666 comments

[ 4.7 ms ] story [ 368 ms ] thread
(comment deleted)
Overall very underwhelming... not even really a compelling landing page. Can't write code, can't understand any language besides english. If this is really the best they've got after months of 'code red' to catch up, they really seem to be in trouble.
I imagine it's due to the urgency with which they been trying to push Bard out the door. RLHF seems to be a key component in making an actually useful chat AI, so they want to start collecting that data ASAP. Management is probably breathing down that team's necks too, which also doesn't help.
Hmm. This is not correct. Bard can write code.
In the FAQ[1] it explicitly states that it cannot help you code. Maybe that's a limit of the prompt or just a limit of Google using a smaller/cheaper model? We'll have to test it out to see.

[1]: https://bard.google.com/faq

It looks like they have a filter, but it's pretty easy to get around depending on the language, e.g. I can get it to generate C++ code.
Bard faq says writing is not supported and that bard is still learning. I left with the impression that it can but maybe not well enough to have them even attempt to promote it as beta.
I tried and it was pretty awful.

This is the prompt I use, for which ChatGPT outputs a perfect program [0] on the first try (and can even convert it to use material-ui, with some prodding):

> Write a react program that I can paste into a codesandbox. It should display a counter that increases by one every second. Below the counter is a button, and when I press it, it decreases the counter by two. If pressing the button would result in the counter decreasing below zero, the button should be disabled.

Bard refuses to attempt it. It did attempt it for one variation of the prompt, but it used class components with an invalid render function and no timer; it looked like it was just pasting something from the first StackOverflow answer from a Google search of the prompt.

To be fair, Google isn't claiming to support this yet. But I'm just replying to the comment saying it can write code.

[0] https://codesandbox.io/s/prod-haze-c28sec

Can it? This was the response I got: "I'm designed solely to process and generate text, so I'm unable to assist you with that."
It can write code, but doesn’t seem very good at it and often refuses.
Anyone know what is the origin of the name Bard?

Every time I see it I think "barf"

According to ChatGPT ;)

  Where does the word bard come from?
> The word "bard" comes from the Scottish Gaelic word "bard", which means a poet or storyteller. The term was also used in medieval Ireland and Wales to refer to poets and musicians who were highly respected members of their society. In ancient times, bards were considered to be the keepers of history and tradition, and their songs and stories were often used to pass down knowledge and wisdom from one generation to the next. The word "bard" is also related to the Old Irish word "bárd", which means a poet or singer.
And this nicely shows how LLMs don't do a good job yet of answering questions accurately. It's Scottish. No wait, it's Irish. Oh, wait it was used in Wales. Read the Wikipedia article to see where it got this mish-mash from.
I just got access to Bard, and this is what Bard says.

  Where does the word bard come from?
> The word "bard" comes from the Old Celtic word "bardo-," which means "poet" or "singer." It is thought to be derived from the Proto-Indo-European root "gʷerH-," which means "to praise." The word was originally used to refer to the professional poets and singers of the Celtic cultures, who were responsible for composing and reciting oral traditions, such as epics, genealogies, and songs. The word "bard" came to be used more generally to refer to any poet or singer in the 16th century.

Draft 2:

> The word bard comes from the Old Celtic word "bardos" which means "poet-singer". It was first used in English in the 16th century. The Old Celtic word "bardos" is thought to be derived from the Proto-Indo-European word "gʷerH-" which means "to praise".

Draft 3 (Including a source: Wikipedia):

> The word bard comes from the Celtic languages: Gaulish: bardo- ('bard, poet'), Middle Irish: bard and Scottish Gaelic: bàrd ('bard, poet'), Middle Welsh: bardd ('singer, poet'), Middle Breton: barz ('minstrel'), Old Cornish: barth ('jester'). The ancient Gaulish *bardos is attested as bardus (sing.) in Latin and as bárdoi (plur.) in Ancient Greek.

The word bard originally referred to a Celtic poet-singer, but it has come to mean more generally a poet or singer. It is especially associated with the early Celtic cultures of Ireland and Wales, where the bards were highly respected members of society. They were responsible for composing and reciting poems, songs, and stories, and they often played an important role in preserving the oral traditions of their people.

The word bard is still used today to refer to a poet or singer, and it is also used in the names of several fictional characters, such as the Bard of Avon (William Shakespeare) and the Bard of the Highlands (Robbie Burns).

