piggy [they/them]

I know an artist that got super rich off of NFTs, she didn’t own any or had anything to do with the crypto side she just made the “apes” though I think hers were mostly fairies. She’s very good at the whole “industrial artist” gig. NFTs honestly seemed like a gold rush for people with the ability to navigate that space. She cleared half a million one year.

There’s a tradition of the Soviet intelligensia being temporarily embarassed nobles, see big fans of Bulgakov.

His only mistake is that the gold bars were stamped “Made in Egypt”. Just like TikTok if they were of American provenance you get a pass.

What you’re afraid of is precisely what was tried with outsourcing dev jobs. That proved to work in some areas where you have very boring crud apps, but was a complete failure in others. I expect LLMs are just going to work out in a very similar fashion.

Okay but like again, I’m not afraid of losing my job. I’m afraid that we’re going to lose real capability as a society. It’s how our oligarchs are practically morons compared to past oligarchs who built hundreds of libraries, or how we don’t have the real capacity in the US to build rail.

I’m currently working as a platform architect coordinating 5 teams over multiple products building a platform for authoring, publishing and managing rich educational courses across multiple different grade levels. I do most of the greenfield development still, I personally manage a DSL and tools for it, while figuring out platform requirements and timelines for other teams including my own. I used to work on a real time EEG system doing architecture and signal processing. I’ve architected and implemented medical logistics platforms. I’ve been a first engineer at a couple of startups. I’ve literally written purpose built ORMs, schedulers, middleware frameworks, and query frameworks from scratch. I’ve worked at almost every major common role at a principal level except security (which is mostly fake) and embedded so front end, back end, database optimization/integration, infrastructure, machine code on JVM and X86, and distributed computing. I haven’t work in niches like networking, industrial, ML or quantum, I’d only really want to explore quantum or networking in reality. But quantum is something you typically need PhDs for otherwise it’s gonna be a bit grunty. OSS may bring up engineers for some of these roles, but in practice the majority of OSS projects don’t reach the level of complexity that I’ve worked at – the ones that do aren’t community projects they’re corporate ones.

Very few people can step into my shoes, most principal engineers I’ve met average out at a large project where they implemented a strangler once or twice. The system currently has a hard time reproducing me, if the bottom falls out it’s gonna be good game. I’m happy that LLMs are helping you rediscover your passion, but the kind of stuff you’re talking about are toys. Personally they’re not fun, they’re mostly boring, I enjoy building large technical systems in complex problem spaces in a high level reproducible way. Everything else gets stale quickly. I’ve built out systems where if you blow on the code the tests turn red without test maintenance and creation being a burden. The goal was high value test in 5 minutes in that system. The future I see is that everything is just shittier because the skill that is hard to find and is dying is understanding the essential complexity at the 10,000 ft view, the 100 ft view, the 1 ft view, and the 1 micrometer view. I can barely find developers who can innately understand essential complexity at one of those view points. I’ve met about 20 who can do all 4 and I’ve met maybe like 400-ish devs in my life.

The only passion project I wanted to start I basically decided to call off because if successful it would be bad for the world. I wanted to build a high level persona management software that could build swarms in the tens of thousands without being discovered.

If LLM removes programming as a job, might be nice, but in practice it’s just gonna mean more people on the struggle bus.

And as I already pointed out above, the problem here isn’t with automation but with capitalism. In a sane system, automation would mean more free time for people, and less tedium. People are doing these jobs not because they want to be doing them, but because it’s a way to survive in this shitty system.

There are certainly bad programming jobs, but programming jobs in general are extreme labor aristocracy. Yes people are chasing the bag, but they’re certainly not “survival jobs”. Within the system until you reach senior levels is no real discriminator between “bag chaser” and “person who is trying to learn”, both these are going to get squad wiped.

There’s certainly still going to be a path to being a SE. But it’s going to be autodidact hobbyists who start extremely young. As a person who has been running Linux since 5th grade, who got a CCNA at 16, who has only had programming or network jobs since high school, this is the worst path because the reality of the career at scale murders your passion. If I don’t age out I’m betting my next 10 years are going to be uncomfortably close to Player Piano, and that’s something that’s entirely dreadful. Instead of teaching juniors to program at scale while giving them boring CRUD tasks, I’ll be communing with machine spirits so “they” can generate the basic crud endpoints and the component screens.

