|
Google Bard is now in open[1] access. [1] Users[2] must sign up for the waitlist and wait an unspecific amount of time. [2] Only some users, in some countries. Boris Galerkin fucked around with this message at 16:02 on Mar 21, 2023 |
#
?
Mar 21, 2023 16:00
|
|
|
|
| # ? Jun 1, 2024 15:07 |
|
|
Ghost Leviathan posted:Okay maybe the funniest possible outcome is specifically because of treating image and text scrapings as bypassing copyright, it makes it effortless to train your own AI on scraping another AI's output and then sell yours at a fraction of the price, but then your competitor pays that once and trains their AI on it... Realistically, how much sample input and output from an optimized model do you need to train another LLM to be 95% it's equal? I actually have the Alpaca Lora installed on my local LLaMA instance and it's neat, but nowhere close to even non-GPT4 ChatGPT.
|
|
#
?
Mar 21, 2023 16:05
|
|
|
LASER BEAM DREAM posted:Realistically, how much sample input and output from an optimized model do you need to train another LLM to be 95% it's equal? I actually have the Alpaca Lora installed on my local LLaMA instance and it's neat, but nowhere close to even non-GPT4 ChatGPT. At this point, your guess is as good as anyone else's. The Alpaca results are impressive because they were obtained on the 7B LLaMA, using 52Kb of scrapped data. Both of these numbers are ridiculously small, and are showing that conventional wisdom with regards to required model sizes, fine-tuning dataset sizes, etc. don't seem to apply here. It's going to take a while for the dust to settle and new rules of thumb to be established. Aramis fucked around with this message at 16:23 on Mar 21, 2023 |
|
#
?
Mar 21, 2023 16:21
|
|
|
Did I miss something or did no one really consider the possibility of their AI being used to train someone else's AI and then being undercut?
|
|
#
?
Mar 21, 2023 16:29
|
|
|
Electric Phantasm posted:Did I miss something or did no one really consider the possibility of their AI being used to train someone else's AI and then being undercut? It was considered, but the general (erroneous) consensus was that this would require millions upon millions of queries to the "source" AI, and so this could be prevented at the API throttling level.
|
|
#
?
Mar 21, 2023 16:35
|
|
|
Speedrunning cleanroom AI duplication. 
|
|
#
?
Mar 21, 2023 19:55
|
|
|
https://www.adobe.com/sensei/generative-ai/firefly.html Adobe's AI generator announced and stands out by having a dataset comprised entirely of public domain and Adobe owned/licensed works.
|
|
#
?
Mar 21, 2023 20:08
|
|
|
Smart, a lot of Adobes bread and butter is artists. Its not like they are beloved by them already,
|
|
#
?
Mar 21, 2023 20:21
|
|
|
Boris Galerkin posted:Google Bard is now in open[1] access. I signed up, 'cause why the hell not. I'm not sure if they write a different poem for each sign up. quote:May your day be bright,
|
|
#
?
Mar 21, 2023 20:32
|
|
|
https://twitter.com/saeeddicaprio/status/1638248160597364736?s=46&t=VqGDTpMBvtmBLtOHIUgZdw lol isn't this based on that AI where some google ethics dude was convinced was sentient?
|
|
#
?
Mar 21, 2023 20:57
|
|
|
This is like complaining that it's difficult to shoot a bird with a hammer.
|
|
#
?
Mar 21, 2023 20:58
|
|
|
cinci zoo sniper posted:This is like complaining that it's difficult to shoot a bird with a hammer. I asked Bard “is it difficult to shoot a bird with a hammer?” and it says quote:It is very difficult to shoot a bird with a hammer. A hammer is not a weapon and is not designed to be used to shoot birds. It is a tool that is used for a variety of purposes, such as driving nails and breaking things. If you try to shoot a bird with a hammer, you are likely to miss and injure the bird. You could also damage the hammer or yourself. It is important to use the right tool for the job and to use it safely. So there you have it. It’s difficult to do so, but also a hammer wouldn’t be able to kill a bird, and it’s hard to aim so you’ll miss and injure the bird.
|
|
#
?
Mar 21, 2023 21:14
|
|
|
Boris Galerkin posted:I asked Bard “is it difficult to shoot a bird with a hammer?” and it says 
|
|
#
?
Mar 21, 2023 21:30
|
|
|
Someone must have let it read the Aliexpress thread.
|
|
#
?
