Bard is a user-facing product, Gemini is an internal/developer facing model.
Bard is now backed by Gemini and its various forms, but used to be backed by PaLM/2. Similarly, ChatGPT is a user-facing product, and GPT3, 3.5, 4, etc, are the developer facing models. OpenAI expose a bit more of the GPT branding to users allowing them to directly select the model in ChatGPT, but Google hasn't done this, I would guess for product reasons.
The main distinction is that Bard/ChatGPT are fine-tuned to be conversational, whereas the raw models are not necessarily. For example, in Google Workspace you can have an LLM draft an email for you, but this is not Bard, it's not conversational, it doesn't chat back and forth, it takes a prompt and generates an email, which is what it was fine-tuned to do. Bard/ChatGPT can write emails, but they aren't explicitly trained to do so and will discuss the email with you more.
The Nanny model stuff has basically eliminated gpt-4-1106-preview from my stable of go-to models, which is a shame because it’s pretty good (not the undisputed best, that’s marketing).
Today I wanted to drop some generated files into a Bazel workspace (just like Gazelle and stuff do all the time) and it was just like “builds are hermitic for a reason…” and after like my 3rd time trying to force it I was like “I’m arguing with a toaster, it’s come to this.”
M-x dolphin-region-task
“Write an Eminem style diss track about how much this sucks and then tell me the magic spell for the Starlark thing”.
Every time I try Bard and the "new" Gemini it hallucinates so bad I get scared to try it again. Its circumstantial and I only use it for tech stuff and each time its immediately returned a very believable parameter that makes me laugh to myself at how stupid I was for not knowing it. I only use GPT4 so maybe 3.5 does the same thing.
In my experience, Llama2, mixtral, bard/gemini hallucinate so much more than gpt3.5 that they are unusable.
For trivial things and small examples you could be fooled. But large scale when you need somewhat accurate results, they aren't great
For example stuff like manual classification/feedback manual "testing" various release candidates (more for which feels better then which is better, benchmarks are likely automatized)
or stuff like "filter out all offensive text/pictures", which could include some pretty bad stuff
rate this X answers
are this filtered answers which where filtered out as inappropriate actually inappropriate?
> Our search quality raters provide us with insights and evaluate pages against our guidelines to help make sure our systems — and proposed improvements — are working as intended.
> What that looks like in practice is often a “side-by-side” test where a rater will look at two sets of Search results, one from the current version of Google and the other from an improvement we’re testing. Raters will review the pages in each set of results, and evaluate if the pages are a helpful match for the query based on our rater guidelines.
They have armies of workers from low wage economies manually flagging issues with AI models, and moderating content now. It's not as magical as it all looks...
This is unremarkable. Companies use contract labour for the express reason that they can dip in and dip out. Sounds like Appen were on a gravy train and didn't work hard enough to diversify their portfolio of clients.
I agree, it looks like they will be back at the numbers from 2019. Which is of course a stepback, but I imagine they will learn something from it. Overall they are still growing, for example, in 2015 they had "only" $43.83 Mil in revenue.
Joking aside, in this situation the contract is canceled in the sense that the plumber is not given an extension to fix anything else in that person's house. He or she must find other work.
> Fast Company wrote last year that some Appen employees who are members of the Alphabet Workers Union had been petitioning Appen to increase wages from $10 an hour to $15. While the union won wage increases, the final number fell short of its goal. Many of these workers were then laid off, with Appen citing business conditions.
"Human workers at companies like Appen often handle many of the more distasteful parts of training AI and are often the lower-paid, often ignored backbone of the entire industry. At Appen, contractors help rate data quality and answers from AI models. Fast Company wrote last year that some Appen employees who are members of the Alphabet Workers Union had been petitioning Appen to increase wages from $10 an hour to $15."
Notably not Accenture, which is where the conflict over whether Google is coemployer for unionization purposes is happening. Previous HN discussion here - https://news.ycombinator.com/item?id=38873878
35 comments
[ 4.0 ms ] story [ 45.4 ms ] threadThis feel likes a 'good enough' vs 'premium' battle at the moment.
You'll probably get better results with pay-for-play chatGPT ... but if Bard/Gemini is 'good enough', why bother?
Jury's still out.
Bard is now backed by Gemini and its various forms, but used to be backed by PaLM/2. Similarly, ChatGPT is a user-facing product, and GPT3, 3.5, 4, etc, are the developer facing models. OpenAI expose a bit more of the GPT branding to users allowing them to directly select the model in ChatGPT, but Google hasn't done this, I would guess for product reasons.
The main distinction is that Bard/ChatGPT are fine-tuned to be conversational, whereas the raw models are not necessarily. For example, in Google Workspace you can have an LLM draft an email for you, but this is not Bard, it's not conversational, it doesn't chat back and forth, it takes a prompt and generates an email, which is what it was fine-tuned to do. Bard/ChatGPT can write emails, but they aren't explicitly trained to do so and will discuss the email with you more.
Today I wanted to drop some generated files into a Bazel workspace (just like Gazelle and stuff do all the time) and it was just like “builds are hermitic for a reason…” and after like my 3rd time trying to force it I was like “I’m arguing with a toaster, it’s come to this.”
M-x dolphin-region-task
“Write an Eminem style diss track about how much this sucks and then tell me the magic spell for the Starlark thing”.
Yeah, that was better.
Does anyone have an example on what these people could have been doing for an $10/h?
- Graphic violence - Child abuse material - Non-consentual pornography - Fetish/misc. offensive material
And I'm guessing those employees don't get mental healthcare benefits.
or stuff like "filter out all offensive text/pictures", which could include some pretty bad stuff
rate this X answers
are this filtered answers which where filtered out as inappropriate actually inappropriate?
etc. etc.
> Our search quality raters provide us with insights and evaluate pages against our guidelines to help make sure our systems — and proposed improvements — are working as intended. > What that looks like in practice is often a “side-by-side” test where a rater will look at two sets of Search results, one from the current version of Google and the other from an improvement we’re testing. Raters will review the pages in each set of results, and evaluate if the pages are a helpful match for the query based on our rater guidelines.
https://static.googleusercontent.com/media/guidelines.raterh...
Joking aside, in this situation the contract is canceled in the sense that the plumber is not given an extension to fix anything else in that person's house. He or she must find other work.
"business conditions"