Maybe they could use Bard to help them fix all of the bugs they’re loathe to ask humans to deal with? That was one of the under appreciated-mistakes which tanked new Google products: many of their most enthusiastic users signed up for Google Apps for Domains, and then Google blocked them from using and evangelizing everything they launched for the next decade.
I’ve seen another variation of this at work: we have two domains registered with G-Suite, Google professional services ghosted on a consolidation process, and this means that only one of those sees the licenses. They block all access to services when you’re logged in to the other domain, even just to view a third-party’s shared document unless you buy a second license for the same person. This has been a slow gravitational win for Office 365.
And once paying customers may find a workaround by using their personal Gmail, just make sure they really can't use it by telling them: "Bard isn’t currently supported in your country. Stay tuned!"
This is the reason why I will always recommend to not go with Google at my company.
I asked Google during an RFP when I can use my paying personal workspace for things like reminders or Google Home (and invite my family). It was "right the corner". The corner is now what, 5 or 7 years away.
This is a company I am with for a bit less than 20 years, but I will leave.
IDENTIFICATION DIVISION.
PROGRAM-ID. FIZZBUZZ.
DATA DIVISION.
WORKING-STORAGE SECTION.
01 COUNT PIC 9(02).
01 MESSAGE PIC X(08).
01 BLANK PIC X(01) VALUE SPACES.
PROCEDURE DIVISION.
INITIALIZE.
MOVE 1 TO COUNT.
PERFORM UNTIL COUNT > 100
EVALUATE COUNT
WHEN 3
MOVE "Fizz" TO MESSAGE
WHEN 5
MOVE "Buzz" TO MESSAGE
WHEN 3 AND 5
MOVE "FizzBuzz" TO MESSAGE
WHEN OTHER
MOVE COUNT TO MESSAGE
END-EVALUATE
DISPLAY MESSAGE
ADD 1 TO COUNT
END-PERFORM.
GOBACK.
This program prints the numbers from 1 to 100. However, for multiples of 3, it prints "Fizz" instead of the number. For multiples of 5, it prints "Buzz" instead of the number. And for multiples of both 3 and 5, it prints "FizzBuzz" instead of the number.
2. I don't know COBOL. Would be good if someone can check if the code is legit. I get a feel that it checks for the numbers to match (3, 5), not be divisible by 3 and 5.
It's obviously wrong, it's not using the modulo function anywhere or any such construct. It'll print text only for '3' and '5', and then numbers from then on.
By the way, GPT 4 when asked to critique this listing immediately found the problem, and produced a corrected version as a bonus.
Bard feels like GPT 2 to me. Even GPT 3.5 can solve COBOL FizzBuzz correctly.
PS: Just to check that GPT 4 is understanding the problem and not just regurgitating something it saw, I asked it to solve the problem in different ways, such as unrolling the loop or avoiding the use of the MOD function. It can do that too.
Yeah, that’s one of my favorite hidden features of GPT-4, if it gives me a function that doesn’t really fit my coding style I’ll tell it how I’d like it and it will adjust.
I got two methods in one response, first a walkthrough of installing and using the FPDF library, and then again for PyPDF2.
But, the code for FPDF was incorrect, and pypdf2 is so old that even the deprecation warning causes an error. I fed the FPDF error back in and the output was still wrong, but close enough that I could figure out an answer. Not faster than just googling for sample code though.
> Show me a clever hack for using deque instead of vector in C++
Ans:
// This function will reverse the order of a vector.
void reverse(std::vector<int>& vec) {
std::deque<int> deq;
for (int i = 0; i < vec.size(); i++) {
deq.push_back(vec[i]);
}
for (int i = 0; i < vec.size(); i++) {
vec[i] = deq.front();
deq.pop_front();
}
}
This is a clever hack because it avoids the need to write a custom function for reversing the order of a vector.
Doesn't this preserve the original order? The deque is used as a queue, first in first out. In order for the elements to be reversed it should be used as a stack.
Why on earth is there no link to whatever this "bard" actually is on that page - I need an on-ramp/slip road?
There is a link to "Learn More" which is also ... eventually a link wonders off somewhere ... nope some wankery about Lamda thingie. [random futile clicking]
It's the basics kids. No matter how far up your arse you are about your funky technology, do enable people to find the bloody thing easily, from all related pages. If you are a little rusty on spelunking then have a look at WAIS and Gopher (int al) for how to do it properly.
Obviously I can find it because I'm a fucking IT consultant. Why not make it easy for everyone?
Waitlists are magic at destroying any chance of word-of-mouth helping spread a product.
“Hey Bob, try out this new Bard thing!”
“I can’t, there’s a wait list.”
“Never mind then, you can just use ChatGPT instead.”
