Can someone in the know explain how Bard could be as bad as it is, considering Google has been investing in AI for as long as they have? I thought they were earlier than most to begin serious work on AI? How could Bard suck this bad?
This is the premise that a mature incumbent will usually fail to respond fast enough to disruptive innovations by an upstart because their organization is optimized to continue providing existing customers sustaining improvements to the products they are already paying for.
This is the premise that cross-functional coordination difficulty grows superlinearly as the organization scales.
The relative influence of these phenomena is a matter of debate. And, there may be other factors at play.
But, the mix of needing to respond (against organizational incentives) to a disruptive innovation while also trying to align long-established factions, product teams and PAs - many with a motivation to have an outsized influence over the new effort - within the organization presents a Herculean challenge.
The answer to why BigCo can’t outcompete startup is always the same.
In BigCo you have ego that blocks authentic artistic critique. There are traditions. There are rules. We learned our lesson last time.
In the best organizations, artistic critique is a time-honored tradition passed down from the elders to the padwans. Young Jedi learn how to navigate the matrix with the blessings of their ancestors.
Unfortunately, hacking culture in a big org is outré. What made Apple special was that Steve Jobs created many hidden micro cultures within his asylum that allowed so many unorthodox experiments, some of them even worked out.
Maybe you have some other examples of people who have been able to hold a core business and build the next iteration many many times successfully? They’re rare birds.
Edit: for clarity, Google could have a killerAI product in a year if they would kill their ego. Demis Hassibis (I think that’s how you spell his name, sorry) the CEO @ DeepMind can absolutely build the thing they want, but they’d have to give him carte Blanche, and they won’t.
I put all questions to both Bard and chatGPT. The exact same words.
Bard is almost always better.
For example try: "why does my hole saw have a spring over the drill bit?"
Bard is correct, chatGPT makes up nonsense. (There's no such thing as "pilot bit extension spring", and the rest of the reply makes little sense as well.)
I've been asking science and math questions to Bard and ChatGPT (at about an undergraduate level) and likewise found Bard to be consistently better.
As an example I asked both how many groups there are of order 6. Bard gave the correct answer of 2 and listed what they were. ChatGPT said there were five and proceeded to list seven.
> To determine this, we consider the possible structures of a group of order 6. Since 6 = 2 × 3, where both 2 and 3 are prime numbers, the groups of order 6 can either be cyclic or a direct product of smaller cyclic groups.
> Cyclic Group of Order 6: There is always one cyclic group for any given order. The cyclic group of order 6 is denoted as 𝐶₆ or ℤ₆, which is generated by an element of order 6.
> Non-Cyclic Group: The other possibility is a non-cyclic group. By Sylow's theorems, the number of elements of order 2 must divide 3 and be congruent to 1 modulo 2, and the number of elements of order 3 must divide 2 and be congruent to 1 modulo 3. This means there must be one Sylow 2-subgroup and one Sylow 3-subgroup. The non-cyclic group of order 6 is the symmetric group 𝑆₃, which is the group of all permutations of three objects. It is non-abelian and can also be thought of as the semi-direct product of 𝐶₂ and 𝐶₃.
> So, there are two groups of order 6: the cyclic group ℤ₆ and the symmetric group 𝑆₃.
3.5 is a weak model that is only good at informal conversation. When people talk about using ChatGPT and its potential (programming, vision, etc.) they talk about GPT-4. If you don't want to pay, you should use Bing Chat for your comparison. 3.5 would not be worth $20 a month but GPT-4 is.
It's funny to me how often this mistake is being made here on HN, despite this exact thing being repeated time and again. There is a big difference between 3.5 and 4.
Google has good research. Bard is half baked therefore not a great product as of now & not sure if or when. Maybe they were under so much pressure from openai that they indeed rush released it, way before it had to be seen out by public eyes. To me it seems/feels in beta
They are worse than GPT-4 because nobody other than OpenAI (and their free Azure credits) have been willing to throw $100 million in training a single LLM.
I’m not surprised by my experiences many years ago when trying to use google to support my small app that hit the front page of HN. Not only did they take my app offline giving me no recourse for help thus destroying my momentum , it was because I had “upgraded” my server hosting. Apart from this being funny and ironic what was worse is when I mentioned it a year or so later google engineers came out of the wood work on Hn to accuse me fibbing and making a big deal out of something that they clearly had no relation to. I think it was a because they were scared they might be linked to the incident and it would give them a poor performance metric. It was super creepy but made me understand that the organisation is just vast with little concern for the places where innovation comes from and full of selfish gatekeepers worried about looking bad. I have nothing against the company and the good it’s done for me and the world I still rather like it, but I understand that like all good companies without careful attention to respecting the little guy and allowing care and checks through its structure to root out social issues it’s destined to always lag behind. Just my 2 cents and if anything reminds me of what a terribly difficult task I would have if I was a ceo of my own company ( if it were large ).
