#Gemini speculation
flint remnant
15 MINUTES
flint remnant
It was here 42 minutes ago
modest ermine
@flint remnant make some concrete predictions about Gemini so when it releases we can reference this thread
my predictions:
- it will broadly match GPT-4 and surpass it in some areas but fall short in others
- pricing will be such that Gemini isn't really a significant upgrade over GPT-4
- GPT-4.5 will come out first and be significantly better than Gemini, but also significantly more expensive
modest ermine
my predictions:
1. it will broadly match GPT-4 and surpass it in some areas but ...
I'm sticking to my guns, screenshot this
flint remnant
my predictions:
1. it will broadly match GPT-4 and surpass it in some areas but ...
Honestly
I mostly agree
Not sure exactly how Gemini could fall short though but maybe
modest ermine
imo the smartest thing Google could do (though I'm not saying they will do this) would be to run Gemini at a serious monetary loss for a long time
just to play catch-up in terms of user base
flint remnant
The only thing is that apparently, they're aiming to compete with GPT-5???
I doubt
modest ermine
pffffffft
flint remnant
Severely doubt
But maybe Gemini 2 lmao
I don't think the Gemini structure itself will be abandoned for a little while
flint remnant
Idk
Maybe it could compete with a hypothetical 4.5 just because of its unique features
Kinda like how Claude is competing rn
Again
The only concrete reason to have faith in Gemini is Deepmind
frosty bone
what is gmmnmbmbinimni
jovial sorrel
imo the smartest thing Google could do (though I'm not saying they *will* do thi...
Goggle has monies to burn, I think they'll do this
I mean if you look at the economic equation, Google has lost a ton of brand equity. I believe the majority of the assets of almost all large companies are goodwill/intangible
modest ermine
RLHF data is a major asset too, and one that google is severely lacking
jovial sorrel
Coca-Cola's manufacturing plants, raw materials, bottled soda and syrup in warehouses, etc. isn't worth shit compared to their brand
And Goggel took a huge hit when they should have been synonymous with AI, search, data/life organization, etc.
frosty bone
eh google will do ok but not quite good
modest ermine
my predictions:
1. it will broadly match GPT-4 and surpass it in some areas but ...
making another prediction in addition to these ones:
\4. google will continue to not be the frontrunner because their organizational bloat prevents them from adapting rapidly like OpenAI can do
jovial sorrel
They need to stop huffing ethanol-distilled copium concentrate and use direct integration with very messy user data that they've supposedly been organizing and sorting for a decade or more
Despite the fact that their assistant seems to work less well each year and the only thing I find consistently useful is Google Maps
That's their real problem, that their promise of organizing the world's data has not quite gone sideways, but really does seem to be a mess
modest ermine
Despite the fact that their assistant seems to work less well each year and the ...
they're fucking that up too though. now I search "mediterranean" and the first 5 results are like, fast food places and gas station chains that paid to be at the top of EVERY search
jovial sorrel
they're fucking that up too though. now I search "mediterranean" and the first 5...
See I tried the single letter Bard hallucination trick and since I'm using my main account, it has my location, so it just asks me "are you looking for the weather in (where I live)" in five different ways
This is literally ELIZA levels of mentally well-regarded
modest ermine
See I tried the single letter Bard hallucination trick and since I'm using my ma...
can't wait until bard gets maps integration so that google can send me to a hallucinated hospital on a street that doesn't exist
flint remnant
can't wait until bard gets maps integration so that google can send me to a hall...
Lmaoooo
The problem isn't only rlhf
When you dont have good rlhf data
You can usually rely on grounding
For example
Bing
It has good grounding
Bard?
Has neither
flint remnant
See I tried the single letter Bard hallucination trick and since I'm using my ma...
LMAO
jovial sorrel
LMAO
What the hell, why are my results so hyperlocal
Okay, so something kind of fucking weird is that it gave me a GOOD (kind of) answer to a question about experimental treatments for anhedonia
It was deficient but at least accurate, probably on par or better than 3.5 sans web search
I wonder if it has good data but just isn't able to handle it well at all? That's a huge problem actually unless Gemini uses a totally different algo
I mean they MUST have enough industrial espionage capacity to just copy GPT
flint remnant
I wonder if it has good data but just isn't able to handle it well at all? That'...
Gemini is apparently built from the ground up
So it shouldn't have the same quirks that PaLM 2 has
Hopefully
modest ermine
Gemini speculation
flint remnant
Also, if Google is getting Deepmind into this, which for a while has been doing their own thing, then Gemini must be at least something notable
jovial sorrel
Gemini is apparently built from the ground up
That doesn't necessarily inspire confidence when they're still blasting out ads for Bard like it's a good thing
That makes me think they don't understand how non-experimental something like GPT is
GPT is experimental as in "not perfect" but this approach is fundamentally too stupid to ever be perfect. It's definitely ready and appropriate for enterprise or professional use among maybe the top 5-10% of tech-friendly users
Bard is more like top 1 or 2% can be trusted to use it
Total WAG numbers but way more people than most techies appreciate are catastrophically bad with tech. Look at all the people thinking GPT 3.5 writes good essays or can be used to game the stock market
crimson mango
Hello