As in workload or “skills”?
1 Like
Workload. While new features and projects have to be done with manual/mental work, a lot of things in established projects can be done by AI if the architecture is good and the code is well structured and you give it clear instructions.
I can’t possibly say to my boss that a thing will take one hour when I can give copilot instructions and it’ll be done in 5 mins + 5 mins for me to review it. If I take one hour I will be asked why I took one hour…
Edit: And while I have experience and can review the output and guide the AI properly, a lot of juniors in many companies are producing crap code that will eventually need to be fixed.
5 Likes
This challenge has always existed. A junior developer is, by definition, not yet operating at median efficiency. And yet we know that it is worthwhile to invest in juniors, so as to elevate them above median eventually.
My greatest worry is about juniors foregoing learning through AI. I’ve seen it happen. Efficiency improves short term, but long term growth is stunted. Not a good trade for anybody.
3 Likes
Yes of course, but it’s a whole new magnitude of junior code being produced. And sometimes not even juniors, our very own designer vibe coded an app out of existence and the management finds it wonderful. I can’t wait for the eventual bugs, data leaks, etc
3 Likes
I’m not convinced of this. I’ve seen juniors blindly accept what AI tells them and then spend days trying to figure out why it doesn’t work. Often someone senior steps in to fix it but the junior doesn’t learn from it because they produced so little of the code consciously. Rinse and repeat.
5 Likes
I’ve had that happen, too. A product manager vibe coded a little helper app, couldn’t get something to work, and asked for help. I (as the software department guy) stepped in and took over, as that unholy mess would have caused no end of trouble. In the event, I entirely rewrote it, as the bramble of Python/Tcl/Tk was too tangled to rescue.
In the neighboring software team, a hardware guy had vibe coded a device driver for a €100K machine. What could possibly go wrong? Again, the software department stepped in before damage could be done.
Sooner or later we’ll be too late. It only takes one or two high-profile incidents for a small company to go under.
4 Likes
Don’t know how I missed this thread 'till now…
Danged freaking computer scientists, they keep adding layers of abstraction to their little engines…
I recall during my time as department chair at a local university, had a hard time finding someone to teach the CS digital concepts course. So, I vectored my CS majors to the EE equivalent, a bit more hardware oriented. The hue and cry was epic, but I later had folk come back to me with thanks, that course helped them better-understand their software errors. Understanding the essential dynamics of a Von Neumann computing engine is useful.
A lot of manufacturing and fabrication is being automated, CNC machines, additive printers and such. Thing is, the dynamics of the materials haven’t changed, bear understanding things like a FDM-printed article doesn’t have the same dynamic properties as an equivalent cast from the same material; one is discrete layers of it, the other is a continuous structure of the material. Can’t just program a toolpath and go to lunch…
The whole GPS/mapapp thing has shaped my geographical perspective in rather unappreciated ways. When we travel, I’ve gotten very good at twisting Google Maps to my need, but I find that doesn’t educate me at all about where I’ve been. Last year we took my wife’s aunt on a cruise/overland trip in the US Northeast; on the overland part, we navigated places in the Concord MA area I’d previously learned pretty well from my Air Force travels decades ago, but the same places felt eerily different navigating them point-to-point in Maps.
My recent LLM experience has been both good and bad in this regard. Of note, I’ve been intermittently messing with writing CAD programs, and I had a good experience with Microsoft CoPilot building an OpenGL render engine. It ‘knew’ all the tools I use, including wxWidgets (?!?), and offered both concise explanations and usable code for things like VAO/VBO management, the heart of recent OpenGL. I spent a lot of time farting around with writing a de-duplication STL file reader, got frustrated and finally said, “CoPilot, give me a de-duplication STL reader”, and it did, with all the current thinking of the game engines behind weld-transform de-duplication. In this episode, I was able to use a LLM to learn things, yay.
5 Likes
This seems to fit Dan Davies’ “trying to stuff 200 end user apps into a trenchcoat so they can pretend to be a system”.
2 Likes
One of the links, probably from this thread, has lead me to
Which was I think the most important thing I read today.
We humans can not fully understand reality. Therefore we must trust others. We have become pretty bad at selecting reliable sources. In fact, reliability itself has lost importance. But to solve this problem, we don’t need good arguments, or better sources. We need to re-establish the value of reliability.
So many important ideas to unpack in there.
1 Like
I stopped at the author’s Luck based account of Knowledge premise.
It seems immediately plausible that an insistence on social epistemology (as a normative claim) begets the same exact phenomenon that he bemoans. If you set up a clergy, don’t be surprised that sectarianism emerges.
4 Likes
6 Likes
Damn I was just going to link this. Hilarious and true. I LOL’d.
2 Likes
“the 2 people on your team that actually tried are now flattened by the slop code everyone is producing, they will quit soon”
… but they will not mention this in the exit interview (if applicable), so the company will double down on AI, because now it has to replace the senior engineers too.
I don’t mind people using AI as a tool. It can be very useful for repetitive tasks, to get started on something new, or as a kind of sophisticated search engine. What amazes me is the lack of ownership in some contexts on AI-produced work: when people hand in something that contains a grave mistake, they now blame it on the AI. Or when asked about code in a PR during a review, they ask the LLM to “explain” what it did in the first place.
2 Likes
Genuinely asking an LLM why it did something is a sure sign someone doesn’t understand AI’s well enough to use them safely.
5 Likes
In retrospect, it was genius of Google to provide “five blue links”, but leave the evaluation their worth to the user.
Humans similarly couch their statements in adjectives to tag their reliability: We “think that” something is true, we “have heard somewhere”, we “read an article that said…”.
It is strange to me that LLMs do not use similar safety nets. Why do they state everything as fact? Even when asking explicitly, they seemingly can not gauge the reliability of their information. I wonder why that is so.
1 Like
Isn’t that coming from the way they are trained - to give an answer at all costs? Someone compared it to answering a test question instead of leaving it blank, where perhaps here, the confident answers get more “points”? Idk
It is not done to say “I don’t know.” Closest one hears/sees is “I am not sure about blah blah.” Which is bullshit.
Even my dad advised me, sixty-plus years ago, to never say, “I don’t know.” My dad taught me many wonderful things, but I have always ignored that one.
I’m sure the AI-company marketing folk would forbid anything other than a show of complete confidence. Otherwise know as bullshit.
4 Likes
There was an observation early on, that for some questions, the 1st chatgpt models answered with ‘I don’t know’. Then, if you recall, there was a thumbs up/down indicator. People responded to ‘I don’t know’ with thumbs down; then, the network was retrained based on feedback.
With current models, if you prompt them carefully, they can say things like ‘in order to answer, I need the definition of function x’, but if you don’t instruct them to do so, they usually go ahead with assumptions / generalisations. Which is not uncharacteristic of humans, either, I’m afraid.
3 Likes
Found a brilliant comment:
The only worse thing than AI hallucinations are people’s hallucinations about AI.
3 Likes
I see how “I don’t know” is a bad thing to say in business and AI. But saying “Interesting question, I haven’t thought about that. At first glance, I’d think that…, but I’ll have to research it to get you a definite answer”, is one of the most powerful and trust-building things you can say.
An AI might answer “according to X, it seems that…, would you like me to verify this claim from other sources?”. Even a simple “this claim is contested”, or “doesn’t have many reliable sources” would help.
3 Likes