The reality of being a greybeard is that if you’re close to retirement in this industry like my dad is, you’re gonna do the same shit jobs as the bag chasers. They’ll stick you in the basement and steal your stapler if you even make it past the vibe check interview. The only way to avoid this is to be a lifer somewhere, but that in itself is a challenge.

The difference between the previous developments and now, is that it may improve productivity now in your case and the case of the 1000 juniors, but tomorrow it’s going to actually undercut demand for people. Building a system that builds and deploys applications has been the goal of several public and private projects I’ve been privy to. I agree this exact use-case that you linked is an example of a way to not have to learn ANTLR or how an AST works and flip a coin if it works. In practice though, this is step 1. Code generation has improved significantly in the last year alone across the whole LLM ecosystem. The goal isn’t’ to write maintainable code or readable code, the goal is to write deploy-able code with 90% feature coverage. Filling the last 10% with freelancers or in house engs. depending on scale. To me that’s a worse job than the job I have now, at least now I can teach others how to do what I do. If that’s taken away from me I’m not fucking doing this job anymore. I don’t care about computers because in reality this job at scale is about convincing morons to stop micromanaging how you build things.

Can’t wait for RFK Jr. to let the free market decide how much listeria people choose to ingest.

I’m so excited for this

Yes and?

1. They’re getting paid.
2. It’s a job.
3. They’re humans who can choose to be better.
4. They’re humans who can choose to fight their bosses out of some idiotic love of the game to the detriment of their own mental health because they’re crazy. (I’m describing myself).
5. They’re humans who can stall or break awful things from coming to pass by refusing to work on something or sabotaging it.

This is about a door to those possibilities closing, not about how many software developers are forced through it. I’m not going to cheer on an awful totalizing future dark age of technology simply because the current odds are bad.

And yeah this won’t actually kill higher end devs in my understanding of the world, I’ll be able to find a job. But, it will kill the social reproduction of people like me. In the same way that the iPad killed broad user-focused technological literacy from zoomers to millenials, LLMs will ultimately destroy the current level of developer-focused technological literacy. There won’t even be guys who can’t code their way out of a paper bag using StackOverflow or guys who memorize LeetCode solutions. It will just be old-heads powerful enough to avoid the cull and nobody else, until we die.

Every large coporation uses this method because they want to have fungible devs. Since developers with actual skill don’t want to be treated as fungible cogs, the selection pressures ensure that people who can’t get jobs with better conditions end up working in these places. They’re just doing it to get a paycheck, and they basically bang their heads against the keyboard till something resembling working code falls out. I’ll also remind you of the whole outsourcing craze which was basically exact same goal corps want to accomplish with AI now.

Damn that’s crazy, imagine working a coding job for a paycheck! Soon you won’t even be able to!

Also, for devices like airconditioners or televisions that use IR remotes and presumably some subset of a standard code is there a structured way to build an interface for them?

You can DIY with https://tasmota.github.io/docs/Tasmota-IR/

Some people recommend Broadlink RM3/RM4 and just walling it off from the web.

If your PC/Raspberry PI is in range you can get a cheap linux compatible IR Emitter and use LIRC https://www.lirc.org/

StackOverflow copypasta wasn’t a productive processes that was seeking to remove the developer from the equation though.

This isn’t about a tech scaling strategy of training high quality high productivity engineers vs “just throwing bodies at it” anymore. This is about the next level of “just throwing bodies at it”, “just throwing compute at it”.

This is something technically feasible within the next decade unless, inshallah, these models collapse from ingesting their own awful data, rather than improving.

For everyone of you there’s 1000 junior engineers running copilot.

The way that you’re applying the tool “properly” is ultimately the same way that middle managers want to apply the tool, the only difference is that you know what you’re doing as a quality filter, where the code goes and how to run it. AI can’t solve the former (quality) but there are people working on a wholesale solution for the latter two. And they’re getting their data from people like you!

In terms a productive process there’s not as much daylight between the two use cases as you seem to think there is.

Nobody is arguing for using the AI for problems you keep mentioning, and you keep ignoring that.

This is absolutely not true. Almost every programmer I know has had their company try to “AI” their documentation or “AI” some process only to fail spectacularly because the basis of what the AI does to data is either missing or doesn’t have enough quality. I have several friends at the Lead/EM level take too much time out of their schedules to talk down a middle manager from sapping resources into AI boondoggles.

