You're only allowed to believe LLMs are a kind of digital messiah or that they're complete hype garbage, when of course the answer is some point between.
There is obviously a lot of potential here, but there is also a lot of solution looking for a problem.
My current red flag is if an argument hinges on a trademark breathless frisson for the growth potential. Statements like "models are getting smarter every month" that hasn't been true for a year. If your excitement over AI is based not on what we can do today, but a presumed future expansion for which we have no evidence, that's silly.
But what they can do today is cool. We worked for a long time to get computers to understand natural language intent, and LLMs demonstrably solve this problem.
It's going to take much longer and more human(?) effort to apply DL/LLMs in worthy applications by adding constraints and human-injected workarounds to make them work with more useful and fewer unpleasant surprises.
This has happened before. From 2000-2004, this was the world. Companies were going out of business or cutting headcount to the bone, and the folks with 7+ years of experience were getting picked up (often at a discount). If you started your first job coding in 1999 and then got laid off half a year later, sucks for you.
2004 both Salesforce, Google, and Blackboard (they were big then!) IPO, and Facebook comes screaming onto the scene. Greenspan monetary policy had already made capital nearly free, and the 2008 financial crisis kicked us into 0% interest rate territory. It costs us nothing to invest in talent, so why not? If we invest in 100 startups, each with 100 employees at $150K salaries, and just two of those unicorn exit, we've made our money back, and it costs financiers nothing to wait.
2004 kicked off the simultaneous rise of "software as a service" and "social media," both of which were highly lucrative. But not only that - SaaS allowed traditional (think General Mills or Procter and Gamble) to have high-quality, cutting-edge software products without needing to employ a lot of engineers to run them. They could just pay a line item to Salesforce and let them concentrate the devs.
Just like in 2004, I think we will have a major industry shift that unlocks jobs for lots of these junior folks. I don't think it will be AI - just feels too obvious. I suspect it'll be something to do with climate change.
It’s a fight to the bottom until we see the gaps. Senior associates think they can replace junior associates, managers think they can replace seniors, partners think they can replace managers and so on. My expectation, however inconvenient for me, is that juniors will be safe, mid levels will suffer as their subject matter will rebalance, seniors will collect the profit. Juniors will suffer longer term on a path with more selective progression.
As a senior I'm really not feeling any FOMO here. If the tools are going to get better and better, is there any need to be an early adopter, or build expertise in using them?
And who (without equity or significant direct incentives), in their right mind, would work on, invent, or deploy an automation tool that would make themselves and/or millions of others redundant? Some may say "productivity" assistant, but then it's not much of a hop to layoffs and reduced salaries.
The format of the article is creative and I don’t like it. The use of movies seems to be trying to channel a16z with movies instead of hip hop quotes and it isn’t working.
I stopped reading midway through because I realized I was more focused on (unsuccessfully) trying to figure out how the movies connected with the sections following them than on the actual article.
That's what's great about a16z's blog and what's bad about these movies. They don't connect in a fun or thoughtful way, that I can decipher. If they did, I might like it.
For me, the sweet spot is simple but time consuming tasks where the execution is very clear, or the result is very clear but I have to first read a ton of d3.js docs before implementing it. This stuff LLMs do faster than me typing it. Anything more involved where I myself don't know the final result yet, it's faster to figure out the problem while coding instead of trying to figure it out by describing to an LLM.
LLMs can often stitch together information from blogs and stack overflow like sites. You can view it as a better way to search and find answers to your development roadblocks, instead of something that writes code for you. This is where I get value worth paying for, by saving me time of having to navigate multiple search results and piece things together. Many times it is a mix of the two and the LLM, as a starting point, gives me better leads
The examples about law offices, and writers, is not wrong. This is happening. The article might be too long, or you might not like the style, but the concepts are correct.
While this somewhat seems like pitch for 'Cody', I don't think it is, it's just what the author is familiar with from his own company. Doesn't hide that it is a wrapper on other tools.
"Cody Pro lets you use both GPT-4o and Claude (and others), so you can spot-check all your work with another LLM."
I like some of his earlier writing but this is a straight up sales pitch. From the clickbait title sowing fear to the fact that he describes a system of using these models that just happens to coincide with what his tool does. And he's also such a nice guy he's going to start educating people on how to follow his system.
Just another person trying to cash in on the LLM hype. And he's doing so at the expense of aspiring developers that don't know enough to spot the BS.
All this talk of replacing junior engineers makes me think:
Aren’t interns and junior engineers paid on their potential rather than the work they produce?