Mar 21, 2023 21:37
|
|
|
Playing with Bard for a few minutes, definitely not impressed. The Bing one has given me better results with fewer restrictions and feels more trustworthy with the way it sources. I have managed to get some better results with Bard on getting obscure movie and TV recommendations when being very prescriptive, though, so that's something.
|
|
#
?
Mar 21, 2023 21:45
|
|
|
Inferior Third Season posted:More worryingly, Bard thinks you might damage the hammer. Let's get Bard to talk about weapon durability
|
|
#
?
Mar 21, 2023 21:46
|
|
|
Electric Phantasm posted:Let's get Bard to talk about weapon durability Be sure to specify which edition you're playing.
|
|
#
?
Mar 21, 2023 21:52
|
|
|
The Bing one sucks for technical questions that you can’t find with Bing/Google trivially, it just repeats the top Bing answers slightly reworded (and patently wrong) then links to them. It’s worse than normal Bing because you need to wait for that stupid pretend typing bullshit and you get only one wrong answer or right answer to the wrong question). I like to eliminate an entire page of wrong or irrelevant links on a guess, thanks.
|
|
#
?
Mar 21, 2023 22:06
|
|
|
feedmyleg posted:Playing with Bard for a few minutes, definitely not impressed. The Bing one has given me better results with fewer restrictions and feels more trustworthy with the way it sources. I have managed to get some better results with Bard on getting obscure movie and TV recommendations when being very prescriptive, though, so that's something. The citations feature for Bing model is an underrated superpower. It makes various multitrack drifting manoeuvres like with “usefully wrong answers” much more tolerable.
|
|
#
?
Mar 21, 2023 22:07
|
|
|
roffles posted:https://twitter.com/saeeddicaprio/status/1638248160597364736?s=46&t=VqGDTpMBvtmBLtOHIUgZdw That's really embarrassing. This is from a 13B version of Alpaca running on my desktop.  edit: Llama will help you murder that bird 
LASER BEAM DREAM fucked around with this message at 02:46 on Mar 22, 2023 |
|
#
?
Mar 22, 2023 02:26
|
|
|
There's a fun one you can do that Bard fails miserably at and ChatGPT (sometimes) gets right: "If I have five bananas and you take away three apples, how many bananas do I have left?"
|
|
#
?
Mar 22, 2023 03:02
|
|
|
At least it's confident. 
|
|
#
?
Mar 22, 2023 03:13
|
|
|
 I mean at least the math in terms of numbers is right, I guess. E: 
Boris Galerkin fucked around with this message at 03:19 on Mar 22, 2023 |
|
#
?
Mar 22, 2023 03:16
|
|
|
Boris Galerkin posted:Google Bard is now in open[1] access. Because I pay for google one, I got sent an email to join, and then when I went to sign up it said "whoops you're not in a country that supports this" so get you're loving lines straight google.
|
|
#
?
Mar 22, 2023 03:39
|
|
|
If we ever cross the boundary of AI no longer being confidently incorrect about things that are trivially solvable, I may actually begin to worry. Seems we're a ways from that point.
|
|
#
?
Mar 22, 2023 04:33
|
|
|
PT6A posted:If we ever cross the boundary of AI no longer being confidently incorrect about things that are trivially solvable, I may actually begin to worry. Seems we're a ways from that point. Being confidently incorrect and ignore facts when corrected makes it far more like real people than something that is actually correct and/or capable of being corrected.
|
|
#
?
Mar 22, 2023 04:40
|
|
|
Tuxedo Gin posted:Being confidently incorrect and ignore facts when corrected makes it far more like real people than something that is actually correct and/or capable of being corrected. AI can successfully imitate a GOP party member... 
|
|
#
?
Mar 22, 2023 05:03
|
|
|
LASER BEAM DREAM posted:At least it's confident. Come Mr. TALL-E man, TALL-E me banana
|
|
#
?
Mar 22, 2023 05:22
|
|
|
Also funny thing that at this point it doesn't seem like there's much meaningful difference between the 'best' AI generators on the market and anything else, even if they're not all using each other to train.
|
|
#
?
Mar 22, 2023 05:32
|
|
|
Remulak posted:It’s worse than normal Bing because you need to wait for that stupid pretend typing bullshit
|
|
#
?
