Show HN: Web search using a ChatGPT-like model that can cite its sources (beta.sayhello.so)

Almost every other search engine is closed source.

Just worked for me: https://beta.sayhello.so/search?q=+fingerings+for+a+c+trumpe...

Unfortunately searches can be non-deterministic, as the links Bing returns to us can change from query to query even if search wording stays constant.

Metaphor is cool, but it's a different product. They use language models to generate links, while we use them to generate answers.

We were temporarily overloaded. Fixed now.

So happy to hear that you've been using Hello already. We've been picking up the pace of improving our models.

Should be fixed.

You're completely right. We have a much better one that we'll be switching to soon.

> how do i mix two colors (hexadecimal rgb string) in js

> ANSWER

> To mix two colors in JavaScript, you can use the color-mix property. This property takes three parameters: the colorspace, the color to be mixed, and the percentage of that color to mix. For example, to mix the colors #FFFFFF and #8FBC8, you would use the following code:

> let newColor = document.createElement;

> // Outputs: #8FBC8

This is all just wrong from beginning to end. It seems to be confusing JS and CSS, too.

They did a Launch HN recently: Launch HN: Hello (YC S22) – A search engine for developers'[1]

Nothing much has changed since. So they appear to be trying to cash in on the interest in ChatGPT.

Interesting they didn't include that they are backed by Y Combinator in the recent S23 cohort. Is being backed by YC a negative for startups here now?

[1] https://news.ycombinator.com/item?id=32003215

Yes, the answer I got was nonsensical too:

> what is the diameter of the sun in sun radii

> ANSWER

> The diameter of the sun in solar radii is 1,392,000 km. This is calculated by taking the mean radius of the sun, which is 432,450 miles , and multiplying it by the circumference of the sun, which is 2,715,396 miles .

The correct answer is, of course: 2.

BTW: Google is also confused, but not in the same way, returning:

2.0018 R(sun)

Well, that's what these language models do in a nutshell. If you think social media and blogs are bad in terms of misinformation, this is the same thing rehashed by a machine that can not actually think or judge for itself (which to be fair, is arguably almost on a level already with how some humans interact with information they consume online).

I asked some controversial questions and the answers are interesting to say the least. The source is often reddit, random sites that are probably the top search engine result or some yellow press outlet.

To the "Why did Elon Musk buy Twitter" question it says things like:

"Additionally, he was offered a high enough price by the shareholders that they were willing to sell their shares to him."

or:

"He acquired Twitter in April 2016 for $44 billion."

what is the path integral of a european swallow

ANSWER The path integral of a European swallow is the sum, or functional integral, over an infinity of quantum-mechanically possible trajectories to compute a quantum amplitude. This formulation has proven crucial to the understanding of how birds fly and migrate. For example, during migration, barn swallows cover 200 miles a day at speeds of 17-22 miles per hour, with a maximum flight speed of 35 mph. In their wintering areas, they feed in small flocks which join together to form roosting flocks of thousands of birds.

The links are just regular Bing results.

This latest model was trained on a distributed multi-node cluster of 8 servers each with 8 NVIDIA A100 GPUs. Need a lot of GPU memory when training with long sequence length.

As a follow up to that question, which I’m also very curious about.

Do we have rough estimates of the training GPU compute needed to train GPT-3 and has OpenAI said anything about compute resources used to train ChatGPT?

We're always actively exploring different prompting - a general question answering prompt most definitely won't work for all searches. Finding what people agree on and disagree about in e.g a reddit thread could be very useful.

I interviewed an author of WebGPT here, OpenAI co-founder John Schulman if you're interested, and he talks about it https://www.talkrl.com/episodes/john-schulman

Thanks, we're getting 10+ requests per second at the moment. Frantically scaling to keep up. Didn't expect this post to blow up so fast!

>I would like a recipe for cookies"

This has been doable on Google for years. Maybe not the same kind of intent understanding but enough keywords are there that you would end up getting lots of results for cookie recipes.

yep, it's super interesting to watch how cheap and plentiful compute starts to change the way we live

> "Please take me to my email"

Results in

"To take you to your email, you will need to sign in to your account. Once you are signed in, open your inbox to check your mail. You can also use the Help tab to contact customer support or see training videos. Additionally, you can right-click an email in the Message List to show more message options and right-click the Reading Pane for additional email commands such as translation. Finally, you can use Search to find other commands or search through your email, contacts, and calendar."

