I'm a big fan of this explanation/rant from Andrew Murphy.
-
@diekehrseite Well, @ulveon doesn't say it explicitly, but this case *was* an interesting example of where we could no longer say the LLM "just generating code."
The fact that it can succeed at that level of sophisticated analysis suggests that when we have clear success criteria (e.g. "vuln found"), the LLM can do very hard things indeed.
Agree this will be really interesting to watch!
-
The fact that we are *not* seeing wildly improving software all around us tells us everything we need to know.
There is no flourishing of value delivery, new product categories, more needs being satisfied better. It’s the opposite.
All we are seeing is decreases in quality, because
code creation is not the problem.It started with visual entertainment, say movies - basically no-one makes new films any more, because the VC suits won't fund it. So everyone does remakes of "shit that worked last time". Now, the same thing is happening to software.
Counter-examples of course exist in the FOSS world.
-
The fact that we are *not* seeing wildly improving software all around us tells us everything we need to know.
There is no flourishing of value delivery, new product categories, more needs being satisfied better. It’s the opposite.
All we are seeing is decreases in quality, because
code creation is not the problem.@elizayer I think of LLM’s as pattern re-use; applying learned patterns from old data to new situations. Good for menial tasks that are too boring for humans, but not for coming up with radical new things.
-
The fact that we are *not* seeing wildly improving software all around us tells us everything we need to know.
There is no flourishing of value delivery, new product categories, more needs being satisfied better. It’s the opposite.
All we are seeing is decreases in quality, because
code creation is not the problem.@elizayer @hbons give me some time. I’ve only been using LLMs to code for a few months… so far I’ve only managed to write an operating system https://codeberg.org/dpp/meows
A new scripting language https://codeberg.org/dpp/meowscript
An eBPF to FPGA converter https://codeberg.org/dpp/lycaon
And some misc utils
But this is weekend work -
@cigitalgem Yeah, I also want to be honest with ourselves.
At least in the US, people change jobs so often -- and promotion practices are so shonky -- that the jr dev-architect flow was already under threat at scale
@elizayer agreed.
-
@elizayer @hbons give me some time. I’ve only been using LLMs to code for a few months… so far I’ve only managed to write an operating system https://codeberg.org/dpp/meows
A new scripting language https://codeberg.org/dpp/meowscript
An eBPF to FPGA converter https://codeberg.org/dpp/lycaon
And some misc utils
But this is weekend work -
@syntheticmind_ai @BmeBenji could you explain to me how to write code that outputs a seahorse emoji
-
@elizayer @hbons give me some time. I’ve only been using LLMs to code for a few months… so far I’ve only managed to write an operating system https://codeberg.org/dpp/meows
A new scripting language https://codeberg.org/dpp/meowscript
An eBPF to FPGA converter https://codeberg.org/dpp/lycaon
And some misc utils
But this is weekend work -
@elizayer@mastodon.social Claude Code found a 23-year-old Linux vulnerability, the kind a regular human security auditor would have taken weeks or months to find (or in this case, 23 years). https://mtlynch.io/claude-code-found-linux-vulnerability/
-
I generally agree!
On the narrow Waymo point, a few things have made me reconsider recently:
- Cyclists who feel Waymos are more predictable and less likely to make the equivalent of attentiveness mistakes. Or to be actively hostile.
- Women and older people who've said they feel vulnerable alone in a car with a driver.
@elizayer
The problem with both Uber and Waymo is that they don't solve the basic problem of cars (the bottleneck if you will): the geometry problem. Cars need too much space and make places, especially cities, less livable for humans.The answer is, and always has been, viable alternatives to driving. Namely, good public transport (with level boarding), good cycling and walking infrastructure. Automated vehicles may have their place in the world but not as mass transport.
-
@elizayer @hbons give me some time. I’ve only been using LLMs to code for a few months… so far I’ve only managed to write an operating system https://codeberg.org/dpp/meows
A new scripting language https://codeberg.org/dpp/meowscript
An eBPF to FPGA converter https://codeberg.org/dpp/lycaon
And some misc utils
But this is weekend work@dpp and somehow, all these boatloads of incredible new projects (not picking specifically at yours) managed to have approximatively 0 impact. (I’m talking about positive impact of course, the negative ones are well documented).
But as long as you have fun helping destroying society, you do you! Otherwise, maybe pick up playing ukulele or something? -
So why are we still trying to optimize code creation?
For decades, people with power - executives and product people - have been shifting the blame for strategy failures and poor market insight onto development "productivity."
This AI moment should be incredibly clarifying. Like, it should be the reductio ad absurdum of a productivity-centric approach.
@elizayer oh wow yes that is a really good point
-
@dpp and somehow, all these boatloads of incredible new projects (not picking specifically at yours) managed to have approximatively 0 impact. (I’m talking about positive impact of course, the negative ones are well documented).
But as long as you have fun helping destroying society, you do you! Otherwise, maybe pick up playing ukulele or something?@ced @elizayer @hbons we are posting via a tool that was developed by an imperial government to ensure communications could survive after a nuclear war.
We are using a toxic form of communication (oh… yeah… mastodon is toxic light) that has destroyed the trust in institutions and is the proximate cause of the rise of authoritarian regimes
Yes, I use tools that have negative externalities.
I use these tools to explore and create.
You don’t like it, ignore me.
-
@ced @elizayer @hbons we are posting via a tool that was developed by an imperial government to ensure communications could survive after a nuclear war.
We are using a toxic form of communication (oh… yeah… mastodon is toxic light) that has destroyed the trust in institutions and is the proximate cause of the rise of authoritarian regimes
Yes, I use tools that have negative externalities.
I use these tools to explore and create.
You don’t like it, ignore me.
-
-
-
I'm a big fan of this explanation/rant from Andrew Murphy.
Taken as a whole, there are many bottlenecks in a corporate software development process. The "load-bearing" calendar is a great example!
Speeding up code creation just increases pressure on the bottleneck, which decreases throughput.
@elizayer it's such a stressful read, honestly, for somebody stuck in the middle of all this!
-
@elizayer @beep I was literally just talking to someone about #Waymo for this same reason. Tech has reached the point where it has become more than abundantly obvious to anyone who dares to ask a single question that the objective is no longer the improvement of anyone’s life but the #EpsteinClass’s. Why is taking a Waymo better than taking an Uber? Because now someone’s out of a job. Why is #AI better than a software developer? Because now someone’s out of a job
-
So why are we still trying to optimize code creation?
For decades, people with power - executives and product people - have been shifting the blame for strategy failures and poor market insight onto development "productivity."
This AI moment should be incredibly clarifying. Like, it should be the reductio ad absurdum of a productivity-centric approach.
@elizayer a feature-idea stress test app would be cool.
-
@elizayer a feature-idea stress test app would be cool.
@elizayer for new biz, Ash Maurya has an interesting tool.
https://leanspark.ai/