Doesn’t it take a while (like 6 months - 1 year) for any associate to start contributing significantly to the organisation’s output?
A lawyer might delegate the same task to ChatGPT rather than a junior but is the purpose of delegation primary to get the task completed or to give an opportunity for junior to learn the specifics of the task?
Well, it’s difficult to study books to become senior developer. You have to participate in, experience the problems and solutions, at the end it’s still people that are working with you. Ai assistant is another tool in the box.
Imagine the engineer going oncall for the code this mediocrity generation tool produced. Are they gonna reason with chatgpt “why the duck did you write it this way?” or are they gonna zoom the person whose name git shows on commit during the incident?
The tool is getting better every day, although it will not replace the operator.
Wow I have never read something that I feel and live the total opposite of quite like this piece. To address the title first: My company only hires juniors, straight out of high school, sometimes HS juniors (3rd year). I myself was hired out of high school as an intern and now I am part owner (1/3rd of the board). It has gone fabulously for us and I wrote about this and why companies should consider it: https://simonsarris.com/p/growing-software-developers
As for the rest...
> You remember that big manufactured drama, right, about the new OpenAI "4o" GPT model supposedly having Scarlett Johansson's voice? That's the one. That model changed everything about programming overnight.
> ...
> You can quibble over the numbers, but it's clear that programmers using CHOP with these new models are getting a turbo boost.
I just don't think this is true. I think its a bit hyperbolic at this rate. And I don't think other people think its true which casts some doubt on part of this narrative.
> Things are changing fast. I'm an optimist, and I generally think, or at least hope, that as companies become more productive in the coming months and years, they simply get correspondingly more ambitious.
Again I really don't think we've seen any real productivity boost yet to companies. Maybe a few individuals, but not companies. And I am suspicious that we will, given the nature of things.
> All I can tell you is this: Get there early.
I actually think the opposite! You can safely ignore most of the AI hype cycle and tools right now. You don't have to play with them that much. You can wait a few more years and when they are more polished, use them more if you like them. There's really not a big rush.
I take umbrage with the style of this article, trying to convince of its point by using conspiratorial tones while gesturing towards movies. Maybe its amusing to some, but it feels like a crunch in its convincing-ness of the central arguments.
Yup yup, in the words of Olivia Rodrigo, it's brutal out here. I've found it nigh impossible to sell my programming skills as a web dev. Instead I've found freelance work doing technical writing and making Urbit apps on the side. Wish me luck. Anyone wants to code together to train for that imaginary dream junior dev job in the sky, hmu! Always happy to learn with friends :)
Companies are still looking for junior developers (in the US). If you're dead set on being a developer then go out and build things that you want to build. It will hopefully be an enjoyable process and you'll continue to improve along the way.
If that sounds like a terrible idea then consider switching to a different field, the trades are booming right now.
Steve is right. Things are definitely changing in the space thanks to LLMs. Question is how quickly and in what way. I tend to think the junior devs are the ones embracing the LLMs.
As a side note: I use Cody and the Sourcegraph core search product at work. I recommend both (I particularly like Cody’s context building using repos). I also think they are in a competitive hard spot, because their own tools will get commoditized by the big bruisers in the space (GitHub, Gitlab, and Codewhisperer).
basically this is a lot of seemingly AI-generated drivel directly related to driving you to Cody, whatever their thing they are trying to sell with this click baity headline is doing.
I could not find anything of substance in this article, and was disappointed, because I feel this is a topic in dire need of discussion.
I agree. My takeaway from this article is that I need to sign up for Anthropic so that I can bounce ideas off of GPT-4o. Ultimately, I’m seeing diminishing returns in the output of LLMs over the last 18 months.
Probably the biggest red flag to me is anything that requires me to rapidly skim to get to their point. Good writing usually establishes their point early.
Five years later: Why can’t I find any senior developers? Shocked pikachu face.
There is obviously a lot of potential here, but there is also a lot of solution looking for a problem.
My current red flag is if an argument hinges on a trademark breathless frisson for the growth potential. Statements like "models are getting smarter every month" that hasn't been true for a year. If your excitement over AI is based not on what we can do today, but a presumed future expansion for which we have no evidence, that's silly.
But what they can do today is cool. We worked for a long time to get computers to understand natural language intent, and LLMs demonstrably solve this problem.
2004 both Salesforce, Google, and Blackboard (they were big then!) IPO, and Facebook comes screaming onto the scene. Greenspan monetary policy had already made capital nearly free, and the 2008 financial crisis kicked us into 0% interest rate territory. It costs us nothing to invest in talent, so why not? If we invest in 100 startups, each with 100 employees at $150K salaries, and just two of those unicorn exit, we've made our money back, and it costs financiers nothing to wait.
