Let's imagine you ran a business. Say, a game development company. Would you use AI for voice lines and asset creation?
What about advertising? Would you use AI generated ads?
I'm curious what peoples' boundaries are in regard to AI in their lines of work.
I work in research. I happily use AI to help me generate code, explore datasets, summarize papers, do literature review. I am uncomfortable using AI to help me write, other than as a spelling/grammar checker. I'm uncomfortable letting AI do the core analysis, preferring instead to use AI on data cleaning & normalization.
I will use AI sometimes for feedback on my work, but I find that I can't really trust it to give a balanced take. I often feel like it's either too nitpicky or too sycophantic, and I can't calibrate well between the two.
AI ad is a error because the audience could taking more time looking for AI details and complaining about how shit it is than paying attention for the message. Bad publicity is a publicity anyway, not in this case.
Im sawing nany small merchant ysing AI banners ad as a pandemic , at same time its good see more elaborated and visual enjoyment ads its a mess, because looks like everyone is using the same artist to do their banners and its a prolific one that pollutes the message with slop information.
I use AI a lot with tons of oversight when coding. Usually I end up producing a prototype in a day or two, then I spend a week (often more) interactively reviewing and reworking the prototype (with AI and in a divide-and-conquer fashion) into something I'm confident in the quality of. At the end of the process, I generally feel like the features are more complete and higher quality than I'd make by hand in the same amount of time.
I'm still learning how to do this more efficiently though. I made the mistake of trusting AI too much in a few recent PRs. Only after getting deeper into review did I realize several parts were over-engineered and practically incomprehensible (slop).
A process I'd like to try is:
I suspect I'd get a lot of the benefits I get now but with ~100 prompts rather than ~1000. And, perhaps, I'd still be able to enter a flow state while programming.
So, I would use AI wherever I could so long as I could maintain full oversight. I will never feel comfortable using it for writing and environments where earnestness is 99% of the value though.
AI-use for mindset may be a good litmus test for what the writer's mindset was when he/she was writing.
Someone who is mostly interested in expressing themselves and establishing a human connection with someone else is not gonna use AI in their writing.
Someone who views the writing as instrumental for some other purpose, e.g. influencing opinions, getting followers, having a line on the resume, passing a class, etc, are more likely to turn to AI.
That's what I've been testing for a while now with js code and I like it better than amending and replacing the slop. I still sometimes make the mistake of keeping something generated - c&p it from a bot PR or comment - so I'm being extra careful to not be lazy and just think everything through.
I feel immensely slow though! Especially bot->write->bot review->amend->test and then have a finding.. fix it->bot->amend->retest -> another finding ... very repetitive and sometimes I feel like I cannot get it right. At one point I was thinking that I should test without the review first, but I am scared to then get regressions later or miss an edge case.
For game development, I'd be pretty comfortable using AI for all of the above. That's because I like the idea of developing games (I'm working on one now) but I don't know how to do just about any of those things well.
I'm curious what you mean by "letting AI do the core analysis".
by that I mean letting the ai unilaterally decide the methodological framework which will be used to establish a result
also, I don't trust the ai to interpret the degree to which a result does or doesn't support a hypothesis
100% agree. I never let an LLM decide anything. If it made a decision then it means my instruction was incomplete.
After Opus 4.6 release, and experimenting a bit, I added instructions to my intel gathering processes on 3rd party code to summarize the findings to me. I don't use this as definitive but it cleanly warns me in case there is a finding that is important to look at. It's not so much that I will only read that finding, but that I want to make sure that I won't miss it. This works very well with GPT 5.x and GLM 5.2 too.
For example, when I woke up this morning there were 56 findings waiting for my inspection regarding the new signal release, 2 of which were highlighted in summary. Both were in
webrtcwhich is written mostly by Google, and were important to validate, and understand context on. Without the nudge I could have read over at least one of them in the long list of things to do, especially with my first coffee.This makes me think that you can have it flag things up to you, but don't trust an "all good" verdict (so don't ask for it - it took me quite some effort earlier this year to make Claude to stop guessing the severity of an issue it spotted, because every bs item it finds was
HIGHif it was the most important thing it found.)Yeah, AI seems pretty bad at calibrating the severity of a problem.
I've tried using it for feedback on my work. I find that the feedback it gives you actually depends a lot on whether it thinks this is your work or someone else's. If it thinks it's reading your work, then all the issues are minor and fixable. If it thinks it's someone else's work and it's helping you write a referee report, then the same issues are paper killers.