I think most large companies have the same problem and the only solution is to literally test different teams with how they handle specific case studies. It would require alot of careful planning and studying but it would be the only way to resolve such problems. An alternative would be to have an anonymous feedback that is constantly pushed onto customers and constantly checked and mined for insight. if was a major company I would hire myself or someone who had my experiences to do this. But I’m already retired/funemployed lol. Sorry for the rant.
This is also because in Google you tend to get more easily promoted once you launch a new product, than when you do maintenance.
So the most logical course at this stage, based on typical politics at Google, would be to shutdown Bard, and launch a new product instead (like what has happened with all the messengers / social network apps)
Makes sense to me. Google clearly has a culture/reputation where projects are routinely killed off because they aren't deemed "successful" by some arbitrary metric. This fosters a "failure/success" mentality among employees, which influences what types of projects they take (obvious, immediately profitable) and how long it takes to develop them (as quick as possible). While they have a lot of smart people working there, I think big G has some organisational problems which prevents them from completing large scale experimental projects like this.
For creative writing, I prefer Bard over the other ones for a few reasons. First off, it’s free with any Gmail account and secondly it has draft (3) that I can compare. It may not seem like much but I have a creative workflow and like it much more than Bing AI. I’ve been trying to use GitHub copilot and it is a lot less reasonable to use for technical work. But throwing an idea at Bard is kind of fun once you get in to promoting
Have you tried using Claude for creative writing? I find its creativity to be quite far ahead of GPT-4. You can try it for free on claude.ai, although it needs a phone number and isn't available in all countries (but that can be solved with a VPN, even if the phone number is from a different country).
I’m trying it out now based on your suggestion. It might be limited in allowed messages, which I think is 10 per 8 hours for free tier and twice for pro tier. It seems comparable to bard without the draft feature
> simply type the name of the product that you want to access after the "@gmail" prompt. For example, to access Google Docs, you would type "@gmail docs".
Me
> @gmail docs
Bard
> Unfortunately, Bard Extensions is still under development and currently doesn't support full integration with Google Docs through Gmail.
Where did you see that? It seems this is integrating two extensions together, I'm curious how that would work. I asked it "@gmail summarize my last 5 emails" and that worked decently.
Edit: Actually, I'm not completely certain how the "@<extension>" syntax works. Are the results limited to that app?
I just tried "@gmail summarize my last 5 emails" and got:
> I don't have access to your Gmail account, so I can't summarize your last 5 emails. However, I can help you summarize your emails if you provide them to me.
You must enable Bard history for these features to work, and you must go to the extensions page and make sure they’re turned on. Arguing with a model that doesn’t have access to the extensions won’t make it suddenly use the extensions.
Summarizing my last 5 emails worked just fine with bard using that query.
> You do have access via the new "extensions" feature.
And it said:
> Thank you for reminding me about the new Gmail extensions feature. I am still under development and working on integrating with those features. Unfortunately, I am currently unable to access your Gmail account even with your permission.
>
> However, once the integration is complete, I will be able to access your Gmail account with your permission and summarize your last 5 emails. I apologize for any inconvenience this may cause. Is there anything else I can help you with in the meantime?
I've tried using it for generating articles based on provided fact sheet, and it somehow managed to invent a bunch of facts while referencing the fact sheet, without directly mentioning anything in the fact sheet. This is just so bad compared to OpenAI 4 or even 3.5.
How is that possible given that some of the smartest people work for Google?
If you arre logged in and gave accepted the terms but your primary Google account is a workspace account that doesn't have bard enabled it also redirects.
Not bad. It’s limited in what it can do but I would probably check Bard if ChatGPT went down.
I have the “AI search” extension enabled in Google search as well and it’s pretty unreliable. Seems like half of my coding questions it hallucinated some function that doesn’t exist.
But if Google release Bard in its current state a year ago most people probably would be using it now, even if it’s inferior.
I've used it to find flights and it was actually really nice. I didn't have a destination for my vacation so I asked for flights for cities at least X km away from me and costing at most Y, in the next Z months. It promptly answered.
