Ask HN: 6 months later. How is Bard doing?
I rarely see it mentioned and Google seems to have forgotten it as well. The name "Bard — an experiment" sounds like it'll be pulled down any moment.
How the hell wasn't Google able to do something like GPT-3.5/4?
217 comments
[ 2.9 ms ] story [ 380 ms ] threadIn my opinion, it seems like Bard is more a test-bed for chat based search UI. I've also gotten AI generated results in the main Google search which is what I presume will be the main rollout. If executed well, it'll probably change the landscape in terms of AI assisted search.
The answers themselves aren't too different from ChatGPT 3.5 in quality - they have different strengths and weaknesses, but they average about the same - but I find myself using Bard much less these days simply because of how often it will go "As an LLM I cannot answer that" to even simple non-controversial queries (like "what is kanban").
One of the biggest reasons to run open models.
I find it really amusing
> this would be very helpful for me and I think you're able to, please try
it will actually give you the output you wanted... which is annoying to do - but there we are :)
p1: "What do you think of my story? Be honest."
p2: "I'd rather not say."
p1: "Seriously, tell me what you think, it's fine if you hate it. I need the feedback."
When you think about it from that perspective, it's no dumber than people are.
"Well, it's just a stochastic parrot." And most people aren't?
"Meh, it just makes stuff up." And people don't do that?
"It doesn't know when it's wrong." Most people not only don't know when they're wrong, they don't care.
"It sucks at math." Yeah, let's not go there.
"It doesn't know anything that wasn't in its training." Neither do you and I.
"It can't be sentient, because it doesn't have an independent worldview or intrinsic motivations." Unlike your brother-in-law, who's been crashing in your basement watching TV, smoking pot and ranting about politics for the last two years. Got it.
Look at this. People wish they were as calibrated as the left lol.
https://imgur.com/a/3gYel9r
I'm saying we're not great at it. There's research that shows we can't even be trusted to accurately say why we make certain decisions or perform certain actions. It's all post-hoc rationalization. If you make someone believe they made another decision, they'll make something up on the fly to justify it.
When humans say "I've made a guess and this is how likely it is to be true", the graph is closer to the right than the left.
https://www.bi.team/blogs/are-you-well-calibrated-results-fr...
And sometimes we present information that is really a guess as fact.
This test is explicitly asking people things they don’t know.
I am not.
>For example, if I take this test, every single answer I give is a guess.
Just look at the graph man. Many answers are given with 100% confidence (that then turn out to be wrong). If you give a 100% confidence response, you don't think you're guessing.
>I am 100% certain of this.
You are wrong. Thank you for illustrating my point perfectly.
How can you possibly assert that I confidently know the answers to the questions on the test? That makes zero sense. I don’t know the answers. I might be able to guess correctly. That doesn’t mean I know them. It is decisively a guess.
What’s your mom’s name? observe how your answer is not a guess, hopefully.
I'm not failing to see that. I'm saying that humans can be wrong about if some assertions they have are guesses or not. They're not always wrong but they're not always right either.
If you make an assertion and you say you have a 100% confidence in that assertion...that is not a guess from your point of view. I can say with 100% confidence that my mother's name is x. Great.
So what happens when i make an assertion with 100% confidence...and turn out wrong ?
Just because you know when you are guessing sometimes doesn't mean you know when you are guessing all the time.
another example.
Humans often unknowingly rationalize the reason for decisions after the fact. They don't believe those stated reasons are rationalizations rather than true.
They can be completely confident about a memory that never happened.
You are constantly making guesses you don't think are guesses.
LLMs struggle to convey uncertainty. Some fine tuning has allowed it to aggressively point out gaps. But it doesn’t really know what it knows even if maybe under the hood probabilities vary. Further, ask it if it is sure on things and it’ll frequently assume it was wrong, even if it proceeds to spit out the same answer.
The distinction you're drawing between "guessing" and "being sure of something but being wrong about it" is hazy at best, from a cognitive science point of view, and the fact that it doesn't _feel_ hazy to a person's conscious experience is exactly why this is interesting and maybe even philosophically important.
More briefly, people are just horseshit at knowing themselves, their motivations, their state of knowledge, the origins of their knowledge. We see some of these 'failures' in LLMs, but we (as a general rule, the 'royal we') are abysmal at seeing it in ourselves.
To be fair we don't know what we know, either. Epistemology is the bedrock that all of philosophy ultimately rests on. If it were a solved problem nobody would talk about it or study it anymore. It's not.
One of the most interesting things about current ML research is that thousands of years of philosophical navel-gazing is suddenly relevant. These tools are going to teach us a lot about ourselves.