Citing resources from google and microsoft support.

It's much faster now :)

It is faster than ChatGPT normally. We're frantically scaling our infra right now.

The p50 time is around 2s end to end not under load. Bear with us as we scale our systems!

  INPUT X
  IF X > 10 THEN
      PRINT "X is greater than 10."
  ELSE
      IF X < 10 THEN
          PRINT "X is less than 10."
      ELSE
          PRINT "X is equal to 10."
      END IF
  END IF

Great! All the best scaling :) I'll give it a shot again in a bit.

It looks like you’ve stumbled upon the next big “impossible question” that is going to require future technology to help us solve.

It is interesting to see how it attempted to “solve it” though (I put that in quotes since it isn’t actually doing the math)

Thinking more on this... I don't think any of these sites will live if they get big enough. And if enough of them pop up it'll draw tons of attention from content sites.

If you want to show that data you'll end up having to work out a license from StackOverflow. Possible, but far more difficult than the current ease of plug-and-play GPT drop-in.

Do we really think Google hasn't thought of this exact thing already?

But it's siting the sources, how is it a copyright violation?

You're confused, copyright != antitrust violations.

Both sources you provide have zero mentions of the word 'copyright' in them.

Those lawsuits have to do with Google's dominating the search market and using that to their advantage in ways that are allegedly unfair.

Copyright law actually allows a service like Google to exist in the first place.

The reason is likely simpler:

- It is expensive (~0.5c per generated answer)

- It is (currently) slow (2-3 seconds to result)

- It is hard to place ads inside direct answers (probably the most important)

What's funny is that most of this ground breaking LLMs you see now are based on Google published research about transformers, and they have better performing models in house than anything publicly available on the market.

> Pulling the meat of the content from a site like StackOverflow ends up as a copyright/anti-trust violation.

Then how did ChatGPT do it?

Note that pulling the meat of the content from StackOverflow isn't copyright violation though, as long as you follow the license (which is Creative Commons something-something but probably fine for this particular application).

I can't find the citation, but someone related to OpenAI said that ChatGpt cost a few cents per query.

Anecdotally, I looked at running an open-source GPT competitor from Meta (OPT), and the requirements for running that are vaguely on the scope of "8 GPUs with at least 400gb of GPU memory" which would be VERY expensive, even in the cloud.

An 8gpu AWS EC2 instance costs $3.40/hr (p2.8xlarge), and if the model runs for 20s per request (roughly how long chatGPT takes me to get a response), that gives you $0.02 a request, not including things like storage or idle capacity. So a request could likely be $0.05-0.1, not including overhead like dev costs, training, etc.

The answers are explicitly based on information found on the web instead of some knowledge found within the model parameters.

Thanks for trying it out - we're still quite early so the model isn't going to be perfect + we're focused on programming related queries at the moment. What is your use case? Are most of your searches related to history?

There's zero attempt to cache answers or cited sources, I suppose, because I keep getting different answers even from the same cited source just by clicking the search icon on the bar with the same query:

"The French-speaking population of France during the French Revolution was estimated to be around 25 million people" (wrong, citing that same Statista graph that doesn't mention language use)

"It is difficult to accurately estimate the French-speaking population of France during the French Revolution, as there was no official census taken at the time." (honestly the most accurate answer yet, but citing that same Statista graph)

"It is impossible to accurately estimate the French-speaking population of France during the French Revolution, as there is no reliable data available. However, it is estimated that around 80% of the population spoke French in 1800." (first sentence is correct, second is wrong, citing that same Statista graph)

"At the time of the French Revolution, less than half of France's population could speak French. This was due to a combination of factors, including the fact that French was an extremely popular language with the elite and higher classes, as it was adopted by nearly all European courts and it even reached the other side of the Atlantic." (cites a babbel.com article; I think the first sentence is the first time the answer approaches both accuracy and consistency with both the question and source, largely by plagiarizing part of the source. It still omits useful context about the lack of literacy/fluency that's in the same sentence lifted from the source. Then it goes back off the deep end with the second sentence, which the source uses to emphasize how broadly French was spoken, and not as a reason why so few people in revolutionary France spoke French.)