I've had success with it. The integration with other google services could be a killer feature, but I think it needs some more polish and control. For example, right now it can't add an event to your calendar. I'd think it could at least generate a link to create one. It sort of seems like they're being overly cautious when compared to the way ChatGPT was launched.
I'm impressed. I asked it "How much is a flight from San Francisco to the rapid & blitz tournament over Christmas?" and it figured out which tournament I was talking about and showed me ticket prices.
Ugh, I tried Bard too, but wasn't as impressed. Granted, I had a specific request with a stop-over for a couple of days, but it wasn't able to complete it, only the first leg. A follow-up question then prompted it a look-up a round trip flight for the 2nd leg.
Surprised to see this here, was just reading about it on the bard page a few minutes ago since I been killing time waiting for chatGPT to come back up for Plus Users. I think it would be really cool if we could get a Bard extension for Google Colab, maybe some integration with the Colab Pro would be really interesting analogous to the Data Analysis Plugin
does it still just say that it will then not actually do anything?
It will now say:
I can't directly create calendar events for you, but I can help you create a
calendar event description. Please provide me with the following information so
I can create a detailed description for you:
* Event title
* Event date and time
* Event location
* Event attendees (if applicable)
* Event description
* Any additional notes or reminders
Once I have this information, I can provide you with a calendar event
description that you can then add to your calendar.
And if you provide that information, it'll just format it, add a few tokens and just display it to you.
Just checked out Bard for the first time because of the OpenAI outage. It sucks. It feels worse than GPT-3.5 with worse reasoning and instruction following. This is embarrassing for Google.
EDIT: Just to be clear: I want it to be good. It's unfortunate that it's underwhelming.
Bard is utter junk. I used to proclaim to everyone how it had its place as an equal tool in my daily AI tool chest specifically for web browsing. When they updated it and said I could access YouTube transcriptions etc it never really could and still can't. It hallucinates far more than any other llm premium model like Claude or Perplexity. Thus far they better get their s** together for Gemini or they're toast and this is coming from somebody who uses AI llms all day long for half a year
LLMs don't retrieve anything unless there's infra in place so that they can go and do that. It's called Retrieval Augmented Generation (RAG) and it's still pretty new enough that it hasn't been integrated in all workflows. There was a bunch of copy early on claiming LLMs "read the internet and find you results" but it's weaselly wording and was until remarkably recently complete BS.
Bard has had access to Google caches since launch. I asked it about my 8-follower Twitter account and it quoted my first Tweet and the registration month correctly. It also hallucinated my gender and described me as "a popular figure in the French-speaking Twitter community" (because my name is a French word, I guess).
Thanks, I know. I built RAG systems with vector databases and related infra. And yes, that's exactly what it claims to do, even with "results verified by Google" in the interface.
Blame your government.
And I don't mean that in a hostile way. People need to temper their expectations when using a product that is in the crosshairs of their government. Alphabet and Meta aren't startups and they have tangible regulatory obligations.
I think you're referring to Canada's Online News Act, which would require companies to pay news outlets for the content they share or repurpose. I'm a Canadian familiar with this story, and Google's opposition to it.
It surprise me if Google would limit Bard in Canada because of the Online News Act. Wouldn't the importance of driving adoption of Bard be much larger than the impact of having to remove Canadian news?
Google says "regulatory uncertainty". It sounds like they're unhappy with Canadian regulators. Be thankful your government is giving them enough grief.
The government is involved in a shakedown of Google for the benefit of bell and Rogers (whose lobbyists wrote it). I’m glad they are pushing back.
There is a law professor at U of Ottawa who is generally considered the most expert and independent on these topics and he has been blowing the whistle on this bill c-18. Michael Geist.
Rogers and bell have quite the grip on our governments but this one really takes the cake.
Likely the Quebec privacy laws which are gdpr equivalent in many cases. This forces tech companies to create a different experience for Quebec vs rest of Canada. As neither have enough user size, it's often easier to skip Canada altogether till the team has the spare bandwidth for this.
>This forces tech companies to create a different experience for Quebec vs rest of Canada.
That sort of seems like a needless barrier for companies. What is the precedent for that? It almost seems like those in Quebec have established some sort of "We're different, so everyone must suffer" constitutional amendment. But then again, I have no clue.
The Canadian government is picking a fight with Google, Meta, and others right now so that's likely why Google is withholding something like this to help create some pressure.