Sources - Learn more en.wikipedia.org/wiki/Bard

I just can't imagine a less inspiring name. It feels like Bard was a joke, internal name. When OpenAI came out with their stuff, Google decided to rush this to market with little thought given to the branding.
It's not like "ChatGPT" is a particularly good name...
Like Microsoft's Copilot? It may be better than putting a specific use case as name.
(comment deleted)
In the US and UK currently. Title should be modified IMHO
use a vpn?
eh don’t need, it’s google so we know it sucks compared to GPT.
We know raw LaMDA sucks compared to GPT, but we also know that a lot of the performance of GPT3.5, GPT4 and ChatGPT comes from the reinforcement learning added on top of raw GPT, so it will be interesting how Bard performs.
LLaMA is not google
He meant to say LaMDA, which is a family of Google's conversation tuned language models, a smaller of which is used in Bard.
right, corrected my post. These models are named too similarly
Modified. Thanks!
I appreciate the honesty about problems with Bard and LLM in general
I'm still waiting for Hacker News to band together and simply start building an open model that does not require a corporate daddy like OpenAI, Microsoft, Baidu, Google, or anyone else. These models have no place being owned by corporations. The services that use these models are certainly up for grabs, but the models should be open. Who is trying to rally the community to this cause?
HN seldom bands together to do anything but quip at things, or discuss how it should have been done better. Oh, wait.
More exciting software projects get started on 4chan than Hacker News.
Most people around here have a job.
Or want to have/generate one.

What I mean is, HN is a place where people keep good ideas to themselves in the hopes of launching a business on its back. It's a hustling/entepreneurship community, not a "let's build stuff" one, like, say, the Open Source communities. Collaboration has an opportunity cost around these parts.

I'll take "exciting" in technical terms rather than "exciting" in legally-challenging, raging mob ones...
(comment deleted)
This isn’t folding@home. These models need as near to synchronous parameter updates as possible. You need a real facility to do this.
Yeah but we could definitely crowdsource enough money to make an extremely effective finetuned version of LLaMA. The challenge there is probably more about finding and cleaning data.
Didn’t they spend hundreds of millions running the training data or something? How are some open-source hackers going to do it?
There are at least two completely open models comparably scaled to GPT-3, and at least one open distributed-hosting project.
Check out Llama, Alpaca, Langchain, Dalai.
(comment deleted)
(comment deleted)
Fully agree, we need a competitive (even if slightly behind) truly open initiative. The potential (and current) censorship of these models is an oligarch’s wet dream.

With google vs DuckDuckGo etc you can see what websites are returned, and get a sense of the kind of authoritative sources used

With this kind of tech, there’s no real way to know except by comparing precise prompts with one another. But if they’re all behind corporate walls, and they share incentives to say similar things, we’re none the wiser

Bing Chat provides sources as in its response. I think that is not perfect, but directionally correct in terms of identifying what web results were used as authoritative sources.
Have you set up the GitHub repository and done the initial commit yet? If you think “we” should take the first step, then go-ahead, take it. Lead the way.

I am being semi-snarky because often people talk about some grand ideas about things “we” should do, but rarely do those advocating for some utopian outcome actually do anything towards achieving it.

If it is to be, it’s up to me.

Considering that the CEO of OpenAI is the former president of Y Colbinator…
(comment deleted)
Looking forward to a new API to build with.
>It’s grounded in Google's understanding of quality information.

I almost spit my coffee reading that, seeing what Google's search results have become.

It looks like some journalists already have access. If folks see any takes on how this product compares to GPT4 on quality of responses, please share.

Maybe there's an embargo? Because normally if reporters had access, they'd immediately post results to twitter.

Here's one take: https://twitter.com/jjvincent/status/1638182196283609092

- it's faster (though presumably that's because of limited users)

- unlike Bing, it doesn't cite sources (unless it directly quotes them)

- it seems relatively constrained, giving answers that are anodyne and often dull

(comment deleted)
we’ve also built in guardrails, like capping the number of exchanges in a dialogue, to try to keep interactions helpful and on topic

So it's simultaneously more neutered and less stable than what OpenAI already offers, in addition to being years behind. What exactly is the benefit to being an early adopter of this when there are better alternatives?

nothing....Sundar won't get fired...oh that's a negative...hmmm...dunno
Not sure, i could see it with Google Mail, but with this one i'm not sure since the alternatives seem to be better. Just having the name Google won't make it a better product.
They're behind their competitor and they're already hobbling themselves. Their "Big brother knows best" attitude isnt going to help them much here. Its like they're stuck in the last decade.
Some (serious) competiton in this space would be very welcome. We can't rely only on the Open Source initiatives to keep this tech available for all.
I just tried accessing it. When you visit bard.google.com on Firefox it just gives you a generic intro after logging in. I was a little confused and logged in with Chrome, where I now saw a "join waitlist" button. Guess they don't want Firefox user using it.

Overall this just seems underwhelming. Most HN comment's I've read are people saying it doesn't work with their account, doesn't work in their country, doesn't work for some other reason, or works and is very limited compared to ChatGPT.

Google what have you been up to?

I'm on Firefox and I could join the waitlist.
(comment deleted)
> You’ll receive another email when it’s your turn to test Bard.

> In the meantime, we asked Bard to write you a little poem while you wait.

> May your day be bright, Your mood be light, And your heart be filled with delight. ~ Bard Thank you for registering.

Cute

Would be better if they sent everyone a different poem. I got the same one as you.
I got the same poem, so off to a good start on the vapourware front.
I'm envisioning an office in Bangladesh full of people copying Bard prompts into ChatGPT and pasting the results.
Google spent years touting its leadership position in AI with DeepMind and Google Brain

now its playing catch up and seems to be flailing around a bit

Pichai really is just looking like a good-times placeholder CEO

Are you discounting the work on AlphaGo, AlphaFold, etc.?

