hammerhead

@sphyrna

shark

www.pillowfort.social/sphyrna

hammerhead @sphyrna8/14/2023, 2:44 PM

nora @nora

Misty @dog8/14/2023, 4:28 AM

Checking in on google

(ht Imran Khan)

#chatgpt sucks #google sucks too #why is this what the future is

Tom @iiotenki8/14/2023, 5:18 AM

I've never taken on this sort of our work out of principle, but in the translation realm, a lot of agencies, even game specialists, are trying to posit MTPE (machine translation post-editing; ie: only coming in to edit machine translated work) as this way for us to more efficiently do our jobs and churn out more translated content faster (because obviously that's what you always want from a translator, the person who can put out the most material quickest) when so much of those jobs in reality involves fixing generated translations with mistakes as catastrophic as this, all because the algorithms are essentially programmed to play linguistic jazz the moment they run into any textual patterns they haven't been trained upon ad nauseum with their sample data. Which, when we're talking about a base language with grammar placement as flexible as Japanese, happens extremely often. With all but maybe the dead simplest material (which is of course never what these AIs are being deployed on at agencies), the end result is a project that takes more time to complete than it would've to just have us do it all from scratch because we're spending more time manually evaluating strings and putting out fires and less time doing the actual work.

Pattern recognition applied in these sorts of ways isn't computerized intelligence making an informed decision no matter how intricate the data links get; it's computers taking a gamble based on probability on things that either shouldn't need it or result in demonstrably worse output when that's the overriding consideration. That these companies think they can money and data volume their way out of these problems when it's built into the very paradigms they're reliant on speaks to how deeply these efforts are being spearheaded by the people least equipped to appreciate the issues and limitations in what they're selling.

You must log in to comment.

in reply to @dog's post:

adam le doux @ldx8/14/2023, 4:31 AM

oof :/

noted ear-haver @rachelmae8/14/2023, 5:05 AM

i just fuckin' can't with AI evangelists anymore. anyone that seriously suggests it as a solution to anything is dead to me

Misty @dog8/14/2023, 5:09 AM

God, right?

Aepoh @Aepoh8/14/2023, 5:21 AM

Goodbye search engines, hello library (if you have one :(

Internet Janitor @internet-janitor8/14/2023, 5:35 AM

Waiting in pained exasperation for the apologists who are about to Kramer in and insist that humans make mistakes sometimes and therefore it's fine for computers to fuck up constantly.

Aepoh @Aepoh8/14/2023, 5:41 AM

That definitely will not happen on an ASSC space. This place is repulsive to techbros and cryptobros

fingerless @fingerless8/14/2023, 6:23 AM

IT teacher had a saying. To err is human. To err spectacularly is machine.

@screamingturnip8/14/2023, 8:58 PM

sometimes I remember these posters in our computer labs in middle school with a friendly computer saying "to err is human, sorry" and like yeah cool I didn't requisition low spec 486s load them windows 98 and pretend that its a 12 year old's fault that they're busted coffee stained garbage
edit: this ain't like a rebuttal just a weirdly connected memory

Christian Simmers @Xuelder8/14/2023, 1:06 PM

I'm working towards a degree in machine learning and I'm just head desking trying to figure out how tech bros managed to cause another AI winter when we were finally starting to make progress again.

I should have gone into bioinformatics.

Demand-side Sheepdog @kojote8/14/2023, 3:57 PM

This is great because it is basically a “GPT-3 fucked up”-only excuse. Like if the bank accidentally charges you an overdraft fee absolutely none of them are going “yeah, well, I bet you’d be even more inaccurate guessing how much money is in your couch cushions, so you’re just a hypocrite blaming software for doing the same thing.”

Watching autopilot fly a jetliner into a mountain and shrugging my shoulders because it's completely unreasonable to expect computers to be perfect when, after all, most people can't do vector algebra, now can they? It’s not an autopilot failure, it’s just “spatially disoriented!” Like a human pilot! That means it’s learning!

sabertoothed jungle doofus @pawcelot8/14/2023, 5:58 AM

Hey Google, uh, you good?