I've even had it tell me to upsell a particular claim more in one session, and then in another session tell me that the claim is undeserved and should be taken out.
Well... which is it?
Oh yes. Everything I do is through a forge so it always reasons about everything in third person. "Maestro (the handle I use) requested that...." Or more fun: "The user golden-apples (another bot account) said: ...". Benefit of most of the training data being done from personal access tokens and the bot expects that its own user is the operator's account. I like this form of automatic confusion - it helps.
Haven't seen that in a while. I don't like that behavior but I never ask for styling advice. Only facts.
There was a time where I would give Claude 2 separate accounts and critique itself but I found that it works better to let another model do the review. Different training, different results. Currently Claude does mostly review for me. GLM does most of the work. I cancelled my GPT subscription so now that I pay top dollar for it over API, I rarely use it. Not worth the cost.
There wouldn't be much point to us if we could trust AI to do those things. Of course, there might not be much point to us either way.
I do use it in my small business for price research. I regularly have it run market research to see what something is worth and whether I can arbitrage it reselling or not. The speed at which that research comes back gives me the ability to make buy decisions in seconds and minutes versus taking hours to do it myself. I also use it for coding but the import library risk is a problem. Seems to be best to let it run in VMs when testing coding versus anywhere else.
Do you find that the price research results are pretty accurate?
Since I do a lot of resell of used stuff it does come in handy. Mainly, there's far more information available online, so what the AI provides in comparisons and references pan out far better than if it had to project unique, new stuff in market research.
Yes. Isolate them. I also automatically cycle workspace and homedir per session. If you can't, at least cycle it manually once a while. Lots of crap accumulates and its very hard to keep track of everything ever installed.
Token spend on installing tooling does rise but can always script the common items. I.e. I have pinned tools like
uv,pip-auditandnpmin my container image to save confusion. Painful to maintain the pins though. Also of the docker base images.An AI generated ad has a tone and feeling to it, irrespective of the actual content of the ad. I can imagine some projects where this would really kill the project. For instance, I'm thinking of Lyn Alden's recent work of fiction: this is something that I think probably didn't suffer too much from having an AI generated cover (although I think it would have been more appealing if it hadn't been) -- but using an AI ad for such a thing was risky. In general, perhaps human made content requires human made ads.
It probably really depends on what the product is: if it's anything that needs to feel high touch or human, I wonder how much its use needs to be almost hidden.
That kinda makes sense.
I definitely think "hiding AI" is going to become a thing. In the games industry there are already some high profile cases of backlash that developers got for having used AI in their internal processes. I personally thought that their AI-use was mild and within the realm of acceptable usage, but some people are just vociferously against AI and will make a big ruckus about it.
I've used AI earlier this year for a huge (in terms of document scope) policy assessment project, but that was not generative: it was only for processing massive amounts of nasty looking scans from typewriter docs made in the 70s and 80s and making it searchable, and then in some cases helping me resolve references to documents published 20-50 years in the past, and what the "changelog" was on it from then to now.
I don't really believe in creating the end product with AI automagically, not in code either - at least not for things that are seeing daylight. I've been amazed at some software that is (or was) high quality, high reputation, and when they start shipping a claude.md in their repo, I start seeing a bunch of crap when I review. For the well maintained stuff, it's mostly only dead code or oddly structured functions. For the poorly maintained stuff... it's costing me a ton of time to make sure that there really isn't an error in "perfect-looking" code, and then I often find what seems to be assumption overload, i.e. poor prompting.
It does get a bit better with the newer models, because they've been trained to do some of the self-adversarial reflection, but I still feel like the one-liner claude code prompts and conversational LLM usage in general are not helping in quality; it all depends on the skills and attention to detail exhibited by the operator.
I've never had an end-to-end successful AI automagic experience. It's definitely iterative.
I like to believe that I'm a somewhat skillful operator, and that's partly because of my long years of experience with the work domain, and doing the workflow without AI.
I do wonder the extent to which young people can become skilled operators without having the same years of domain experience.
That being said, I'm not a pessimist on that. People probably wondered if calculators would make people bad at math; and in a way they do, but in another way they also expanded the ability of more people to use math as inputs into bigger projects.
Be a pal and develop an SOP. I'm slowly developing a huge set myself. I've built reference material from books about how to not just write but evaluate and maintain standard operations and I have a long list of jobs that I have standardized within my own framework that need to be documented and evaluated. Lots of work to do!
To the degree that it doesn't create a dependence.
AI usage != quality control
A shitty AI generated ad is unacceptable. But there are shitty human made ads too. There is AI slop code. But there is shitty human made code too.