I’ve had to talk people off of this ledge, and lead that works under me (I’m technically a platform architect across 5 platform teams) actually decided to try it anyway and burn a couple days on a test run and guess what the results were garbage.

Beyond that the problem is that AI is a useful tool in IGNORING the problems.

I’ve given you concrete examples of how this tool is useful for me, you’ve just ignored that and continued arguing about the straw man you want to argue about.

I started this entire comment thread with an actual critique, a point, that you have in very debate bro fashion have consistently called a strawman. If I were a feeling less charitable I could call the majority of your arguments non-sequitors to mine. I have never argued that AI isn’t useful to somebody. In fact I’m arguing that it’s dangerously useful for decision makers in the software industry based on how they WANT to make software.

If a piece of software is a car, and a middle manager wants that car to have a wonderful proprietary light bar on it and wants to use AI to build such a light bar on his wonderful car. The AI might actually build the light bar in a narrow sense to the basic specs the decision maker feels might sell well on the market. However the light bar adds 500lbs of weight so when the driver gets in the car the front suspension is on the floor, and the wiring loom is also now a ball of yarn. But the car ends up being just shitty enough to sell, and that’s the important thing.

And remember the AI doesn’t complain about resources or order of operations when you ask it do make a light bar at the same time as a cool roof rack, a kick ass sound system and a more powerful engine, and hey if the car doesn’t work after one of these we can just ask it to regenerate the car design and then just have another AI test it! And you know what it might even be fine to have 1 or 2 nerds around just in case we have to painfully take the car apart only to discover we’re overloading the alternator from both ends.

I’ve never said that AI is the cause of those problems that’s words you’re putting in my mouth. I’ve said that AI is being used as a solution to those problems in the industry when in reality the use of AI to solve those problems exacerbates them while allowing companies to reap “productive” output.

For some reason programmers can understand “AI Slop” but if the AI is generating code instead of stories, images, audio and video it’s no longer “AI Slop” because we’re exalted in our communion with the machine spirits! Our holy logical languages could never encode the heresy of slop!

This is a quantization function. It’s a fairly “math brained” name I agree, but the function is called qX_K_q8_K because it quantizes a value with a quantization index of X (unknown) to one with a quantization index of 8 (bits) which correlates to the memory usage. The 0 vs K portions are how it does rounding, 0 means it does rounding by equal distribution (without offset), and K means it creates a distribution that is more fine grained around more common values and is more rough around least common values. e.g. I have a data set that has a lot of values between 4 and 5 but not a lot of 10s. I have lets say 10 brackets between 4 and 5 but only 3 between 5 and 10.

Basically it’s a lossy compression for a data set into a specific enumeration (roughly correlates with size), so it’s a way to given 1,000,000 numbers from 1-1000000, of putting their values into a range of numbers based on the q level How using different functions affects the output of models is more voodoo than anything else. You get better “quality” output from higher memory space, but quality is a complex metric and doesn’t necessarily map to factual accuracy in the output, just statistical correlation with the model’s data set.

An example of a common quantizer is an analog to digital converter. It must take continuous values from a wave that goes 0 to 1 and transform them into digital values of 0 and 1 with a specific sample rate.

Taking a 32 bit float and copying the value into 32 bit float is an identity quantizer.

You’re making up a giant straw man of how you pretend software development works which is utterly divorced from what we see happening in the real world. The AI doesn’t change this one bit.

Commenting this under a post where an AI has spit out a dot product function optimization for an existing dot product function that’s already ~150-250 lines long depending on architectural implementation of which there are about 6. The PR for which has an interaction that is two devs finger pointing about who is responsible for writing tests. The PR for which notes that the original and new function often don’t give the correct answer. Just an amazing response. Chefs kiss.

What a wonderful way to engage with my post. You win bud. You’re the smartest. This industry would never mystify a basic concept that’s about 250 years old with a 716 line PR through its inability to communicate, organize and follow an academic discipline.

I know what I want to do conceptually, and I have plenty of experience designing applications.

How does AI help you actually traverse the concepts of React that you admit you don’t have nitty gritty knowledge of how they work in terms of designing your application? React is a batteries included framework that has specific ways of doing things that impact the design and concepts that are technically feasible within React itself.

For example React isn’t really optimized to crunch a ton of data performantly so if you’re getting constant data updates over a web socket from multiple points and you want some or all the changes to be reflected you’re gonna have a bad time vs something that has finer grained change controls out of the box such as Angular.

