Bard coding update with Colab exporting
blog.google
blog.google
Good that Bard is aware of latest version of software (eg. NextJS). Helps when APIs of a software evolves quickly.
I use a lot of Google products and really hope that they manage to nail it as at least #2 player in the space. As a company, they have not proven to come back and take the lead from behind.
I've found copying and pasting chunks of the API context into the prompt solves 99% of it. I'm now able to get it to generate modern workers just fine. You can copy an example method signature, some docs, etc. and it picks up on that fairly reliably. I've even just copy-and-pasted an entire worker implementation (100+ LoC) at the beginning of a prompt and that worked.
I have spent a lot of time reading through prompt engineering, but coding examples to adapt to latest versions are rare to find.
The big “breakthrough” for me was unlearning all of my search engine habits of thinking I need to be precise and talk to it like a machine. I use compliments like “that’s great” and it seems to reinforce good answers well enough.
Just treat it like a human on the other side of the screen and adjust as necessary.
Also pay the $20 - GPT-3.5 is not good, GPT-4 is junior engineer good. Most of my time is now spent describing problems to it and code reviewing what it does.
Edit: I do use prompt engineering when I’m building AI enabled products to help keep things on the rails. But for my personal use of ChatGPT I just let ‘er rip.
Have you tried and compared phind.com?
The complexity of the bugs you get are usually 10x as complex as the code's main logic you wrote. So you need to write very simple code.
Unfortunately, these GPT models are very complex. The bugs they produce will take an order of magnitude better model to fix.
Today's programming is not SICP in scheme, it is more plumbing between very large bodies of code. Using GPT is a great way to short cut needing to learn lots of details about code bases you only need for one specific feature. Recently though, by integrating two pieces of code I do know a lot about, I lost a lot of confidence in GPT coding.
For example, I asked it to implement a small script in emacs that would loop through org text files in emacs, and run some commands based on keybindings. That's it.
It messed up the usage of recursive-edit (which is a complicated feature itself), it produced many bugs in how Org utilizes indirect buffers, it sends windows around instead of buffers in the variables, and it wrote lots of contradictory code.
I understand how all these features work, and I went through and started fixing all the bugs. I would tell it about each bug, but it was producing so many subtle bugs with org-mode, inventing functions that didn't exist and it couldn't define... I gave up and just re-implemented it myself.
So no, I do not understand 100% of all code I refer to when I write code, that's why APIs actually exist.
This alone will help Google destroy the competition.
Bard can share with you any code that it is licensed to distribute (which includes GPL).
And if an architecture decision is so significant, retain a repository of architecture decision records.
// Credit Foo Project per MIT license
// See <https://github.com./foo/project/path/to/file
fn some_foo_fn() {
Microsoft and OpenAI got a lot right but also a lot wrong, I think. This could have held Copilot back, not making proper composite projects.
> I have a 3 dimensional black bitmap (voxel map) with a white straight line drawn in it. The 3d bitmap is saved as a bunch of tiff files by layer eg. layer_347.tif Write me python code to load this data and return the x,y,z location of each end of the line. The bitmap is too big to load all of it into RAM at once. Use a binary search to avoid loading unnecessary files.
Bard came up with a solution which was close but didn't follow the instructions (it just loaded all the data). It also assumed layer_9 sorted before layer_10.
GPT-4 came up with a solution which followed the instructions, but had a buggy binary search implementation, and a suboptimal O(n).
The GPT-4 solution was closer to what I was after, but there wasn't much in it. Neither implementation handled edge cases like missing files or had tests or anything.
Did you tell them to do those things after they came up with the solution?
```
// Basically still not handling edge cases
```
I don't see more drafts as a real differentiator.
Google has invested in custom AI hardware for some time now and does not run their workloads on nvidia cards
To do so, you need to split the matrix multiplies across the new machines. You also need more inter-machine network bandwidth, but with GPT-3 that works out to 48 kilobytes per token predicted collected from every processing node and given to every processing node. Even if Bard is 100x as big, that is still very doable within datacenter scale networking.