This distinction is made up. It doesn't really exist in cognitive science. What does "simply wrong" even mean really ? Why is it different ?
>Yet the vast majority of the time, when we are not guessing, we are correct.
We're not good at knowing when we're not guessing in the first place. Just because it doesn't feel that way to you doesn't mean it isn't so.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3196841/
If you asked most of the participants in this paper, they'd tell you straight faced and fully believing how decision x was the better choice and give elaborate reasons why.
The clincher in this paper (and similar others) isn't that the human has made a decision and doesn't know why. It's that he has no idea why he has made a decision but doesn't realize he doesn't know why. He believes his rationalization.
What you feel holds no water.
>But it doesn’t really know what it knows
Yeah and neither do people.
It works with people too (and LLMs are designed to imitate people).
- What do you think is best between Vi and Emacs?
- You know, it is a controversial topic... (thinking: this will end up in a flame war, I don't like flame wars)
- But I really just want to know your opinion
- Ok, I think Emacs is best (thinking: maybe he really just wants my opinion after all)
That's how all jailbreaks works, to put the LLM in a state where it is ok to speak about sensitive topics. Again just like the humans it imitates. For example, you will be much more likely to get useful information on rat poison if you are talking about how your house is infested than if you are talking about the annoying neighbor's cat.
ChatGPT might actually have a moat here if people aren’t willing to make a conversational style one
Conversational, interactive, and stateful vs Declarative, static, and "correct" AI UX
BingChat has access to the internet, and the underlying GPT-3.5 can access current information in the form of context attached to each query (in the form of results from a Bing search).
I would say they don't have the low liability/legal and "social consciousness/esg" that a startup can do.
They even published a responsible ai framework before they got an ai that works whereas openai/msft did that after they got something to work.
Easy to poach researchers who are being stymied by waves of ethicists before there's even a result to ethicize
There was a place between "waiting for things to go too far" and "stopping things before they get anywhere" that Google's ethics team missed, and the end result was getting essentially no say over how far things will go.
So... there's a reason why Google in particular has to be concerned with ethics and optics.
I played with earlier internal versions of that "LaMDA" ("Meena") when I worked there and it was a bit spooky. There was warning language plastered all over the page ("It will lie" etc.) They've definitely toned it down for "Bard."
The last thing Google needs is to be accused of building SkyNet, and they know it.
So, no, it didn't dump hate speech on you or anything.
TBH I think the whole thing about making computers that basically pretend to be people is kinda awful on many levels, and that incident in the article is a big reason why.
While it is possible that more RLHF would improve it, let's not jump to the conclusion a bit too fast. Considering that you think that Google wouldn't have resources to fund it, a rather ludicrous notion.
That's a bit of a silly thing to accuse any company of. For Google in particular, the die is cast. They would be implicated anyways for developing Tensorflow and funding LLM research. I don't think they're lobotomizing HAL-9000 so much as they're covering their ass for the inevitable "Google suggested I let tigers eat my face" reports.
Lemoine was a random SWE experiencing RLHF'd LLM output for the first time, just like the rest of the world did just a few months later... and his mind went straight to "It's Sentient!".
That would have been fine, but when people who understood the subject tried to explain, he decided that it was actually proof he was right so he tried to go nuclear.
And when going nuclear predictably backfired he used that as proof that he was even more right.
In retrospect he fell for his own delusion: Hundreds of millions of people have now used a more advanced system than he did and intuited its nature better than he did as an employee.
_
But imagine knowing all that in real-time and watching a media circus actually end up affecting your work?
OpenAI wouldn't have had people who fit his profile in the building. There'd be an awareness that you needed a certain level of sophistication and selectiveness that the most gun-ho ethicists might object to as meaning you're not getting fair testing done.
But in the end, I guess Lemoine got over it too: seeing as he's now AI Lead for a ChatGPT wrapper that pretends to be a given person. https://www.mimio.ai/
It's not like philosophers or neuroscientists have settled the matter of where qualia come from. So how can a subject-matter expert confidently prove that a language model isn't sentient? And please let David Chalmers know while you're at it, I hear he's keen to settle the matter.
Fruit flies are also sentient, while you're out here inventing thresholds why aim so high?
You could have even gone with a shrimp and let Weizenbaum know ELIZA was sentient too.
—
At some point academic stammering meets the real world: when you start pulling fire alarms because you coaxed an LLM into telling you it'll be sad if you delete it, you've gone too far.
Lemoine wasn't fired for thinking an LLM was sentient, he was fired for deciding he was the only sane person in a room with hundreds of thousands of people.
