can't, not yet)

[@ldandco]

@@mac2105 my point being, at this stage, chatgpt is definitely not a solution to replace programmers. There are so many risks a company would also take on if they decided to let an AI code their entire code base. I think in the next 10 years, if anything, it will enhance developers skills, but ... think about this... it will be paradoxical in a sense. Many people that once thought about becoming software engineers, will rethink their career due to fear of AI taking their jobs, creating a subsequent lack of developers in the market that can skillfully deal with these technologies. In a sense, there will be a shift.

[@mac2105]

@@ldandco The AI isn't going to code their entire code base, the AI is going to make individual programmers many times more productive so they'll end up needing only one or two really good AI coders instead of 100 regular ones. Obviously AI output needs to be supervised and tested, just as it is today.

[@jason_m2003]

Disagree, ChatGPT is like ask me anything documentation for C++, Java, Python, Javascript (React), Shell, Docker, SQL, Etc... I think what needs to be learned here is how to interact with LLMS to optimize workflow, not bash LLMs for their current limits, heck it's helped me 10X my development in DevOps

[@fennecbesixdouze1794]

I am so excited for ChatGPT to keep getting better, it's already great at drudgery bullshit that I don't want to do. I needed some stupid CassandraDB migration in C# today for some shitty portion of a system I don't care about but absolutely needed to be done. I don't know how to use Cassandra or C# so I just asked ChatGPT to write it for me, and what it spat out looked like about what kind of bullshit I expected, except it got all the annoying shit I didn't feel like looking up right. I noticed a couple things I didn't like so I literally wrote out to ChatGPT what I didn't like and it fixed those things for me. Then I asked it to write the unit tests, and I read the unit tests and confirmed they were doing exactly what I wanted except like two lines I changed. Then I ran the tests and it didn't work because there was a missed null check so I added that. Then I ran the migration locally and it worked so I committed it. This entire process took me 14 minutes, which is less time than it would have taken me to even describe the task to a junior dev, who would then likely proceed to take 4-6 days reporting "sorry my local env is giving me a hard time, I reached out to so-and-so to help get me set up", and then at the end hand in something that would likely crash the whole production system. That alone saved my company thousands of dollars, and I'm only finding better and better ways to use ChatGPT daily. That being said, I've seen several other devs attemp to use ChatGPT, and they miraculously do a terrible job at it. Indeed, it turns out that writing prompts for ChatGPT is itself a form of programming, that takes intelligence and expertise and taste and skill to do it well. In other words, ChatGPT is not in any deep way different from other types of programming tools we've seen before. It's another programming tool that good programmers will be able to use well to automate away boring tasks, and bad programmers will use poorly and make the world a worse place.

[@AJ213Probably]

I do think we underestimate it's future capabilities. I do agree though. I am waiting for automatically catching bugs and code analysis. Because we easily make the same mistakes. No way that can't be caught.

[@asdfqwerty14587]

I actually think that LLMs have a pretty hard limit to their capabilities. Fundamentally, a LLM is always going to be limited by the amount of data available for it to be trained on - there's a lot of data out there, but not an infinite amount of data. For simple tasks this works well enough... but the problem is, that as the tasks it's being asked to do get more complicated, 2 things happen: 1) It requires more data to be able to find out what the pattern is, because the pattern itself is more complicated and 2) There is less data available for it to be trained on.. because the pattern is more complicated so fewer people are doing it. Put those 2 things together, and you can see why it's always going to struggle with more difficult problems. It requires more data to understand the pattern *and* there's also less data available at the same time - it's kind of like the equivalent of trying to use something like an O(n^2) function on a database with billions of entries. Something like that can work for small datasets without much trouble, but the algorithm is just always going to struggle on bigger datasets even if you improve the hardware.. and likewise, LLMs will always continue to struggle with solving more difficult problems for a similar kind of reason (except that you're looking at something different from time complexity) There might eventually be an AI that can do the kinds of things people are imagining.. but I can't imagine that it will be a LLM that does it.

[@AJ213Probably]

​@@asdfqwerty14587 I saw a video explaining how they were able to overcome this data limitation for math/logical proofs by generating hundreds of thousands of proofs. So while the available data pool of proofs was low, they were able to generate them leading to the AI to be better than most people for solving geometrical math problems except the best. At least this was my understanding from watching the logical reasoning AI video from Sabine Hossenfelder. I am not sure how this would apply for programming though.