How does AI help you choose between functional and class based React components? How much of your application is doing typical developer copy-pasta instead of creating HOCs for similar functionalities? How did AI help you with that? How is AI helping apply concepts like SOLID into the design of your component tree? How does AI help you decide how to architect components and their children that need to have a lifecycle outside of the typical change-binding flow?

This in my opinion is the crux of the issue, AI cannot solve this problem for you nor can it reasonably explain it in a technical way beyond parroting the vagaries of what I said above. It cannot confer understanding of complex abstract concepts that are fuzzy and have grey areas. It can tell you something may not work explicitly but it cannot educate you realistically on the tradeoffs.

It seems to me that your answer boils down to “code monkey stuff”. AI might help you swing a pickaxe, but it’s not good at explaining where the mine is going to collapse based on the type of rock you’re digging in. Another way of thinking about it is that you could build a building to the “building code” but it will still collapse. AI can explain the building code and loosely verify that you built something to it, but it cannot validate that your building is going to stay standing nor can it practically tell you what you need to change.

My problem with AI tools boils down to this. Software is a medium of communication. It communicates the base of a problem and the technical process of solving it. Software Engineering is a field that attempts to create strong patterns of communication and practices in order to efficiently organize the production of Software. The software industry at large (where most programmers get exposed to the process of building software) often eschews this discipline because of scientific management (the idea you can simply manage a process through fiduciary/managerial knowledge rather than domain knowledge) and the need for instant development to maintain fictional competitive advantage and fictional YoY growth. The industry welcomes AI for 2 reasons:

1. It can code monkey…eventually. Why pay programmers when you can ask CahpGBT to do it?
2. It can fix the problem of needing to deliver without knowing what you’re doing… eventually. It fixes the problem of communication without relying on building up the knowledge and practice of Software Engineering. In essence why have people know this discipline and its practical application when you can continue to have the blind leading the blind because ChadGTP can see for us?

This is a disservice to programmers everywhere especially younger ones because it destroys the social reproduction of the capacity to build scalable software and replaces it with you guessed it machine rites. In practice it’s the apotheosis of Conway’s Law in the software industry. We build needlessly complex software that works coincidentally, and soon that software will be analyzed, modified, and ultimately created by a tool that is an overly complex statistical model that also works through the coincidence of statistical approximations.

Okay let me ask this question:

Who is this useful for? Who is the target audience for this?

That’s just a straw man, because there’s no reason why you wouldn’t be looking through your code. What LLM does is help you find areas of the code that are worth looking at.

It’s not a strawman because classifying unperformant code is a different task than generating performant replacement code. LLM can only generate code via it’s internal weights + input it doesn’t guarantee that that code is compilable, performant, readable, understandable, self documenting or much of anything.

The performance gain here is coincidental simply because the generated code uses functions that call processor features directly rather than get optimized into processor features by a compiler. LLM classifiers are also statistically analyzing the AST for performance they aren’t actually performing real static analysis of the AST or it’s compiled version. It doesn’t calculate a BigO or really know how to reason through this problem, it’s just primed that when you write the for loop to sum, that’s “slower” than using _mm_add_ps. It doesn’t even know which cases of the for loop compile down to a _mm_add_ps instruction on which compilers and which optimization levels.

Lastly you injected this line of reasoning when you basically said “why would I do this boring stuff as a programmer when I can get the LLM to do it”. It’s nice that there’s a tool that you can politely ask to parse your garbage and replace with other garbage that happens to use a function that’s more performant. But not only is this not Software Engineering, but a performant dot product is a solved problem at EVERY level of abstraction. This programming equivalent of tech bros reinventing the train every 5 years.

The fact that this is needed is a problem in and of itself with how people are building this software. This is machine spirit communion with technojargon. Instead of learning how to vectorize algorithms you’re feeding your garbage code through a LLM to produce garbage code with SIMD instructions in it. That is quite literally stunting your growth as a Software Engineer. You are choosing to ignore learning how things actually work because it’s too hard to parse through the existing garbage. A SIMD dot product algo is literally a 2 week college junior homework assignment.

Understanding what good uses for it are and the limitations of the tech is far more productive than simply rejecting it entirely.

I quite literally pointed several limitations in the post you replied to and in this post from a Software Engineering perspective.