#AI spend

1 messages · Page 1 of 1 (latest)

pastel magnet
#

@unkempt charm

One of the things I read into Adafruit's recent LLM streaming stuff is that hiring of new CircuitPython devs is much less likely than increased LLM spend.
I have a lot of my own qualms about LLM's, which I won't bring up here. But I'll say that our spend is still way less than the cost of a developer. I don't see it influencing the hiring of humans at Adafruit. And I see a lot of current layoffs being justified as "for AI", when in fact that's a convenient excuse to cut staff. The productivity gains have yet to be demonstrated for most non-programming tasks.

unkempt charm
#

Part of my take on this is formed by having spent some time working on recent projects with Claude's Haiku model on the free tier. Feels like on the order of 10x or 100x productivity boost compared to the time it would take me to research the volume of docs it plows through before I wipe out my token quota. To me it feels like people who aren't using AI are gonna get left in the dust. Like, just totally smoked.

#

Also, it will be really hard for junior devs to anticipate where the AI is likely to go off the rails. It's gonna be hard for new folks to see the efficiency gains (in the context of working reliable code) that will be possible for more senior devs coding with AI assistance.

blissful slate
#

I agree with Dan. The hiring slowing is companies using AI as an excuse. The reality is that the economy isn't great due to the uncertainty of US policy. For Adafruit, the struggles with covid, chip shortage and tariffs have had a much bigger impact than AI.

blissful slate
#

re: Junior devs. I think what that means will change. When I graduated in 2009, out of school most of my peers had gotten proficient with typing, learned a few language syntax, seen data structure stuff and had started to learn to read code. They'd learn a bit of theory but had little experience reading code and understanding large systems. I was lucky enough to start coding before university and have projects of my own to practice my typing and syntax knowledge. That got me a head start into the higher level stuff.

With LLMs, typing proficiency and knowing syntax matters much less now because LLMs can do that for you. Now folks learning to code can jump into learning systems engineering and testing. Generating coding is easy. Knowing it's the code you wanted is hard and knowing exactly what you want is hard too. Those things were previously gated on prerequisite skills (typing and syntax knowledge).

unborn mountain
# unkempt charm Part of my take on this is formed by having spent some time working on recent pr...

But have you actually measured the productivity impact? From what I've seen of studies on LLM productivity impact on devs, anyone who's not measuring it is likely to be substantially overestimating the actual benefit.

A recurring theme in studies about LLM usage in development is perception gaps. Devs feel like the LLM is letting them push out better work faster, but when someone's actually timing the work or evaluating its quality, the measured productivity and performance turns out to be significantly worse than the dev using the LLM thought.

The literature's far from complete enough to be conclusive, but it's enough to warn us to be careful about measuring and evaluating the results, rather than taking our gut instinct at face value. Especially if you're going to be basing major business or career decisions on the results.

blissful slate
#

I haven't measured mine. It is definitely shifting what I work on though. Less time typing, more time reading, reviewing and testing. How much time I spend on the latter depends on what I'm going to use it for. A one off helper script doesn't need review. Something I'm going to PR I want to stand by and not waste time so it requires my own review.

I'm also doing more automated testing than I would have before because it was time intensive to setup and maintain. Now, the LLM can generate it fast and use it as a tool for itself.

unkempt charm
#

The difference in quality of results achievable by people with limited experience in a domain vs. results achievable by people with significant experience is part of what concerns me about AI adoption. There's potential for significant disruption to the pipelines of mentoring that previously allowed junior devs to gain the domain experience that senior devs acquired over many years of working with their seniors.

blissful slate
#

I think junior devs will be able to gain this experience much faster now because they can create code to experiment with much faster.

#

@orchid vector You disagree with me?

pastel magnet
#

do we want to move this discussion into the new channel?

blissful slate