crazypython

Have you tried Phind? I'm curious how you find the two (Metaphor and Phind). I'm debating dropping Phind, i weirdly prefer raw ChatGPT. Curious how Metaphor is

rushingcreek•

on Sept 8, 2023

Phind co-founder here. May I ask why you’re considering dropping Phind and what we can do to improve?

unshavedyak•

on Sept 8, 2023

Two main things:

1. It sits in the "money for AI" fund from me. Ie it competes with ChatGPT. While it uses OpenAI in the background, i have not found it to be as complete and useful as ChatGPT directly. I did cancel ChatGPT for a couple months to try out Phind instead because Phind can search, and i find that missing from ChatGPT of course (at least until they have the module back). I'm on the pro plan of Phind, fwiw.

Hypothetically it could (maybe, i'm assuming) be as valuable for my ChatGPT use cases, but i found that i couldn't just give it the same input i'd give ChatGPT, as Phind would then search for an answer and come back with.. well, things from the internet which wasn't what i was wanting from that ChatGPT interface. I struggle to think of an example, though.. so hopefully this vague description is still of use hah.

2. Where Phind excels of course is digging answers out of the web, finding API solutions, wiring those results up into code, etc. However unfortunately this still seems a complex task. I do not find Phind to be accurate enough. It usually does a decent job at inferring the solution from the results, but it struggles with knowing library versions, context of the search results, etc.

I find myself using Phind primarily as a quick jumping off point for search results.. but i still often need to still search myself to figure out the actual code. So when i feel like i still need to search anyway, and that i feel i gave up ChatGPT for Phind (due to item #1), then it's just a tough position to justify Phind over CGPT for me.

For context usually i'm using and searching about Rust.

Fwiw i've still not decided one way or another. Though i'm definitely on the fence. Hope this vague and off the top of my head post helps a bit.

rushingcreek•

on Sept 8, 2023

Thanks for the detailed feedback.

We see Phind as a superset of ChatGPT. For the problems where search results may not be relevant, it should still function as GPT-4. Do you see it performing worse in your use case?

As for mistaken library versions, we’re working on developing our own index as a long term fix. But in the meantime you should be able to tell it the proper version and it should work.

So our hope is that Phind is a complete ChatGPT replacement for devs. Soon we’ll be rolling out our own models alongside GPT-4 that will bring additional speed and quality improvements.

I’d be happy to chat further with you directly, my email is michael(at)phind.com

unshavedyak•

on Sept 8, 2023

> We see Phind as a superset of ChatGPT. For the problems where search results may not be relevant, it should still function as GPT-4. Do you see it performing worse in your use case?

I did, but honestly i've not tried it in a while. I'll keep this in mind and try to give it a go again before i cancel/go back to CGPT.

> As for mistaken library versions, we’re working on developing our own index as a long term fix. But in the meantime you should be able to tell it the proper version and it should work.

That is true, i rarely give it my library versions. Though, i feel what's needed is frequently versions are ambiguous in results. In the case of something like Rust, i would take a search result, go to docs.rs/foo and see if those results exist in the lib. Usually that's the mismatched step, something i'm unsure if Phind is doing at all.

I'll try giving it versions more frequently, though.

Unrelated side note: i'd love if Phind had nothing user/pass auth or at the very least much longer valid logins. Using email on multiple devices is a bit annoying for me. For me (i use anonymous emails) it's all the downsides of auth, but plus email validation every time hah. Though this is not making me cancel or anything, i just had to login again and figured i'd vent my small annoyance :)

rushingcreek•

on Sept 8, 2023

Gotcha, would love to hear once you do try it again.

Does it really kick you out every time you visit? Login sessions should last for at least two weeks. Do you have any privacy settings that might be interfering with our cookie?

unshavedyak•

on Sept 8, 2023

Oh definitely not every time i visit, it's probably working as designed. It's just that across ~5 devices it feels frequent. Plus logging into other devices is made more difficult by this email-only option. Just my own personal gripe, it's probably unique to me haha.

swyx•

on Sept 10, 2023

you can also use https://github.com/smol-ai/GodMode to run chatgpt and phind side by side fyi

yett•

on Sept 8, 2023

I've usually clicked on the links that Phind provides as a source. Don't you find them useful?

unshavedyak•

on Sept 8, 2023

I do, in so far as it usually reveals where Phind failed for me. Ie "Why does it think that method should exist? Oh, this search result has an unrelated library using that method - and it got confused" sort of thing.

