Tech CEO applauds an employee selling their dog to accommodate return-to-office

I am childfree and 15 years into my career.

Over that time span almost everyone I worked with closely has had children. And for the life of me, I cannot think of one instance where it affected their overall productivity.

If anything, kids tend to be a stabilizing influence. I much prefer working with the person who is AFK every day from 3:00-4:00PM because he is picking up his kids versus the person who drops offline at 3:00PM randomly because they scored some concert ticket and the venue is 2.5 hours away.

Your org sounds different from mine though. It sounds like immediate responses to short notice requests are very important. Maybe folks are getting rewarded for heroics and not for planning and proactive communications? That is not an atmosphere that is conducive to parenting. Or anything else, really.

They build the tools and platforms for digital marketing, but Google at least, outsources the actual digital marketing to third parties. If you reach out to their sales team for "enterprise" GMP solutions, you'll be connected with a reseller.

I found it really helpful for porting an extension I'd written for Atom (now unsupported/defunct) to VS Code.

Now, if I were writing VS Code extensions on a daily basis, I'd want to spend the time to learn all the details, but (ideally) I'll never have to write one again. This is the only one I ever had to write for Atom, so it's a decent bet.

And yeah, most of the code was boilerplate-- but still ChatGPT saved me from having to learn how to write that boilerplate. At least a day of time, I'd reckon. Possibly several days.

The task was to recognize the Scheme expression (atom or pair) preceding the current caret position, sign it using a password hash, send it to a remote Scheme REPL for evaluation, then insert and select the result of the evaluation in the editor.

ChatGPT essentially wrote everything for me, except the part that parsed out the Scheme expression. It wrote a really horrible regex for that. Fortunately I already had a proper parser implemented in the previous Atom extension, which dropped right in.

I think there's a very narrow window where I'm getting useful results out of it, enough that I'm not completely writing it off as a way to help out with some productive work. I do expect I'll make use of it occasionally while programming in the future. But all of the times I've used it, it hasn't been an extremely clear productivity boost (certainly not at the x2/x10 level), and I do need to babysit it way more than people are suggesting.

I go back and forth in my opinion on it. There are some topics where I get useful sessions, there are some topics where it's a mess. The problem is I don't know in advance which are which. It seems to be most useful in scenarios where I either need some extra motivation to get over the initial hump of starting a task (making an task feel more approachable is I think a great use-case), or where I know just enough about a task to know if it's generating nonsense, but not enough to be able to knock out the same results in less time myself.

The other problem is that in areas where I'm not immediately familiar with what's going on, I sometimes have to spend a lot of time debugging its code. And reading and debugging other people's code is (to me) a lot harder than writing code myself. So there's an inherent slowdown there where I have to check its work as if I am reading a junior programmer's code and checking it for errors. Code reviews are annoying and take a lot of mental effort to do well. You can ask GPT-4 to explain itself, but there's no way to know without actually stepping through the code if its explanation is correct.

I also worry that it's railroading me a bit. I had it prototype an algorithm for me and debugged it and helped it expand the algorithm and get it working with the results I wanted. Only afterwards did I realize there was a much simpler way to get the result I wanted without invoking the algorithm at all. If I wasn't in a conversational loop with an AI focused on getting its code to work, I might have noticed that earlier. But my brain was only focused on debugging the code for that task, it wasn't focused on going back and rethinking the algorithm and cutting out the useless bits.

BUT... sometimes it seems helpful with prototyping? So it's not all bad. I don't necessarily want to dismiss it entirely.

My take is potentially useful, and for some people who are particularly suited for this kind of rubber-duck peer-coding style of work, maybe a big increase in productivity? But certainly not universally useful for every programmer, and I can still see the seams. GPT-4 is significantly better than GPT-3, but it's not as good as people have hyped it to me, and I think people are vastly over-estimating its capabilities. I went in skeptical, but even with that attitude from what people had told me beforehand I still expected GPT-4 to be better than it is.

Am I worried this is going to steal my job? No. That's honestly a kind of laughable idea to me after seeing the kind of code it writes and how unreliable it is.

----

It's hard to put this into words, but it also kind of hints to me at a much better system, and I can see problems here around the auto-regressive nature of the LLM and the need for alignment training and around the whole structure of the thing and the slowness of generating text over and over, and the sort of forced linearity of the chat session that makes me think that maybe GPT-4 is just kind of hinting at a debugging/prototyping methodology that would be a massive improvement to productivity but that doesn't necessarily involve LLMs or involves them to a much lesser degree. It's in some ways a very clumsy/brutalist approach to simulating a feedback repl and accessing documentation, and because it's clumsy/brutalist it over-emphasizes the natural language aspect. And I'm curious to see if future projects can identify what specifically is helpful here and deliver that without all of the inherent downsides of needing to use a conversational model.

That's all ignoring the privacy/data concerns, but they're a separate problem.

> , all of whom are remote and not on PTO, have not even opened their laptops in a month?

But in the call, he said "I'm not saying they haven't done any work...."

So- these people can work without a laptop. Can they work as well? Who knows, he's not giving us any details - but this isn't a case of them doing nothing over the last month, or he wouldn't have added that caveat.