PS: it doesn’t fill me with confidence that Azure enabled OpenAI to scale to 10M users in record time, but a GCP can’t even scrape together enough compute to avoid a wait list on something comparable to LLMs hobbyists now run on their phones.
It appears that a large part of OpenAI's success is the large volume of human feedback reinforcement learning that they've used to tune their models. The more feedback they get, the better the model becomes.
It's a bit like Tesla having an advantage developing self-driving AI models by having hundreds of thousands of their cars on the road ("the fleet") gather data about rare events for training the model.
I know we all love GPT-4, and it totally deserves it, but we really need to support other companies building their own large language models even if it's Google. We don't want a monopoly situation.
I don't have access to Bard. I'm skeptical that's a valid approach. Language models are famously incapable of introspection. Does it change the answer if you tell it it's wrong, for example?
I got that message too. Classic Google trying their damndest to handicap the success of their own projects so they have an excuse to bin them in a couple quarters/years.
50 comments
[ 4.4 ms ] story [ 88.0 ms ] threadI’ve seen another variation of this at work: we have two domains registered with G-Suite, Google professional services ghosted on a consolidation process, and this means that only one of those sees the licenses. They block all access to services when you’re logged in to the other domain, even just to view a third-party’s shared document unless you buy a second license for the same person. This has been a slow gravitational win for Office 365.
I suspect this is because they can't use Workspace data for improving the models.
I asked Google during an RFP when I can use my paying personal workspace for things like reminders or Google Home (and invite my family). It was "right the corner". The corner is now what, 5 or 7 years away.
This is a company I am with for a bit less than 20 years, but I will leave.
A> Sure, here is the FizzBuzz problem in COBOL:
This program prints the numbers from 1 to 100. However, for multiples of 3, it prints "Fizz" instead of the number. For multiples of 5, it prints "Buzz" instead of the number. And for multiples of both 3 and 5, it prints "FizzBuzz" instead of the number.Here is an example of the output of the program:
Disclosures:1. I work for Google, but not on Bard.
2. I don't know COBOL. Would be good if someone can check if the code is legit. I get a feel that it checks for the numbers to match (3, 5), not be divisible by 3 and 5.
By the way, GPT 4 when asked to critique this listing immediately found the problem, and produced a corrected version as a bonus.
Bard feels like GPT 2 to me. Even GPT 3.5 can solve COBOL FizzBuzz correctly.
PS: Just to check that GPT 4 is understanding the problem and not just regurgitating something it saw, I asked it to solve the problem in different ways, such as unrolling the loop or avoiding the use of the MOD function. It can do that too.
What do you think they want?
Think about any inovative use of gpt and tell that in your form. Mine got approved really fast. Never implement nothing like what I proposed.
But, the code for FPDF was incorrect, and pypdf2 is so old that even the deprecation warning causes an error. I fed the FPDF error back in and the output was still wrong, but close enough that I could figure out an answer. Not faster than just googling for sample code though.
Ans:
But it did write a custom function?!There is a link to "Learn More" which is also ... eventually a link wonders off somewhere ... nope some wankery about Lamda thingie. [random futile clicking]
It's the basics kids. No matter how far up your arse you are about your funky technology, do enable people to find the bloody thing easily, from all related pages. If you are a little rusty on spelunking then have a look at WAIS and Gopher (int al) for how to do it properly.
Obviously I can find it because I'm a fucking IT consultant. Why not make it easy for everyone?
Nope, stuff to do.
Sundar: “one word: waitlist”
“Hey Bob, try out this new Bard thing!”
“I can’t, there’s a wait list.”
“Never mind then, you can just use ChatGPT instead.”
PS: it doesn’t fill me with confidence that Azure enabled OpenAI to scale to 10M users in record time, but a GCP can’t even scrape together enough compute to avoid a wait list on something comparable to LLMs hobbyists now run on their phones.
It's a bit like Tesla having an advantage developing self-driving AI models by having hundreds of thousands of their cars on the road ("the fleet") gather data about rare events for training the model.
What are the programming languages supported, for example?
Python, JavaScript, Java, C++, C, R, Ruby, PHP, Swift, Kotlin, SQL, HTML, CSS
C++
Go
Java
JavaScript
Python
TypeScript
Google Sheets functions
C
C#
R
Swift
Kotlin
PHP
HTML
CSS
SQL
Bash
Perl
Ruby
Lua
Rust
WTF
Italy disciplined OpenAI on GDPR concerns [1]. Google may not want the grief.
[1] https://www.bbc.com/news/technology-65139406
> Me: Write a golang function to get a http request, following the API of the get function in Python requests library.
> Bard: Sure, here is a Golang function to get an HTTP request with an API similar to the requests.get() function in Python:
It produced a function that takes url and parameters. Decent first approximation.
ChatGPT has trouble with this (when I tried a couple months back)