Mar 22, 2023 06:19
|
|
|
Ruffian Price posted:what's even funnier is it's not pretend, it's streaming the tokens in real time, that's really how long it takes to compute a simple sentence Yeah, that's the thing, still doesn't feel like this is near being able to scale anytime soon. I'm assuming Microsoft and Google are just taking losses on it at the moment until it gets to the point where it is actually profitable at scale.
|
|
#
?
Mar 22, 2023 06:22
|
|
|
Oh for sure. chatGPTs $20 a month is purely a trial to see who will pay. And when they dropped the API price it was just too stay competitive. chatGPT has had to limit tokens multiple times already, even for the paying people. On Tuesday the service was offline for most of it. They are hemorrhaging investment money.
|
|
#
?
Mar 22, 2023 08:48
|
|
|
dr_rat posted:Yeah, that's the thing, still doesn't feel like this is near being able to scale anytime soon. I'm assuming Microsoft and Google are just taking losses on it at the moment until it gets to the point where it is actually profitable at scale. Microsoft was reported rationing GPUs for machine learning teams eventually to make it possible to launch BIng preview while maintaining their Azure contracts. Proper scalability here will need hardware availability (in a more broad sense than Nvidia) improvements on the order of at least an order of magnitude. Mega Comrade posted:They are hemorrhaging investment money. OpenAI's current pricing on the enterprise APIs is a loss-leader, and it would need to be at least a few times more expensive on some of the more popular line items to just balance out the equipment/operating costs if I had to guess. In true commitment to an open and Socractic approach to AI, OpenAI is trying to lock in as much as possible, being first a heating-up market with a single-digit number of total possible participants.
|
|
#
?
Mar 22, 2023 11:19
|
|
|
Didn't Microsoft routinely do this in the 90s? Slash prices and burn through tons of cash just in an attempt to capture market and kill competition before raising prices after?
|
|
#
?
Mar 22, 2023 11:25
|
|
|
Mega Comrade posted:Didn't Microsoft routinely do this in the 90s? Slash prices and burn through tons of cash just in an attempt to capture market and kill competition before raising prices after? Absolutely. Basically every Microsoft product you can think of from the time that’s still around today did that in one way or another.
|
|
#
?
Mar 22, 2023 11:35
|
|
|
and Microsoft a company that probably can figure out how to scale it, and make it profitable if they can get it working reliably in their 365 suite. There's plenty of very generic documents and emails that people write up everyday in offices. If they can get it working right it can be a legit time saver, and that's something companies will pay for.
|
|
#
?
Mar 22, 2023 12:02
|
|
|
Wonder how much they will charge and make a profit,I doubt they can just slap it into their existing pricing model and be fine with capturing new customers. 365 already owns the majority of the market. Didn't Microsoft say every bing AI search costs 10x that of a traditional one? The tokens required for emails and documents are gonna be larger than most simple web searches.
|
|
#
?
Mar 22, 2023 12:13
|
|
|
Mega Comrade posted:Wonder how much they will charge and make a profit,I doubt they can just slap it into their existing pricing model and be fine with capturing new customers. 365 already owns the majority of the market. Microsoft is suggesting it'll be a paid addon to the existing 365 subscription costs: https://arstechnica.com/information-technology/2023/03/microsoft-365s-ai-powered-copilot-is-like-an-omniscient-version-of-clippy/
|
|
#
?
Mar 22, 2023 12:18
|
|
|
dr_rat posted:and Microsoft a company that probably can figure out how to scale it, and make it profitable if they can get it working reliably in their 365 suite. There's plenty of very generic documents and emails that people write up everyday in offices. If they can get it working right it can be a legit time saver, and that's something companies will pay for. Additional costs to running the 365 suite service is arguably not the thing they want to scale unless they have to.
|
|
#
?
Mar 22, 2023 12:27
|
|
|
|
| # ? Jun 1, 2024 15:07 |
|
|
Yeah, and that makes sense. With owning the majority of the market, if they can get it working as advertised I can see quite a bit of buy in. Like honestly surprisingly sensible approach from Microsoft. Googles approach seems more out of desperation and not wanting to be left behind. As has been mentioned using it for search is expensive, like unless they can somehow figure out a way to make it a hell of a lot more efficient -as well as accurate- I just don't see how they can make cash off it based off the minimal amount they get via ads from an individual search.
|
|
#
?
Mar 22, 2023 12:30
|
|