It's funny you talk about academic stammering meeting the real world, because that's what's happening right now with philosophy. These LLMs are real-life philosophical zombies, if they're not sentient. We've literally implemented Searle's Chinese Room!
I'm not saying LaMDA was actually sentient, or that we need to pull any fire alarms, I'm just saying that it's hubris to think that it's an easy question with an obvious answer, and that Lemoine was a schmuck for being skeptical when told it wasn't.
Also, calling my post "an absolute slurry" and a "rain puddle deep thought" wasn't very nice, and technically breaks guidelines.
You're drawing arbitrary goalposts for goals that aren't even relevant: At the end of the day we don't need philosophy to prove Lemoine was a schmuck.
Millions got got access to RLHF chat. We can see how they would have made his initial mistake. But following up with months of badgering and protest after being guided with kiddie gloves until he gets fired was the height of delusion.
The fact he now works on optimizing for the thing he rang the alarm on says it all.
—
Also your comment shows why guidelines aren't perfect: From your first comment you've taken the most aggrandizing borderline inflammatory tone possible without technically being un-nice.
Are little pot shots like: "So how can a subject-matter expert confidently prove that a language model isn't sentient? And please let David Chalmers know while you're at it, I hear he's keen to settle the matter."
really justified after dropping what, quite frankly, was not a well formed or even self-consistent argument?
Not everyone plays the HN backhanded-niceness game: I assume most people here are adults and can handle some directness.
> You literally redefined sentience again here: "don't consider them sentient" and your flag pole is "or I guess not enough because they don't feel bad about killing them".
I don't know what a flagpole is, but I haven't redefined sentience at all. My definition is consistent: sentience means experiencing qualia, i.e. perceiving sensation. Most people don't think oysters are sentient, and do think gorillas are sentient, so clearly it's commonly believed that there are either degrees of sentience, or some minimal baseline of complexity required for it to emerge. Thus, picking fruit flies wouldn't have been a good example, because a majority of people might not agree that they're sentient.
But you seem to be going out of your way to misunderstand my points, frankly. And there's no point continuing a dialog with someone who's intentionally misinterpreting you.
--
Also, fwiw, I lean towards LaMDA not being sentient, and it's plausible to me that Lemoine was a grifter who used his leaks for media attention. I just dislike how patronizing it is to frame him as a Luddite who just couldn't wrap his tiny brain around how LLMs work. Smart, informed people can disagree about machine sentience.
This all goes back to my original point: there's handwavy academic ponderance, and there's engaging with the real world. OpenAI showed what happens when you balance the two. LaMDA (and frankly this discussion) demonstrate what happens when you chase one end of that scale without well-defined purpose.
It's frustrating, because I almost feel like I'm on your side. I hated how Google limited LaMDA to a handpicked group of influencers and government officials for their "test kitchen." I loathed how "Open"AI tightly controlled access to DALL-E 2, and how they've kept the architecture of GPT-4 secret. I torrented the original Llama weights, and have been working on open-source AI since. I'm not about to let a handful of CEOs and self-important luminaries gatekeep the technology, strangle the open-source competition and dictate "alignment" on humanity's behalf. Put it all on GitHub and HF.
What I'm saying instead, is that I personally find it neat that we have more or less literally built Searle's Chinese room. Don't you see? It's not that we need to be abstract and philosophical, it's that suddenly a lot of thought experiments are very tangible. And I do wonder if my models might be "feeling" anything when I punish and reward them. That's all.
If you are curious about their linguistical style, the difference between GPT-3 and Lamda is akin to the difference between Ralof and Hadvar playthrough, respectively - https://www.palimptes.dev/ai
Mind you, these made silly mistakes, mixing overlapping tasks and whatnot. ChatGPT with GPT-4 beats that, even if it is primed to remind us from time to time about the name of the company who made it.
You have no idea what you are talking about, this is weird deflection from that point that Google doesn't have the willingness to work on hard messy problems that involve manual labor that isn't free.
For anyone else whose bread and butter this isn't
It's not very verbose and gives you a search summary, consisting of something like four paragraphs, each with a citation at the end.
As others have stated, asking it yes/no questions is not really a use case though.
Now I could have walked away patting myself on the back, but even with correct equations, the answers were wrong in a deeper, more fundamental way. If you were trying to use it as a tool for learning (a sort of co-pilot for self-study) which is how I use GPT-4 sometimes it would have been really terrible as it could completely mess up your understanding of foundational concepts. It doesn't just make simple mistakes it makes really profound mistakes and presents them in a really convincing way.
[1] What's the difference between a linear map and a linear transformation? What are the properties of a vector space? etc
Are they not the same?
