Being a developer in the age of reasoning AI
near.tl
near.tl
As with every openai release, I’d take it with try a grain of salt until it sees wide public usage.
So far every iteration of gpt since 4 has been pushed as a replacement for developers, but I have yet to feel that these tools are even consistently useful for my day to day.
But I don’t think many people are that worried about gpt4. We’re worried about what it’ll look like with 5-10 more years of progress. O3 is apparently already better than 99% of competition programmers or something - it just costs thousands of dollars of compute per query. But give them a few short years.
I’m still no expert. But I have a lot more confidence in that query & how it works than I would have from cobbling something together from stackoverflow.
I don’t discourage use of LLMs but I don’t encourage it either. If you need to pay OpenAI to get your work done then that’s on you.
Just curious, does it matter? For some simple boilerplate code, simple test cases and simple skaffolding, which is needed in most applications, do you even care?
I’m not sure I understand what you’re asking though.
result, err:= doSomething()
and it'll generate the rest around it. Every time. Faster than I would type it and with more eye for detail (I am lazy, I will forget things as I did it now the 10004th time). So you can imagine some boilerplate scaffold thing that would do this err handling for you, but if you have a custom one, for 'old' tools you would have to tell the tool; now it just does it.
It especially becomes clear in frontend tasks: it just generates 100+ lines of react that work and look good; who likes writing they type of thing manually? I know 0 people who do: after that you will want to tweak it, but the 100k+ lines of react we generally have in a project are not something we would want to write these days from scratch. Which old school scaffold etc tool does that? So that it is 80-90% there for SaaS first shot? I know nothing (doesn't exist) without the lovely verboseness of having to type a million tags?
It saves us a lot of time we don't have to spend on boring stuff. We only have to write the actual business logic and data models; the rest drops out. And of course we only have to write that very loosely, not formally.
What is boring is the time I’ve wasted on a recent project fixing all of my coworkers LLM mistakes. I think I’ve nearly rewritten the entire php wordpress plugin. So his work really was a waste of time.
For golang I have never had a problem writing if err != nil. It takes a few seconds, paying some LLM company to write something I could write myself but I don’t want to because “I’m bored” is ludicrous use of time to me.
Optimizations give better results when care is taken first about the most costly parts.
Ever heard of snippets? They do that too, but much more energy efficient and faster.
create-react-app and yeoman
I also feel they’ve made responses fairly wordy to make it appear to know stuff.
No one does that. You go to school and learn numbers and basic arithmetic, which gives you all the formulas you need. And even illiterate people work out quickly the minimum they need to do transactions.
> This ability to abstract logic is what separates developers from users
Then everyone is a developer as everyone can write a serie of instructions which abstract action from context. It just that the computer is very limited and you need to be very precise. But what it can do, it can do it very fast. What is instant for us is an eternity in CPU time, and you can do a lot with simple operations if you can do a great number of them.
> Just like a developer's brain can think about a problem and come up with a solution using code, o3 generates a program (i.e., metadata) on the fly to solve the problem
Again that's not happening. You think about a problem and solves it by constraining yourself to the basic operations you have available (the APIs, languages, libraries, platforms). The code is just the written form of that solution. You can't invent operations that does not exist, you can only group them under a new name. Each set of names is another layer of abstraction that make it easier to write the code. But the constraints don't go away.
The code is not how you create a solution. It's how you instruct the machine to replicate the solution you came up with.
As a self-taught developer I absolutely promise you don't need to go to any school to learn the requisite skills to achieve a variety of complex things. You absolutely can learn from both trial and observation.
Yes, everyone actually is potentially a developer, because there are no baselines to the contrary. I have met many developers that should never have been there professionally. Not everybody is capable of actual engineering though, as engineers follow processes and measure things. So many people writing code for a living absolutely cannot measure things.
I think your fundamental problem is loss of empathy. Perhaps you understand your own limitations but do not seem to understand other people learn differently than you do or may have wildly different experiences than you have.
I do agree that you can learn with trial and observations, but it's way faster to have someone transmit the knowledge either in person or in books. And once learned, it is learned, unless you don't use that piece of knowledge for a long time and you forget about it. And it's exponential, learning things enable you to learn more things. You don't have to restart from scratch for similar situation. Like you don't have to learn letters again to learn a new Latin languages.