@Happytoaster8/14/2023, 7:37 AM

Creating spam articles to be indexed by Google is sadly the most practical use of this technology.

Daniel Martin @fizbin8/14/2023, 10:38 AM

What I like best about this is that when I tried it, I got a generative AI response from Google claiming the same thing, so then it's Google's AI snake oil repeating garbage generated by OpenAI's snake oil, instead of just OpenAI's stuff fooling Google's "authoritative answer" algorithms.

Christian Simmers @Xuelder8/14/2023, 11:21 AM

So when that USA Today article about those scientists that made a proposal that the universe might be twice as old as we originally thought, the AI algorithm started to cite that as fact and would keep saying the universe is 26 billion years old.

philosophy major (derogatory (reclamative))@nullpat8/14/2023, 1:58 PM

sometimes it feels like the current "AI" lurch is entirely motivated by people who are angry the public has access to facts

Aepoh @Aepoh8/14/2023, 7:21 PM

I don't really think it's malicious. Maybe a few small bad actors are. But I mostly see it as a symptom. In our current attention economy, clicks mean ads, and ads make websites money. If you want to make passive income with minimal time investment and effort, making bullshit websites is a pretty good way to do that.
It's not really a difference in kind but a difference in scale. Websites similar to the on in the OP have used ghostwriters to spew out poorly written factoid articles at a breakneck pace. The little profile picture and name you see next to most articles on the web hasn't had much real accountability for years now. It's just that now instead of a human getting paid very low wages to write redundant and poorly made articles, they're getting large language models to do what they see as roughly equivalent for even cheaper or free.

At least when they were using human ghostwriters they would usually at least try to compile somewhat accurate information. Even if it was just the same stuff that 50 other websites already had.

D W the Gutless @dwoboyle8/14/2023, 2:43 PM

The thing is, how difficult would it actually be to create something that could answer this question?
Generate a list of countries in Africa, filter out to the ones that start with "K", and return the number.
I gotta imagine the engineers at google are smart enough to make a thing that could parse a question like this and others without LLMs or statistical analysis.

Matthew "Queso" Niederberger @Queso24698/14/2023, 3:11 PM

Building programs that do one thing are easy. Building things that generalize a problem space like that are hard. Building something that decides which of many problem spaces a question falls into is nearly impossible. Building something that both decides what a problem space are without an extant solution, and then generates a solution in that space is a literal general artificial intelligence.

D W the Gutless @dwoboyle8/14/2023, 5:15 PM

I understand what you are saying. But surely you can make something that looks at a simple question and can parse it.

"How many", "What are", etc.

Their search engine already passes simple questions like "burger places near by".

Most simple questions like the one in OP are basically just coming up with a way to convert a sentence into what essentially is SQL query.

Internet Janitor @internet-janitor8/14/2023, 5:40 PM

There is a combinatoric explosion of possible "simple questions" and approaches to solving them. You could chip away at the problem space for thousands of man-years, one special case at a time, and still only cover the tiniest fraction of the territory. In full generality it is simply not tractable.

Lizardguy @Lizardguy648/14/2023, 3:44 PM

Yeah, this is why I use Duckduckgo.

Aepoh @Aepoh8/14/2023, 7:11 PM

That will not save you. Just the other day I was looking up some laundry information on DuckDuckGo and was getting suggested articles that were clearly AI written.
https://theinteriorevolution.com/should-you-use-fabric-softener-on-bed-sheets/
https://sleepation.com/before-you-use-fabric-softener-on-your-bedsheets-pros-and-cons/
These were page 1 search results on DuckDuckGo btw.