[@bram_adams]

this take is well formulated but is obviously wrong on a number of points and comes off defensive. 1. he says no code that is probabilistically similar is interesting. this is false. the letter "e" is used four times in the last sentence alone. does this mean it is not used in different ways to create an interesting sentence? there are 26 letters total that can form any word that has ever existed or will exist in the english lexicon. there are only three control structures in programming (for-while, if/else, function calls). these three alone make up every single piece of software on any turing machine that we have ever used. probability in a token set is not the same as probability of a whole set. 2. his point on typing was contradictory, at the start he empathizes with ppl who hurt their hands from shitty keyboards but spends the rest of the video saying the speed up on keys typed from gpt are useles? hunh?? 3. he obviously has formed an opinion before using these tools, if he did he would know that hallucination is trivially solved in small datasets with embeddings and as context windows increase you can preload more data. its like hes asking a 64gb computer to hold a HD version of skyrim, four copies of ableton and fcpx, and all of his gh repos. you run out of space bro!! speaking of space, any one mathematically minded or trained in neuroscience will understand that the latent space of probability is all there is for any type of intelligence, be it human or machine, or evolution based.

[@maertscisum]

True, but writing code doesn't always mean creating new code. Most of the time it is for customizing or modifying existing code to support new scenario. So don't dismiss GPT too easily, it is still useful for automating a lot of tedious coding work. Even for new code, and even if GPT can't get it right, but it has capability to combine things and suggest quite amazing code.

[@llothar68]

And we for sure don't need AI for refactoring. We need deterministic code transformations.

[@FIIRdesu]

Probabilistic word prediction already requires some understanding. While the current ChatGPT level of understanding is not enough for non trivial progamming, i would not presume that will stay that way forever, i see no fundamental reason that would prohibit it leading to general ai eventually.

[@noxabellus]

I use chatgpt and copilot all the time. It is legitimately useful if you understand the points Jon is making and work with the system rather than just throwing up your hands like a frustrated child

[@nowayjosedaniel]

I would love to see the code or projects of people who use ChatGPT all the time. I predict with extreme confidence that either the projects are insanely mundane or the programmer cant actually program. Do FizzBuzz without ChatGPT's help.

[@metropolis10]

Seems like this take is missing two things. YOU might not write this code every day, but over the whole of all programmers, someone out there has probably done this before. Maybe even in a different language. It doesn't need to be predictable for YOU, but for all programmers. Second, almost all computer programs can be generalized as copies of some kind of pattern. Most apps are some kind of CRUD database frontend. A platformer game has so many common elements that is shared by all platformers. What makes these apps unique is usually a small twist or flavour, or something specific to their problem space. But that isn't the bulk of the code. The bulk of the code is typical and near identical, maybe with a few variables tweaked, to a similar app. I feel that is lost in his take.

[@supernewuser]

I highly doubt the reason we type ~40 minutes at the ide per day is because we're reasoning about our problems the rest of the time. Also, I think he'd be very surprised at how far things have actually come.

[@nowayjosedaniel]

3 months? Why would he be surprised at something that has gotten worse, not better?

[@pugbread2873]

i'm surprised he could be so incorrect about this.

[@Muskar2]

Not a very insightful perspective honestly. Sure, ChatGPT sucks at questions that require a certain level of context or technical knowledge, but that doesn't mean it's categorically useless, nor is programming things that other people have done BS. Not everyone today started programming before the 1980's and have basically learned everything that's worthwhile to learn from others. Or if somehow most other people skip the entire educational stage of standing on the shoulders of giants, I'm much worse off than I expected and will never amount to anything meaningful. Fortunately, I'm naïve enough to not believe that. To me ChatGPT is best at being a reverse dictionary, when an expression is at the tip of my tongue but I just can't remember. And often also as a sparring partner, kind of like having a streamer chat: even though it's remarks usually aren't helping, just voicing them to someone/something that seems interested and getting something back sometimes helps my brain getting unstuck. Perhaps it might be better to have a streamer chat (or friends/colleagues who care to listen), but not everyone has an audience and there's always a percentage of people who are toxic, so ChatGPT is an attractive "junior assistant" to me sometimes.

