I’m a software engineer at a startup with impossible deadlines - I’ve used GPT4 for months to generate huge amounts app/server code, and much like your IDE, once you learn to use these tools you don’t want to go back to the days without it.
Speed
- Bard is very fast- similar to GPT 3.5 Turbo
- You need to multitask two GPT4 instances side by side to compensate for how slow GPT4 can be
Reliability
- Bard lies and makes up fake API calls more than GPT4
UI
- Bard UI is garbage - You have to keep manually scrolling down the chat window, and for some reason the largest button on the page is “stop” (???)
- You can tell Bard to modify its response to be longer/shorter and a few other options - I thought this would be useful, but it never ended up helping
Memory
- Bard has really short memory - Forgets details from last response!
- GPT4 memory is also unreliable, any details that are important you have to repeat
Intelligence
- GPT4 is objectively smarter
Internet Search
- GPT4 Internet search is garbage
- Bard has “Verify with Google” - I had high hopes for this, but never actually had a use for it
Willingness to give full code
- GPT4 is bad, but Bard is worse. Both need to be begged/threatened to return more than 100 lines of directly paste-able code.
Generating Useful Code
- Bard can give more concise medium complexity functions
Adding tougher features
- Bard hallucinates and lies
Dealing with lies
- When you tell GPT something doesn’t work, GPT will try something else
- When you tell Bard something doesn’t work, Bard will lie, claim to fix it, then give back the same code
Following Instructions
- GPT4 sometimes doesn’t follow instructions, but improving the prompt will fix that. Bard will happily ignore instructions, as clear as they may be.
Summary:
- GPT4 is still objectively better than Bard. Quite frankly, the prompts Bard couldn’t handle, GPT3.5 could.
- The cons of GPT can be worked around, but for Bard, it’s almost faster to do it yourself. Unless Bard was used like Copilot for short 1-2 lines of autocomplete, I wouldn’t trust it.
PS: If you’re not using AI yet for development, I highly recommend it - It’s like using an IDE instead of Notepad. AI can easily 2-3x your output, but you have to learn how it works so you can prompt it correctly, and you have be good at fixing its mistakes.
You should try copilot if you haven’t yet.
Copilot is great. I stopped paying for it after using ChatGPT so much (one subscription is expensive enough, as-is!) but I do miss it. Maybe I’ll buy it again…
Copilot regurgitates code verbatim and strips the licensing (see that one video of it spitting out Quake inverse sqrt). Don’t use it if you care about legality.
Quake inverse sqrt is one of the most famous algos in the world. Although I agree, if politicians actually stood for public interest they would force any AI derivative work to be SSPL or AGPL
Yes, its very famous. It’s also GPLed and Copilot had no right to use it in that way, and strip the license.
I’m working on convincing my boss to get licenses for the team lol. I just use a personal license but it saves a lot of typing and usually has a good idea of what I want.
I managed to convince mine, and it’s been great. I miss it at home when I’m doing personal projects
It’s honestly worth it tbh, even paying out of pocket. Our salaries are how much an hour? Save that and go home early 😄
I’m currently eyeing this PR to CodeGPT, a free plugin alternative to ChatGPT - This PR would add the autocomplete feature: https://github.com/carlrobertoh/CodeGPT/pull/333
CodeGPT lets you run a model locally, and my work computer has a GFX card strong enough to do that. The local models have gotten as good as GPT3.5, so needless to say the money-saving part of me is very excited about this!
Haha now I have to convince him to upgrade our laptops to gaming laptops too