Google may be playing "catch up" in releasing these things as products, but remember that ChatGPT, etc. are all based on Google's paper on Transformers ("Attention is all you need").

How do you account for the apparent discrepancy between research and product? It's not like Google is low on resources when it comes to engineers or product managers. If ChatGPT owes so much to Google's research, then why is their product so far ahead?
>> Race to Create A.I. Chatbots

I am glad finally we know what these LLMs are made for.

(comment deleted)
$today + 1 year: "Google shuts down Bard, its AI chatbot"
I laughed; but, this is truly a sad state of the trust in them. :(
(comment deleted)
Just asked chatgpt to write it.

Google recently announced that they will be shutting down Google Bard, their new AI-powered writing tool, after less than six months since its launch. The tool, which used natural language processing to help users write poetry and song lyrics, was met with mixed reviews and failed to gain significant traction among users. In a statement, Google cited the lack of adoption as the reason for the shutdown and expressed their commitment to continuing to explore ways to use AI to enhance creative expression.

Nah - 2 years, because that's the amount of time you can juice a new project on a promo packet.
ChatGPT chose the timeline. haha
Please don't post shallow, dismissive clichés. They're repetitive and therefore the opposite of what we want here.

You may not owe faceless BigCorps better but you owe this community better if you're participating in it.

https://news.ycombinator.com/newsguidelines.html

The parent comment just reached legendary status:

https://twitter.com/juanbuis/status/1638289186351456257

Hall of fame level here. This is hilarious.

Put me in the screenshot, I mean the hallucinated Bard response.
It would be even more hilarious if mod scoldings made it rate the parent higher. But the text it quoted was from https://news.ycombinator.com/item?id=35247109 - maybe it treats subthreads as a single thing?

It's pretty funny that ChatGPT trolled Bard.

Can you elaborate on what you found wrong with the above comment?
Reflexive "Google will just shut it down" reactions were already a cliché 10 years ago. It's been tedious for a long time, and therefore is not driven by curiosity, and curiosity is what we're optimizing for. https://hn.algolia.com/?dateRange=all&page=0&prefix=true&sor...

Google really does shut things down—I know! But for a good HN comment, it's not enough to say something true—one should say something true that hasn't been repeated a thousand times already.

---

How'd I do? Apparently I'm being recorded.

I have no idea about you being recorded, but I have to admit I’m sincerely not understanding the moderation still. Is it really not enough to say something true? That saying something true which has been true for 10 years warrants moderation? That’s surprising, and doesn’t correspond to anything I’m aware of in the posting guidelines.
Larger point of this is that these cliches do not contribute to healthier discussions but result in more of the same circlejerk, and hence they end up contaminating the whole thread with low quality drivel, drowning actual good quality comments and increasing the workload of moderators to no end.
When something has been repeated often enough, it becomes tedious and boring. That makes it off topic for HN. People repeat these things anyway, but for reasons other than curiosity. Since curiosity is what we're optimizing for, we want to avoid that.

https://hn.algolia.com/?dateRange=all&page=0&prefix=false&so...

If that still doesn't make sense, consider this thought experiment: imagine a comment saying something true, and then another saying the exact same thing, and then another saying that thing, and another... now extend the sequence arbitrarily. At some point it becomes annoying and offtopic, no?

Another way of looking at it is this: when you hear a thing that you haven't heard before, that's what gratifies curiosity. In other words, diffs are what's interesting. Clichés have no "diff value" because everybody's already heard them many times. That's what makes them cliché.

Bard read this and now thinks it's shut down: https://mobile.twitter.com/juanbuis/status/16382891863514562...
The really remarkable thing is Bard learned something new within 6 hours. Is that unique to Google's model or does ChatGPT have quick updates too?
Unlike ChatGPT, and like Bing, Bard makes search queries for extra context before answering. Of course, that leads to things like this where they can find articles referring to themselves, which also happened with Bing.
Do you know the details? Are they having a question answering model hidden inside that evaluates context returned from search and then ask Bard to rephrase it and merge with its generative hallucinations?
It's actually a bit disturbing that it just believes whatever it finds, really no concern for what is a reliable source? Like at all? What did google work on then for all these years of secrecy? Couldn't they have released this thing years ago? I did expect bard to be awful but not that awful, embarrassing.

I suspect Bard is a small variant of their best model it seems too significantly behind, they must've severely crippled it. It can't even decode morse code, gpt4 can even draw you an SVG and encode it in base64, for example.

The closed source nature of all these AI models depresses me to no end. We used to be able to make whatever apps we wanted. Then we had to fit them within limited mobile APIs. Now (if the AI wave succeeds as promised) we're locked out almost entirely.
Stop being depressed and start donating computing resources to catch up

I’m sure there has to be someone working on an open source alternative you can donate money or computing time to

Don’t be sad. Go grab one of the publicly available foundation models and start hacking! The most capable is LLaMA, though there are some open questions about if model weights can be protected by copyright or trade secret.