2004 kicked off the simultaneous rise of "software as a service" and "social media," both of which were highly lucrative. But not only that - SaaS allowed traditional (think General Mills or Procter and Gamble) to have high-quality, cutting-edge software products without needing to employ a lot of engineers to run them. They could just pay a line item to Salesforce and let them concentrate the devs.
Just like in 2004, I think we will have a major industry shift that unlocks jobs for lots of these junior folks. I don't think it will be AI - just feels too obvious. I suspect it'll be something to do with climate change.
But of course there are newer things to build, buy that is largely affected by 0% interest rate.
I have a genuine query. Are any software engineers getting sane code out of LLM ?
I struggle to conjure good unity or kotlin code from both paid and self hosted solutions.
For me, the sweet spot is simple but time consuming tasks where the execution is very clear, or the result is very clear but I have to first read a ton of d3.js docs before implementing it. This stuff LLMs do faster than me typing it. Anything more involved where I myself don't know the final result yet, it's faster to figure out the problem while coding instead of trying to figure it out by describing to an LLM.
[0] https://wakatime.com/blog/58-chatgpt-prototyped-our-new-feat...
I could never imagine copy-pasting back and forth with an AI in a chat window. At that point I'd just Google the docs and write the code myself.
While this somewhat seems like pitch for 'Cody', I don't think it is, it's just what the author is familiar with from his own company. Doesn't hide that it is a wrapper on other tools.
"Cody Pro lets you use both GPT-4o and Claude (and others), so you can spot-check all your work with another LLM."
I like some of his earlier writing but this is a straight up sales pitch. From the clickbait title sowing fear to the fact that he describes a system of using these models that just happens to coincide with what his tool does. And he's also such a nice guy he's going to start educating people on how to follow his system.
Just another person trying to cash in on the LLM hype. And he's doing so at the expense of aspiring developers that don't know enough to spot the BS.
Pretty appalling behavior.
Aren’t interns and junior engineers paid on their potential rather than the work they produce?
Doesn’t it take a while (like 6 months - 1 year) for any associate to start contributing significantly to the organisation’s output?
A lawyer might delegate the same task to ChatGPT rather than a junior but is the purpose of delegation primary to get the task completed or to give an opportunity for junior to learn the specifics of the task?
Imagine the engineer going oncall for the code this mediocrity generation tool produced. Are they gonna reason with chatgpt “why the duck did you write it this way?” or are they gonna zoom the person whose name git shows on commit during the incident?
The tool is getting better every day, although it will not replace the operator.
As for the rest...
> You remember that big manufactured drama, right, about the new OpenAI "4o" GPT model supposedly having Scarlett Johansson's voice? That's the one. That model changed everything about programming overnight.
> ...
> You can quibble over the numbers, but it's clear that programmers using CHOP with these new models are getting a turbo boost.
I just don't think this is true. I think its a bit hyperbolic at this rate. And I don't think other people think its true which casts some doubt on part of this narrative.
> Things are changing fast. I'm an optimist, and I generally think, or at least hope, that as companies become more productive in the coming months and years, they simply get correspondingly more ambitious.
Again I really don't think we've seen any real productivity boost yet to companies. Maybe a few individuals, but not companies. And I am suspicious that we will, given the nature of things.
> All I can tell you is this: Get there early.
I actually think the opposite! You can safely ignore most of the AI hype cycle and tools right now. You don't have to play with them that much. You can wait a few more years and when they are more polished, use them more if you like them. There's really not a big rush.
I take umbrage with the style of this article, trying to convince of its point by using conspiratorial tones while gesturing towards movies. Maybe its amusing to some, but it feels like a crunch in its convincing-ness of the central arguments.
Which makes this seem like a cynical, if long and winding, sales pitch.
If that sounds like a terrible idea then consider switching to a different field, the trades are booming right now.
But... that's what code review is for. Has the author of this article never heard of the concept?
As a side note: I use Cody and the Sourcegraph core search product at work. I recommend both (I particularly like Cody’s context building using repos). I also think they are in a competitive hard spot, because their own tools will get commoditized by the big bruisers in the space (GitHub, Gitlab, and Codewhisperer).
I could not find anything of substance in this article, and was disappointed, because I feel this is a topic in dire need of discussion.
I'm biased in his favor, but I recognize that brevity is not among his virtues.