[@jason_m2003]

Agreed

[@Muskar2]

@@jason_m2003 Coming back, I have no idea what I said here haha: "Or if somehow most other people skip the entire educational stage of standing on the shoulders of giants, I'm much worse off than I expected and will never amount to anything meaningful. Fortunately, I'm naïve enough to not believe that"

[@nowayjosedaniel]

When students are learning from something as dumb as ChatGPT4, we are already far past Idiocracy and close to total societal collapse.

[@Muskar2]

@@nowayjosedaniel I agree that there should ideally be bigger warning signs than it has now, that says you shouldn't rely on it for information, and that it's basically inviting learned helplessness if you do. But I don't think misanthropy has ever made the world better - providing value to the world, on the other hand - has. So if you think students should learn without being "baited into the easy way" of relying on GPT, frameworks, bloated libraries etc., then the best way to change that is to investigate and solve that problem. I think every positive impact counts - just as every negative does. If you're just having a bad day and ranting out stuff, that's fine too. Take care and good luck

[@creepychris420]

bro you don't know what you are talking about. i have been a developer for 10 years, and now i am using ai to code my code and debug my shit. You are incorrect. Copilot produces working code nearly every time. It now has chat which will generate markdown codeboxes with buttons (and hotkeys) to insert at cursor. The code is great, it understands the context, it sees the files in your ide. I understand the concept of the most relevant answer, but this is magic somehow, you need to try it. I now have access to copilot chat, copilot cli and regular copilot. It's great. I have experience with devops and servers for 10 years but this is changing the game. i don't man anything anymore. i type shit like man echo into chat instead. there is no man echo in linux for bash echo, but chat will tell you how to -ne if you forget. new copilot gives changes to a whole document with a diff and an undo button and a chance to revert parts of the change. it lets me ask stupid things like 'refactor this to not be ugly', or prototype a piece of shit monlithic function 'refactor this to use functions' and get some usable tooling out of it. i write a lot of html for fancy product descriptions and as soon as i start writing it it knows what we are doing and i just tab my way through the day lol. i paste giant tables from my writer bro into comments and copilot creates it for me. sometimes i do cool shit like colspan="3" for little subheadings, maybe that shit links to some product name, i do it once and copilot will tab that shit through 100 rows of spec table no problem. new copilot can see your selection as well as your doc, last night im decompressing some tar'd backup in shell and i highlight and say /fix these args are wrong and im too tired to fix it - i was compressing not decompressing - shit just knows this from the rest of the context and function names and changes it and im like lol oh right. go to bed bro 😅 its my rubber ducky and my documentation and my second set of eyes. no formatter for shellscript? no problem. copilot format this script. ezpz lol. copilot cli lets me type a natural instruction in my cli and get the command back with ?? - i can then just run it in my current shell. This shit is changing my life and giving me more energy to do cool shit or enjoy my life after work. don't hate it man. it works for visual studio too not just vscode, you don;t need to be mad at it coz you can't use it lol. go pay ur $10 and click the buttons to get access to chat and cli. its amazing bro don;t be scared. im on waitlist for copilot chat lol i wana shout at my ide

[@kevinfischer4869]

Glad you’re loving it too! I do more coding side projects now since the tool saves me time. It’s absolutely making me a better programmer!

[@nowayjosedaniel]

Thanks. We now know that despite your experience, you are and have always been horrible at your job. People who think ChatGPT is some amazing tool must be the worst programmers around. You're REALLY lucky most employers are computer illiterate and thus incapable of knowing a programmer's level of competence.

[@painterc15]

I disagree with Jonathan's point that if you're writing "statistically average" code you are doing something wrong. In theory most things we need build as developers has been built before, but in practice even if its been done before it's unlikely that the code is suitable to be integrated into a new project. This is where an LLM like ChatGPT might come in. If it can encode the essence of a particular piece of software and transform that essence between different language representations and adapt it to different interfaces, then we will actually live in a world where we can stop reinventing the wheel without having to accomplish the impossible task of standardizing the entirety of software development under a common set of languages and interfaces. Kind of like the Rosetta Stone of programming. Then we can really turn our focus to the interesting bits he talks about.