However, OpenAI doesn't seem to have done this - I suspect an individual request is simply routed to one of n machine clusters. As they scale up, they are just increasing n, which doesn't give any latency benefit for individual requests.
They are claiming to be the first to achieve >50% saturation during training. Pretty sure I recall Midjourney is using TPUv4 pods too
https://cloud.google.com/blog/products/ai-machine-learning/g...
https://cloud.google.com/tpu/docs/system-architecture-tpu-vm
Well the person I replied to did..
Also, chatGPT always goes the extra mile and includes instructions on compiling/running the code it provides (including sample data). Bard does not.
Bard still has a ways to go to catch up.
Haven't bothered with Bard at all.
Yeah, given the random nature of these LLMs it's possible that I won the lottery in that particular case.
I have a suspicion there might be some... showmanship at work where all kinds of test problems and common tasks are made sure to work well.
Also it generates code slower than I would type (if I know what to write), which makes me mad for some reason.
Because its a much smaller model IMO. Its like comparing a car with 2.0 engine, with a car with 3.0 engine. Also ChatGPT made a lot of errors in code just a few months ago. Google seems to be catching up.
There's almost no way they make a 30T model that's useful at least? That's over 60TB of memory at 16bit, 15TB at 4 bit. That's 15TB that have to go through either VRAM or TPU Memory (TPUv4 is 32GB I think?). Yes you could distribute it and chunk it, but that would be insanely slow/expensive to run for anything more then a research demo.
GPT-4 is most likely 1~2T from what I've heard, and they've been struggling to run that profitably/quickly judging by the crazy pricing and 25messages/3hours if you pay for premium chatgpt.
EDIT: The only way this is possible is if they've pulled off some crazy optimization, which is definitely possible, but a lot of the research from them such as their own Chinchilla scaling suggest smaller models that are more fully trained are better.
According to their chinchilla paper they would need over 330TB of training data for that, and an absurd amount of training time.
So it depends heavily on the task and how much it has stored from the training.
Now if you ask them to do a teapot, it sometimes do it since there are so many examples of that, even if you often have to tinker a bit to get it to compile and run, at least it is a teapot and not some weird random geometrical object.
I have yet to try it with bard, will test when I get home to a proper computer.
I think there will be a settling in period over the next year where providers will take the time fix problems. I have been playing around with AutoGPT on my laptop and that is a very different experience because it exhaustively searches for useful material on the web, saves it to local text files, then works with these local files to finish up preparing text for the user. I don't expect Bard and Bing+ChatGPT to be so explorative and iterative, but they might be when inference technology and specialized hardware architectures get better.
Twitter post coming soon, I will share with HN when it’s live
> Can you write a <HTML> file that POSTs a form with firstName, lastName to some other URL?
> I'm a language model and don't have the capacity to help with that.
To send a POST HTML request with firstName and lastName as inputs to a URL, you can use the following steps:
1. Create an HTML form with two input fields, one for firstName and one for lastName. 2. Set the method of the form to "POST". 3. Set the action of the form to the URL you want to send the request to. 4. Submit the form.
The following is an example of an HTML form that you can use to send a POST request with firstName and lastName as inputs:
```html <form method="POST" action="https://example.com/api/v1/users"> <input type="text" name="firstName" /> <input type="text" name="lastName" /> <input type="submit" value="Submit" /> </form> ```
When you submit this form, the server will receive the firstName and lastName values as POST request parameters. You can then use these values to perform whatever actions you need to do on the server.
Could be making it up but need to test further...
E.g, does SendInput send to a background window? It told me not only does it but it is the only way to do it and then gave me examples where it got a handle to a window but then didn't even use it.
A few months ago this would be super impressive but after using ChatGPT there is no reason that I've found why I'd want to use an inferior AI.
Same question with a different answer.
This is interesting. This version of the question is better than the original answer I got but still not as good as your answer.
I'm not sure if this is what is planned of course, but that would be far more widely used than an otherwise siloed website you have to visit specifically.
Also remember that most people have at least never used ChatGPT, and most probably still haven't heard of it. We live in a tech bubble – none of my family had heard of it a few weeks ago. They all use Google search on a daily basis.