Comment on Driverless car startup Cruise's no good, terrible year
hellothere@sh.itjust.works 8 months ago
I’m not sure why I feel the need to preface this comment, but, I work in software, I get how hard a problem autonomous driving is, how genuine safety improvements over human drivers are highly valuable, and how perfect need not be the enemy of good.
However, the level of sheer blind optimism from the AV crowd is the same as the AI “leaders” and the crypto bros before them. How their statements are not straight up fraud is beyond me.
The reality of them needing to have a remote team of drivers having to intervene every 2 to 5 miles of driving, within an urban setting very much designed for cars, is so far away from the picture they paint.
No wonder the tech industry has a dog shit reputation.
SlopppyEngineer@lemmy.world 8 months ago
People have been saying since the 1950 that this time they really cracked the code of AI and soon it will . I mean some serious steps have been made but in the end it never lived up to the promises made. I’ve got no reason to believe that this time will be different.
wantd2B1ofthestrokes@discuss.tchncs.de 8 months ago
I mean, I think it’s a little different in that there’s tangible AI products that millions of people are already using?
hellothere@sh.itjust.works 8 months ago
I’d describe those products as being in the “heh, that’s neat” phase. At least, the good ones anyway. After a few hours you start the see the issues - like images with very weird fingers - and then you the illusion is broken.
For example, in work I’ve been playing around a lot with various LLMs to generate marketing copy for physical products. I’m being vague here for reasons I’m sure are obvious.
Across the first few hours it felt very impressive, it would pretty much instantly churn out descriptions in a variety of styles, even when the product information we provided was low quality.
The problem though, was that the copy it wrote was actually - if you sat down and read it - shit.
Now, the vast majority of marketing copy is also shit. The good stuff is excellent, rare as hens teeth, and incredibly expensive, but your generic boilerplate crap you see all over the Internet, that stuff, it could replicate perfectly. Even the lies.
If you want sales speak waffle, it’s 11/10 every time.
So we’re currently in a bit of a bind. Do we release a tool that creates boilerplate shit more easily, and will turn even the most inconsistent (and straight up conflicting or impossible) data in to fancy sounding text, and jump on the AI-powered bandwagon, or do we not, because what it creates isn’t actually any good?
CoggyMcFee@lemmy.world 8 months ago
I agree with what you’re saying about your line of work. I code for a living, and Copilot is genuinely useful all day long. I use it now and again to generate a script from scratch, but most of the time it functions as either an incredible autocomplete of whatever I am coding, or it converts a chunk of code from one format to another, with just a description of what I’m trying to do, instead of me having to write a complex regular expression or do it by hand.