> Not everybody is capable of actual engineering though, as engineers follow processes and measure things
They are capable of it. It's just that most people don't care about that if there is another way to get what they need (including getting other people to do it)
This entirely depends on the teacher or the book author.
The "do not try anything new, but repeat after me" is a popular method of teaching things when we talk about skills, but often the next step is not just pursued but entirely discouraged.
There is a particular point where a teacher says "some of you might fail, but it is a sacrifice I am willing to make" & ask people to do something entirely new by themselves.
That is cross-over point where being self-taught has some advantages, because you have external guidance but not instruction.
If you're the kind of person who has failed a lot learning basics, then those new failures are less of a knock on your ego than if you had coasted through the early years without learning the parts which come after a failure.
Maybe if we all had a clear memory of "learning to walk", it would help - six+ months of falling down and not giving up, which most of us have done once.
> Not everybody is capable of actual engineering though, as engineers follow processes and measure things
So how is it that anyone can self learn anything but anyone can't follow processes and measure things? You just have different pet dogmas than the other guy. At least he didn't say you had lack of empathy to feel better.
It will be borderline impossible for anyone to learn to read/write without either learning it at school or being taught by someone.
Likewise for skills for which the information isn't easily available e.g. how to build a fission reactor.
The mostly unsolved problems right now are around build, deployment, and operations.
Code is the spec, and it's far faster to write the code yourself (if you know what you want to do and how to write code) than try to pass all this context to an LLM in the form of lossy and imprecise human language.
Your dataset is all of the 6502 binaries, corresponding manuals, and descriptions of gameplay generated by multimodal AI looking at videos, screenshots, etc. Maybe even screenshots and memory captures of games could be input along with sequences of key or controller inputs.
But if you feed in a large chunk of the Z80 and 6502 programs and as much metadata as you can and for a training signal make sure that the thing runs without freezing on an emulator and maybe shows the correct loading screen or something.. I think there are enough 8-bit programs and games for this to work, at least to some degree. Maybe. Start with shorter simpler programs and work up. Generating BASIC might be easier in a way, I don't know.
But the idea is you would describe your game along with maybe a screenshot or cover art or something and it would generate the machine code.
You could also do something similar but on a frame-by-frame input-by-input basis like Oasis.
Anyone in this thread want to fund that? I would love to work on it.
I listened to this week’s All In podcast and the podcasters (all wildly successful tech entrepreneurs) were quite emphatic that they were seeking a world where they were not beholden to developers. They envision a world where non- or semi-technical employees create the desired software via AI; other AI agents would then figure out how to deploy it.
I feel like we are at a similar point to many other professions in the past facing a rapid technological change. But in the case of software development, developers themselves are hastening their own demise.
Just something to think on.
If they use AI to build an app by just telling it what to do then everyone on the planet can look at how their app works and describe it to their AI thus making a clone.
Even Eric Schmidt said at Standford that we would be able use AI to copy TikTok and steal all of it's content and users in the next few years. Interestingly though he didn't even think to mention Youtube...
Okay, so, the thing is, roughly every decade since 1960, someone has come out with something that promises “use this, and you won’t need those icky programmers anymore”. COBOL was originally marketed this way! And then you had various flavours of implausibly high level language, that Sun thing where you were meant to make software by dragging and dropping flowcharts (quite why that was supposed to be _easier_ than programming was unclear), Macromedia Flash (while this eventually just became a platform that people wrote programs for, it was originally touted as being a bit like the Sun thing, with Actionscript as a mere escape hatch), the recent ‘nocode’ trend, a few similar historical ‘nocode’ trends, and so forth.
The only thing that has come _close_ to doing this is the spreadsheet, though that kind of happened by accident, as people took what was originally supposed to be a simple tool and created nightmares (and this sort of aggressive use leaves you needing icky Excel experts, which may not be an improvement).
Like, I’ve never heard of that podcast, but on Googling it I note that none of the people involved could reasonably be called subject matter experts (also, argh, Jason Calacanis appears to still be a thing). I’m not sure why you’d take their word on it.