Off the top of my head i think if the link is correct, Phind is correct.

It's a tough problem though. I suspect the difficulty lies in a lot of search results not containing the appropriate versions/etc at all. So you have to go to secondary/etc steps of looking up library APIs and knowing that the two are incompat.

sharts•

on Sept 8, 2023

Crazy how much blank space both of these seem to waste on a wide screen.

fprotthetarball•

on Sept 8, 2023

Kagi allows you to provide custom CSS if you want to tweak it to your liking. I've used it to remove some of the widgets I don't personally find useful, and to replace some icons.

https://help.kagi.com/kagi/features/custom-css.html

Zetobal•

on Sept 8, 2023

Crazy that people have their browser in fullscreen if 90% of the pages use a max viewport of 1280...

webmobdev•

on Sept 8, 2023

A common use case for maximized browser window - tab-bar on the left- or right-side of the browser, and websites set to zoom to more than 125+ % to read it comfortably ... With higher resolution and bigger screen monitors everything seems to be getting tinier and tinier on screen. :)

seanw444•

on Sept 8, 2023

Yeah my browser and my terminal always share a 60/40 split on a basic 1080p monitor at work. I actually hate having it take full width now. Everything is condensed and is less strain on my eyes.

romanshrompsky•

on Sept 9, 2023

Kagi ftw I rarely use another search engine.

crazypython•

on Aug 4, 2023

•on: Non-determinism in GPT-4 is caused by Sparse MoE

The GPT-3.0 "davinci-instruct-beta" models have been returning non-deterministic logprobs as early as early 2021. This is speculation. CUDA itself often has nondeterminism bugs.

text-davinci-001 and text-davinci-002 were trained through FeedMe and SFT, while text-davinci-003 was RLHF; the models themselves have more variance at high temperature.

cubefox•

on Aug 5, 2023

What about the foundation models, i.e. davinci and code-davinci-002?

crazypython•

on June 6, 2023

•on: Nvidia releases new AI chip with 480GB CPU RAM, 96...

"™" has no legal meaning. "(R)" means a registered trademark.

HWR_14•

on June 6, 2023

That is not true. (TM) has a legal meaning. It's weaker than an (R), but it is still an enforceable trademark.

It's similar to creating a work covered by copyright vs. registering it with the copyright office.

bdowling•

on June 6, 2023

Affixing (TM) does not make something an enforceable trademark. Trademark rights are established by use of the mark in commerce. A (TM) is used to put others on notice that a word or design is being used as a trademark, which could potentially be evidence of the intent of an alleged infringer in a subsequent lawsuit. Intent is relevant because it affects damages and other remedies.

crazypython•

on June 3, 2023

•on: The NixOS Foundation’s Call to Action: S3 Costs Re...

S3 is a highly reliable storage media. You should use a CDN that transparently caches S3 artifacts to save on reads, such as AWS CloudFront or Cloudflare R2 are good choices.

crazypython•

on Dec 1, 2022

•on: ChatGPT: Optimizing language models for dialogue

How did you come up with the long-term memory idea?

gamegoblin•

on Dec 1, 2022

Seemed vaguely like how the brain does it. You think a thought, and somehow your brain conjures memories that are semantically related to that thought. That sounds a lot like a nearest neighbors algorithm on a language model embedding vector to me.

visarga•

on Dec 1, 2022

There is a paper from DeepMind

RETRO - Improving language models by retrieving from trillions of tokens (2021)

> With a 2 trillion token database, our Retrieval-Enhanced Transformer (Retro) obtains comparable performance to GPT-3 and Jurassic-1 on the Pile, despite using 25× fewer parameters.

https://www.deepmind.com/publications/improving-language-mod...

crazypython•

on Aug 23, 2022

•on: Stable Diffusion Public Release

Setting norms of asking folks not to abuse it reduces the likelihood people will step out of line with those norms.

crazypython•

on Aug 16, 2022

•on: Prometheus: Fuel from the Air

Definitely worth getting excited about, though it seems the current cost of fuel production as of Aug 8 including crude oil cost is $3.40/gal minus taxes and distribution.[0] Prometheus will likely cost less to distribute, since it doesn't need to be shipped across seas, funneled through pipelines, and between refineries, it can be produced next to solar or wind generation.

For reference, Prometheus costs $1.86/gal to operate and they are aiming to reduce the cost of the machine so it can be produced including capex at $3.00/gal.

[0]: https://www.energy.ca.gov/data-reports/energy-almanac/transp...