@molly0xfff I had my boss the other day selling us on the example of an AI that....waitforit..... could take minutes of meetings.
He seemed to think this was impressive, because surely this is a feat no silly bumble-fingered human could accomplish with any degree of accuracy.
Sadly, with our new CEO being apparently a fan of all things AI I reckon my company is going to see a lot of this kind of desperate gold-rushing evangelism from people hoping to appear 'aligned to that vision'.
@molly0xfff Not surprising considering it's the exact same scammers behind both, and even the same business model (selling GPUs/compute time to do utterly pointless shit to fools who think they'll make money on it).
@molly0xfff You use AI every time you make or receive a payment. Without it there would be a lot more fraud undetected and a lot more false positives (legitimate transactions wrongly declined).
@molly0xfff It sure uses a lot of power and cooling water to do a meh-job. Maybe using an intern (or fiverr) would be a better choice because it also helps a person.
@molly0xfff Hi,
If I use my GBP Credit Card, I'll get a $2 currency charge?
Do you know if "LINK" allows billing in non USD currencies? I've never heard of or used "LINK".
@molly0xfff That’s a really interesting piece. I’ve been experimenting with AI in writing for a client a lot over the past year and my findings are close to yours.
Writing from scratch: terrible
Research: sporadically useful but often inaccurate
Finding tip-of-the-tongue words: excellent
Iteration: useful with caveats*
Proofing: mediocre**
@molly0xfff * I have found it possible to speed iteration by throwing drafts through an LLM. But afterwards, I tend to blend output with my original text and rewrite. So I’m removing 1–2 drafts from the process.
** ChatGPT in particular will spot some errors. But I’d say it misses 50–75% of them, and up to a half of its recommendations are flat-out wrong. However, it does find things I might have missed.
@craiggrannell I've found Claude to be far better than ChatGPT at proofreading. Not perfect, but it doesn't try to point out "errors" that don't exist in the original as much as ChatGPT.
Mostly, though, I just use my Mac’s built-in audio system to speak my articles to me at ~1.5x speed. That tends to help me pick up errors I’ve missed better than I’d otherwise do solely with read-throughs – or AI.
@molly0xfff Of all of the text helpers, I think the one I’d really miss is Otter. Not because I don’t want to pay transcribers. But because when I have, the results have never been as good as what I get from Otter.
@molly0xfff I think you make a good point that for experienced workers it can provide some efficiency, but in near term I worry that companies or going to limit hiring and increase layoffs because they think it replaces junior people
@molly0xfff the robotics uses of LLMs seem really cool and what I am most interested in. Giving a robot an approximate understanding of the real world seems like what robots always needed to be useful outside of narrow domains. https://github.com/GT-RIPL/Awesome-LLM-Robotics
@molly0xfff I’ve found Copilot-style autocomplete code tools to be more irritating than useful, but just straight out asking ChatGPT “If I have a table with x and y columns […] what’s the sql query to return all rows where […]” almost always gets me the correct result. Same for “How do I X in Python” or “what’s the zsh command to…”.
@molly0xfff I 100% agree with both your views on positives, shortcomings, risks - and obviously, the grifter hype and overpromise.
However, for me, it's still a huge net win. For example, writing code at work without Copilot I feel less productive or even frustrated (use it at home for personal projects). Even if it's "just autocomplete" and misses the mark a lot - it still is an amazing autocomplete.
Btw., here is my recent post on my uses - a lot overlap with yours https://bartwronski.com/2024/01/22/how-i-use-chatgpt-daily-scientist-coder-perspective/ :)
@molly0xfff and I think exactly like with web3 and crypto stuff - a) time will filter the bad actors and grifters and this part will fall apart suddenly and completely. b) people will understand those tools better and their limitations/overpromise. c) there is more and more push on regulations to make it more ethical - hopefully, those will succeed.
@molly0xfff This is the core message to me... and it's a self-defeating proposal. I see a lot of second-language English speakers using chatGPT to sound more "professional" but it just ends up looking obviously chatGPT. There was a nice post here a while ago framing this as a racism facet of LLMs: https://kolektiva.social/@FractalEcho/111949020988947948
@LeoRJorge yeah, you may have noticed the instructions I gave around proofreading had to include very explicit instructions to just proofread, not rewrite it in that godawful GPT "voice". i imagine something similar is necessary when trying to get it to help with English
Add comment