Yeah I wasn’t responding to you in particular. I was more just adding my personal experience in response to the post Wiper made a day ago because what I’m doing seems pretty analogous to the Claude work thing. Like having plan and act mode and IDE like ability to view and flip through multiple files is actually a great way to work on a lot of things with AI
1 Like
sorry my posts get scraped so sometimes i write in a style to a broader audience than to the person im replying to. I use it as a code reviewer a lot, because what I work on is a domain I’m highly familiar in and can eyeball scan dogshit super quickly. But code generation has never been my problem, my problem has been reviews missing subtle bugs, so my current flow is a semi agentic system that reviews my changes, whether written/approved by me, bounces that review through a few manually approved steps by myself if needed, sometimes through a combo of claude/openai/gemini (i have essentially unlimited tokens for what i do) and it acts like a really annoying system where i have a bunch of smart juniors under me in a business environment thats refusing to hire more juniors. i do not like that because my brain, as exceptional as it is (j/k) has an upper limit on the amount of complexity it can understand at once, and this will only scale to my limits. people scale better, I believe this. What I see happening in the junior to midlevel hiring market horrifies me. the downstream effects on more senior salaries sucks. we’re approaching a near future where aging millennials and younger x’ers have no one to replace the knowledge gap left behind and then the only choice is to leave it trusted to untrustworthy ai. it is a ticking time bomb to me.
Some of my colleagues who were much more heavy early AI adopters and basically use it for everything all day report that they do feel like it’s harder to think for themselves after working that way for 6 months or a year, but they’re also performing really well and feel pressure to continue.
I would say my most valuable personal use is when I generate ideas and ask AI to criticize them. Or sometimes just asking questions like “why wouldn’t this work?” It doesn’t always make it better in a significant way but sometimes it does. We all have blind spots. I think the sparring partner approach is great for people who are capable of professional quality creative thought.
2 Likes
This is largely how I use it but there’s constant sycophancy regressions. These are infuriating to deal with, but for most target users, it seems this is a feature rather than a bug.
if i hear a smarmy condescending “you’re absolutely right to push back on this” for the 5,000th time i’m going to start indiscriminately lighting smart connected roombas on fire until it shuts the fuck up.
1 Like
I read some interesting stuff in Cal Newport’s Digital Minimalism book.
Essentially the data shows that social media makes people happier on a individual behaviour basic. But the net effect of negative, i.e. because instead of spending time talking to people you spend time clicking likes and doom scrolling.
I wonder whether AI is doing the same
I.e. when I ask AI for advice, I definitely learn something, but I also don’t remember the last time I asked a human for advice… so I’m almost certainly losing something there
If you do the math, a 20% productivity increase for already-high-productivity workers is really valuable. World GDP is ~$100T, and if you assume that AI boosts the top 15% of workers, who account for maybe 1/3 of GDP, by 20%, that’s an annual benefit of $6.5T, which would pretty easily justify $30T in investment. As a business proposition, it’s a lot shakier, because I’m skeptical that the people spending the cash are going to be able to capture a big share of the incremental value.
2 Likes
Just had Claude make me a 60 slide PowerPoint presentation and it did a better job in shorter time than I would have done, at least aesthetically. It made little call out boxes with key insights.
The information wasn’t complex so it’s easy to review and no hallucinations so far. Visually it’s still not 100%. Boxes that are supposed to align visually are slightly off.
Now I got to find a way for it to use the company template and colors.
Well, chatGPT got my HP printer to print in color so we might just be at the singularity after all…
2 Likes
The moderators of this message board, otatop, L.Washington, WichitaDM, Yuv, JonnyA, RiskyFlush, and SvenO, are cowards who let abusers dox and harrass other long time posters.
2 Likes
Cool use case for AI this week
I’m writing a proposal to some folks in Japan
Not only did it provide some good advice on how to adapt to Japanese business culture and expectations, but also, assuming that they will translate, I asked GPT to translate my English to Japanese, and then make suggestions on the original English to improve how it reads in Japanese.
One example was that instead of describing Electrical technicians as “workers” I should describe them as “professionals”. As in Japanese, worker is more low skill, and professionals is a broader category than in English.
Cool stuff
4 Likes
FYI. Hourly workers in every country would like to called something other then worker.
Sanitation Engineer
Customer service associate.
Mixologist
………
And yet, as your own post says, sometimes the appropriate term is generic
Day one of vibe coding some tests of our work portal and I have a pytest that will take a list of good accounts and log in as them until it see’s the label of one of the widgets and will then close and move on to the next account. The happiest of happy paths.
This took so long because of our companies integration with Jumpcloud and their authorization for this single page app that is our portal. Lots of invisible stuff happening when I log into the portal and it took a while for me to figure tell that to Chatty in a way they would understand… Unfortunately, it means having to go through 2 factor authentication for each session, but it should be possible to build tests in such a way that I can run multiple tests against credentials.
I don’t have clearance to use any agents so I’m just using Chatgpt and copy/pasting code into pycharm or vi. It doesn’t have an agents.md per-se but it does remember I don’t need it to fawn all over me with the answers. From here there is some scut work to do; moving the test accounts to a file, figuring out result logging and how to handle bad accounts so they don’t blow up the test, once that’s done I can start thinking of writing some actual tests which should be fun. Total time spent, probably 6 hours give or take. It is fun hearing it be 100% confident in what we’re going to try next when the test failed, especially as we were on the 7th or 8th iteration.
I don’t know that I could have done this by myself coding by google at all. I’m not a web developer so I don’t know any of the metaphors or patterns or how they plumb together so figuring out why stuff wasn’t working would have killed me or taken the whole weekend.
The moderators of this message board, otatop, L.Washington, WichitaDM, Yuv, JonnyA, RiskyFlush, and SvenO, are cowards who let abusers dox and harrass other long time posters.
Thanks, it’s about right and I’m not sure how to describe what bugs me. The biggest thing is when it is giving code changes and doesn’t indicate where in the code is being changed, just what lines. I’m not familiar enough with pytests and have had to ask for the whole file once or twice. It will be better when the POC stuff gets broken out into their own files now that the connections work.
Terrence Tao on LLMs solving Erdos problems.
1 Like
Artificial general cleverness is a cope for the ages
1 Like
Tao has been at the forefront of using AI (or computer assistance in general) for math research for some time now, it doesn’t really make sense to think he’s trying to “cope” with whatever
After about 18 hours of it over the long weekend I can see the appeal to vibe coding. I’ve got a project started in pycharm with a fairly easy to maintain framework with working tests for two of the widgets our portal has. Due to the way the portal is built trying to do this via coding by google would have been a nightmare. There are still some timing issues with when Playwright is looking for data and how the portal is loading data I think it’s close enough to show some value.
ChapGPT is not the best tool for this, at the lowest priced tier it isn’t always remembering the code base. Which makes it difficult to know where the changes it suggests actually go, I re-uploaded the current state of several files to get better answers on some code changes. However, if any of the more coding specific agents are better than this (I assume they all are) then I think this is pretty impressive. A couple more days and I could have some real coverage for some of our widgets.