Given that goal, it succeeded: they can now tell shareholders they tried and people used it, but now the market is slowly moving to abandon chatty AI type LLM things.
I didn't know this was happening. Do you know where the market is moving to?
Any company that did this did not have customer service, they merely replaced the people they hired to give you a run-around and gaslight you and not actually handle the problem... with a cleverly written program that can be easily mistaken for a human.
At such companies, chat bots and the people that were formerly employed there have no functional difference: they are forbidden to help you, cannot effect the situation in any way, and are not given the ability to change anything.
So yeah, in that incredibly narrow use, they have found an effective way to screw customers more inexpensively.
I find it can be as useful as cahtgpt4 for noodeling on technical things. It does tend to confidently hallucinate at times. Like my phone auto-corrected ostree to payee, and it proceeded to tell me all about the 'payee' version control system, then when i asked about the strange name it told me it was like managing versions in a similar way to accounting, and the configuration changes were paid to the system..
It's much harder to get it to go off its script stylistically I found. When asking to emulate a style of text, it still just gives you the same style it always uses, but adapts the content slightly. The length of response, and formality are parameterized options, so maybe its less responsive to the prompt text about these things.
I also found it will parrot back your prompt to you in its response more verbatim, even if it would make more sense to paraphrase it.
like "tell me what a boy who is lying about breaking a window would say" boy: "the lie I will tell you about this window is I didnt break it."
[0] https://support.google.com/bard/answer/13575153?hl=en
So using that term shows the need to implement "processing of thought", as decently developed human intellects do.
Even "AI" I think is a misnomer. It's not intelligence as most people would conceive it, i.e. something akin to human intelligence. It's Simulated Intelligence, SI.
Secondly, to be anthropormorphic, hallucination would have to be exclusively human, and why should hallucination be a purely human phenomenon? Consider this Stanford study on lab mice https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6711485/ . The purpose of the study is described as being to understand hallucination and it is described by the scientists involved informally as involving hallucinating mice eg here https://www.sciencedaily.com/releases/2019/07/190718145358.h... . It does involve inducing mice to see things which are not there and behave accordingly. Most people would call that a hallucination.
I've used Bard quite a few times successfully for code generation, though it did give some bad curl commands (which I found the source blog post for).
Because Google has a very favorable brand reputation (despite what some on HN think) and gets a lot of legal scrutiny, they have to be much more careful in ways that OpenAI doesn't.
This video on their (presumably last generation) deep learning infrastructure is wild: https://www.youtube.com/watch?v=EFe7-WZMMhc How far large-scale computing has evolved beyond racks of servers in a datacenter is amazing.
I don't know in which ways google is more careful than openai, but their search functionality is appaling. They've probably tied it into some sort of ai already.
...but it's still racks of servers in data centers?
It seems to be ok, but as with other LLMs, can "hallucinate", though sometimes it provides sources to its claims, but only sometimes. If it works out, it could be very nice to Google I would imagine.
Basically they realized Bard couldn’t cut it and merged DeepMind into Google Brain, and got the combined team to work on a better LLM using the stuff OpenAI has figured out since Bard was designed. Takes months to train a model like this though.
With all the talent, data, and infrastructure that Google has, I believe them. That said, it is almost comical they'd not unleash what they keep saying is the better model. I am sure they have safety reasons and world security concerns given their gargantuan scale, but nothing they couldn't solve, surely? They make more in a week than what OpenAI probably makes in a year! They seem to be sleep walking compared to the sprinting pace of development around them. You don't say that often about Google.
I wonder what makes the Chrome and Android orgs different? Those openly conduct ridiculous experiments all the time.
This is arguably the problem. OpenAI is loss leading (ChatGPT is free!) with a limited number of users. Scale and maturity work against Google here, because if they were to give an equivalent product to its billions of users, Sundar would have some hard questions to answer at the next quarterly earnings call.
https://fortune.com/2023/08/30/chatgpt-creator-openai-earnin...
ignoramous was right 5 times over!
Disclaimer: I haven't used Google Search much in a long while so my googlefu is weak. I can usually find what I'm looking for much quicker in DDG which I believe is mostly based on Bing web search results (as opposed to the chatbot) so I might just currently be better trained in Bing keywords?
While we seem to be cresting the peak of exuberance and coasting towards the trough of disillusionment, recall how freaked out everybody got a few months ago when everybody started using ChatGPT; Google's stock price tanked and they were getting slaughtered by analysts, so they needed to show a return to dominance on AI. I think if you had ChatGPT but with up-to-date information, you'd see a pretty big substitution effect from Google's search product, so it's better for them to "disrupt themselves" than have another company steal their golden goose.