[@stanbme]

okay, boomer

[@nowayjosedaniel]

Zoomer mad at GenX bc they didnt fallfl for room temp IQ hype from silicon valley techbros? Yikes.

[@MrMastrsushi]

"it doesn't work because it's programmed by blue haired people in SF" Tell 'em!

[@maxblank7222]

I've been calling it a glorified Markov Chain for such a long time, hearing the same opinion from someone as intelligent as Jon gave me a huge ego boost

[@arachnid4910]

He’s wrong

[@seriouscat2231]

@@arachnid4910, you should try to understand the math behind the models and the code that runs them and you probably would agree.

[@xn4pl]

Says the guy who made his own programming language and game engine instead of using the miriads that already exist.

[@nexovec]

See? People are not responding to you. They are not responding because you don't know what you're talking about.

[@nowayjosedaniel]

Thank you for showing us you should never again be heard or taken seriously. You have to have the saltiest dumb comment in this whole comment section.

[@seriouscat2231]

The word myriad is already a plural, synonymous with multitude when used as a noun and with innumerable when used as an adjective.

[@boekhoffa]

The model has a trillion dimensions and can understand virtually every human language. Comparing it to a Markov Chain is being very dismissive of the achievement. I know that Jon prides himself on being an excellent programmer and it is probably very uncomfortable to consider that the superhuman version of this skill may soon be widely available. It's not that understanding computers or engineering jobs are likely to disappear, but it does seem likely that things are likely to change rapidly from here on out.

[@nowayjosedaniel]

ChatGPT or any variant is incapable of being superhuman. You're already looking at it at its peak. There isn't anywhere to go because it isn't capable of intelligence in any form. It's literally impossible, as it is literally in the name. GPT. Literally nothing to do with intelligence.

[@seriouscat2231]

A superhuman array of numbers that operates purely on numbers and multipliers between them and then maps those numbers into words at the output. Usually people who understand anything about the code or data structures that make up and run the model also know there's no place for anything superhuman to occur.

[@boekhoffa]

@@seriouscat2231 If the model can perform a task better than any humans isn't that the definition of super human? With sufficient knowledge and compute it would be possible to accurately simulate the human brain with arrays of numbers. What difference does that make?

[@seriouscat2231]

@@boekhoffa, it depends on your definition of performing a task.

[@peersvensson9253]

Blow is way too idealistic. Most work isn't creative, it's rote. Most programming isn't systems programming, it's web and app development. A typical Jonathan Blow rant is basically "this is how I think the world should be", meanwhile the world doesn't conform to his wishes at all.

[@marcocaspers3136]

I'm a developer, but I don't code 200 days a year. Coding is however not my only responsibility. So it's quite insulting to hear from your mouth that you consider developers that code less than 200 days a year to be slackers. It's a rather shortsighted comment from an intelligent person that should know that the remark adds nothing to the conversation or the math.

[@y01cu_yt]

If coding is not ones only responsibility I don't think they are in that slackers category.

[@marcocaspers3136]

@@y01cu_yt But that's the whole point, you set a boundary, but his remark was general and unbounded.

[@doekewartena5729]

I think it's rather shortsighted from you to be insulted by it.

[@theitatit]

"I don't code at all but I'm a programmer"

[@vladinosky]

Pffr come on, you are giving him far too much credit. No one can deny he has made good games but just hearing the statement: "if you are not spending 80% of your programming time doing something completely original, then you're doing it wrong" made me chuckle. No wonder he hardly can work with anyone else and takes a decade to release something. He's undeniably a very self-important person and therefore he obviously is better than everyone else including you, if you catch my drift.

[@noxabellus]

"that should be the vast minority of what you're doing" - guy who has never innovated a single thing in his life

[@nowayjosedaniel]

The amount of seething jealousy in your single sentence is absolute incredible. I dont think I have ever seen someone so jealous and angry at someone for being more competent than them. Why does JBlow make you feel so small?

[@noxabellus]

@@nowayjosedaniel lmao reign in your fandom there bud

[@seriouscat2231]

@@noxabellus, it's spelled rein. Horses have reins. Kings have reigns.

[@noxabellus]

@@seriouscat2231 thanks