These sentence structures and way of presenting information are clearly not human. Contradictions, repetitions, non-sequiturs. All hallmarks of Chat-GPT esque writing.
Search engines--not Google--are degrading, have been degrading, and large language model generated articles are just the final nail in the disinformation coffin. The days of asking your search engine a simple question and getting a simple fact result are going to fade away if search engines do not crack down on SEO abusing sites. Searching up even basic information is going to become such a chore, especially for the information illiterate: those who don't have or weren't taught the skills to parse reliable information from bullshit. Not only will you need to verify the veracity of the information you're being presented with, but also the authenticity of the website you're getting it from and whether the article you're reading is even written by a human or not. We're going to need to start doing that just about every single time for most pieces of trivial information. And I imagine it will become even harder to tell the difference over time as these large language models improve and become better at spewing more fluent bullshit.

What once granted up unprecedented information efficiency is now going to become extremely information inefficient. I feel really bad especially for the neglected kids who weren't taught many basic life skills because their parents just figured that because we have access to the world's largest encyclopedia in the palm of our hands they don't need to do their job as parents.

Lizardguy @Lizardguy648/14/2023, 8:33 PM

Why are you searching laundry info online? You just put clothes in the washing machine and turn it on.

Aepoh @Aepoh8/17/2023, 11:18 PM

My comment isn't even about laundry info. The only reason I included those links was an example of LLM-generated articles that could be found via DuckDuckGo.

I'm not really sure why you laser-focused in on that of all things, but to answer your question it's because my parents used fabric softener growing up so naturally I just assumed you used it in every load, and only recently decided to look into whether it was necessary or what to use it in or not use it in after something didn't feel right with my sheets.

Tempest @sutempest8/14/2023, 4:23 PM

The intent they sold us: "AI as a writing assistant to help actual writers"
The shortcut: "AI as .... ..... .... writers"
The result:

threnody @threnody8/14/2023, 6:58 PM

"What were you doing?"
"He doesn't know..."

in reply to @iiotenki's post:

Christian Simmers @Xuelder8/14/2023, 11:16 AM

I did AI research(technically still am? ostensibly my masters program is in AI) and some of the biggest problems we kept having was because of the public domain resources we used as training data ie court documents and evidence, we would have to filter out some of the weirdest shit to make it kinda sound "normal." Like for a period the language model would keep using the word holocaust to describe everything. And how for the longest time all models were very misogynistic cause a non-insignificant amount of the training data was built on top of Enron's emails.

Nick @4gateftw8/14/2023, 6:01 PM

Remember that time when Florida was using a racist algorithm for fucking prison sentencing, it got exposed, and somehow nobody cared? (for the uninitiated, simply Google "COMPAS recidivism" to see the hellscape we currently live in)

Christian Simmers @Xuelder8/14/2023, 6:10 PM

What's really bad is that unless you got specific, I would have had to ask you which one.

Nick @4gateftw8/14/2023, 5:50 PM

What I wonder about is where exactly the lack of discipline comes from; what leads tech companies to want to present what would be an early proof of concept at best for an academic research project as a finished, ready-to-market product? I say this because all the researchers I know in that field are generally responsible people who are very cautious about the claims they make re: their work. Are they the exception, and a lot of LLM researchers are just insane weirdos? Are business people just pushing projects into weird places? Or is it both? And who are all the pricks online who call everyone who doesn't agree that Kenya isn't real a luddite? Questions for the ages tbh

Christian Simmers @Xuelder8/14/2023, 6:08 PM

I think the Silicon Valley Hype Machine folks are looking for anything to keep the funding flowing into their companies. They failed with NFTs, Crypto, and the Metaverse, so now they are looking for something to pivot to. It just so happens AI researchers have had great leaps over the last decade (a lot of hardware innovations caught up with a lot of the algorithms we designed as far back as the 50s). ML and LLMs can provide solutions for specific problems, but now people see an AI chat bot and think it's a general AI. And the hype machine wants to market that. Doesn't capitalism just rule! /s