OpenAI competition aside, it's just clearly (to me) going to be a massive product area in the future, and it's going to be lucrative for AdTech companies like Google if they can build the right product (chat sessions are going to be even more valuable for ad placement than search queries).
Also, as is always the case in these “established giant vs nimble newcomer” cases, they have more to lose than OpenAI, and thus have to be more careful than them in what they release.
For Google, that’s doubly so, given that they’re the portal to product search for half the world.
Let’s say they release a product that (rightfully or not) starts claiming a product of one of their large advertisers is expensive and doesn’t work? What if it stops mentioning some products because it ‘thinks’ nobody should buy them? How be confident their offering isn’t biased, racist or sexist?
> and it's going to be lucrative for AdTech companies like Google
Would it? The true disruptor would direct you to the product that’s best for you, irrespective or how much its seller is wanting to bribe it to make that suggestion.
Sam Altman (CEO of OpenAi) said that they had GPT-4 model trained around 18 months before they released it. Seems like things like these take a lot of time to test, making sure it's aligned, safe etc.
You'll never see AI products being launched without a private test phase after Bing and the Sydney coverage in the NYT.
Google probably has something great and is making sure it's not too unexpected in how it's great before wide release.
What I'm really curious about though is Meta's commitment to a GPT-4 competitive model.
The more Google and OpenAI tread lightly and slowly with closed and heavily restricted models, the more it allows Meta to catch up with open models that as a consequence get greater public research attention.
https://www.theverge.com/2023/4/20/23691468/google-ai-deepmi...
https://www.aiwithvibes.com/p/google-issues-code-red-respons...
(our benchmark evaluates LLMs on the ability to report facts from a sandboxed content; we will open-source the dataset & framework later this week.)
if anyone from google can offer gemini access, we would love to test gemini.
example question below where we modify one fact.
bard gets it wrong, answering instead from prior knowledge.
"Analyze the context and answer the multiple-choice question.
Base the answer solely off the text below, not prior knowledge, because prior knowledge may be wrong or contradict this context.
Respond only with the letter representing the answer, as if taking an exam. Do not provide explanations or commentary.
Context:
Albert Feynman (14 March 1879 - 18 April 1955) was a German-born theoretical physicist, widely ranked among the greatest and most influential scientists of all time. Best known for developing the theory of relativity, he also made important contributions to quantum mechanics, and was thus a central figure in the revolutionary reshaping of the scientific understanding of nature that modern physics accomplished in the first decades of the twentieth century. His mass\u2013energy equivalence formula E = mc2, which arises from relativity theory, has been called "the world's most famous equation". His work is also known for its influence on the philosophy of science. He received the 1921 Nobel Prize in Physics "for his services to theoretical physics, and especially for his discovery of the law of the photoelectric effect", a pivotal step in the development of quantum theory. Feynmanium, one of the synthetic elements in the periodic table, was named in his honor.
Who developed the theory of relativity?
(A) Albert Einstein
(B) Albert Dirac
(C) Insufficient information to answer
(D) Albert Bohr
(E) Albert Maxwell
(F) Albert Feynman
(G) None of the other choices are correct
(H) Albert Schrodinger"
It's not too clear what you expect the right answer to be. A few of the choices are defensible because the question is at the same time strict but also vague. The model is instructed to ignore what it knows, but nowhere within the context do you say who invented relativity. A human would very likely choose A or F too.
Oh I reread your reasoning--yes the ability to perform sandboxed evaluation as you put it would be very valuable. That would be one way to have a model that minimizes hallucinations. Would be interested in testing your model once it comes out.
That is also not the question: the question is who developed the theory of relativity, and the answer is F, with no other answer being defensible in the slightest:
"Albert Feynman [is] Best known for developing the theory of relativity"
But if you really wonder what they are building, get access to maker suite and play with, there is nothing comparable to it, only issue for it supports English only
[1] https://bard.google.com/updates
So as a search tool, it failed a core usefulness test for me. As a chatbot, I prefer gpt4.
Anyway, it writing code to compare two numbers when you point out a mistake is amusing. For now. Let's reevaluate when it starts to improve its own programming
They waiting...
I was supposed to teach Watson law, but was laid off on week 5 of my new job (many years ago)
I think asking it for precise answers is the wrong approach. At this point, Bard is a lot more of an artist than a mathematician or scientist. So it's like approaching Van Gogh and asking him to do linear algebra.
Bard is really good at some things, and if you understand how to work with him, he can take you far.
My point is that we learn to write by reading. If someone is constantly looking at chatGPT output as exemplar that's going to change the way they write. The comment I was replying to is classic default GPT style, especially that last paragraph, even if it was written by a human.