Why Vivaldi won't follow the current AI trend?
vivaldi.com
vivaldi.com
They don't say that you should feel bad. The author simply says that the characteristics and unreliability of LLMs make them unfit to be implemented in the Vivaldi browser at this moment. They even close off the article with a positive sentiment about the future of the ML field. They don't burn any bridges there.
> While they do this, they also use vast amounts of energy and are happy using all the GPUs you can throw at them which is a problem we’ve seen before in the field of cryptocurrencies.
But is he wrong?
What does "a lot of energy" even mean? It's meaningless. Do they use too much energy for their usefulness? If they did, people wouldn't be using them.
Like with LLMs, some of us do. But still there are places in the world where one is imprisoned without car. Which is not case with LLMs (yet).
Cars enable long-distance travel, which in many parts of the world is now essential for survival, and often that is, ironically, a result of climate change.
LLMs do nothing to aid human survival. In a post-apocalyptic world, cars would be sought after for both transport and shelter, but nobody would give a care that we lost ChatGPT.
Cars use more energy than LLMs because most of the world is built around requiring a car to get to places.
It's not that we don't mind it; quite the contrary, in fact, as we're always reminded of when the price of gas/petrol shoots upwards.
When the cost of living forces you to live 45 minutes away from your job by car because living within walking distance is astronomical, many folks are conditioned to want a house with a garage "for the kids" because apartments are "for the poor," and EVs, in their current state, are not financially feasible yet (though that is changing!), it's more that we're forced to tolerate it.
I know people don't mind the cost relative to the value they're getting, because if they did, they wouldn't use LLMs.
I think it’s fair to ensure that you create something that isn’t subject to possible quotas in the future
There's plenty of documentation on that directly from the source (openai, etc.) and elsewhere.
> While they do this, they also use vast amounts of energy and are happy using all the GPUs you can throw at them which is a problem we’ve seen before in the field of cryptocurrencies.
> While they do this, they also use vast amounts of energy and are happy using all the GPUs you can throw at them which is a problem we’ve seen before in the field of cryptocurrencies.
"Vast amounts of energy", "happy using all the GPUs you can throw at them", and then the random cryptocurrency reference to evoke all the bad stuff associated with those.
Edit: I do think you're making a fair observation, I just feel there are a few reasons not to draw too strong of a conclusion from it
> Despite all this, we feel that the field on machine learning in general remains an exciting one and may lead to features that are actually useful. In the future, we hope that it will allow us to bring good privacy-respecting features to our users with a focus on improving discoverability and accesibility.
Edit: yes it does make mistakes, but it takes vastly shorter amount of time to find them rather than writing something from scratch.
To me, it feels like there will definitely be features or even entire products out there that wouldn’t exist without something that lowers the barrier of entry far enough for someone to overcome the zone of proximal development: https://en.m.wikipedia.org/wiki/Zone_of_proximal_development
Just the other day I was able to introduce some custom ESLint rules into the project to make the import names match the filename (e.g. when someone renames a Vue component but the import names don’t change or you don’t notice those usages, so you end up with bunches of <OldComponent /> all over the place). I wouldn’t have had the time to do so without ChatGPT/GitHub Copilot nudging me in the right direction and giving me examples to work off of, because frankly my searches for about 15 minutes didn’t turn up anything either.
Clearly that counts for something because nobody in any of my projects had done that previously, now it’s obvious that basically we need a local package added to the project (in package.js) but also that we can store it in the file system and just use the file: syntax. It was much the same story with some .cjs Node prebuild scripts that remind the developers that they should make examples of the new components they make in a development showcase page and so on.
Aside from that, LLMs are useful for boilerplate stuff (hello, ASP.NET and Spring Boot), as well as simple one shots when you don’t want to rack your brain coming up with a particular RegEx, amongst other cases.
If someone wants an LLM in the browser, good for them! If someone doesn’t want to add them to their browser project, that’s their rightful choice.
As a dyed in the wool Firefox user, I am of course very used to disappointment about how they direct their development effort. There are a million things I'd rather see them work on than strapping in an LLM. I don't know much about the Vivaldi community but I assume they took this into account when making this decision.
You don't have to cater to every user. Focusing on things you do better for a huge portion of potential users is a much better decision. If features turn out to be so useful that everyone needs to integrate them to not be left behind you have to do it. LLMs are not there yet. They are to unreliable, expensive and difficult to run locally. In a few years when most users might be able to run decent models locally you can reevaluate that decision
If every company and their dog started churning out their own LLM as a business or revenue model, one can easily see the supply overwhelming demand here apart from many other problems.
This certainly seems the case on the iOS/Play stores. Primary AI apps from the big hitters (OpenAI/Gemini/Copilot etc.) but then loads of knock-off ones from those riding the AI hype train. All the knockoffs being basically wrappers around APIs from the big hitters with ads/in-app purchases.
I've been burned by LLMs several times where it always came-up with some answer while it was either misleading or there should be no answer at all.
Generally, I thing that the llm should be its own service and everything else should have an easy way to connect to it, but I'm a lowly user, not a product manager.
Having much better text to voice could also be nice for the blind. While screen readers are fine I don't know how bothersome that robot voice is for longer texts.
Aren't most people using vivaldi for philosophical reasons in the first place ?
> So, as we have seen, LLMs are essentially confident-sounding lying machines with a penchant to occasionally disclose private data or plagiarise existing work. While they do this, they also use vast amounts of energy and are happy using all the GPUs you can throw at them which is a problem we’ve seen before in the field of cryptocurrencies.
Odd to compare to crypto which only has "store of value" as a use. It would be interesting to compare the energy of an LLM doing research tasks vs other means (e.g. a human working 8 hours a day).
In fact, I’d prefer it not to be included as it’s bloatware.
Unfortunately, the lack of AI functionality has made Edge a more attractive browser. Very useful to let AI summarise a large wall of text when browsing Hacker News for example. If only they'd add a 'close all tabs but this one' feature.
For now.
Are they the same people? The vast majority of users are consumers, they don't even think about these stuff, they only complain when the product degrade.
I don't see the privacy crowd who were conscious about their data and algorithmic feeds, are treating commercial LLMs products differently.
???
Right click tab -> O ?
Edit: oh, you're talking about Edge, not Vivaldi
To be perfectly honest, it’s a good browser. Being able to explain console errors with a click and them adding some resource limit functionality (like Opera GX) are all welcome. I even tried it on my MacBook and it works without issues.
I did go back to Firefox later though since I prefer those DevTools otherwise and because a recent update no longer makes Firefox fail to do DNS through my VPN.
This right here is why I'm against LLMs big time.
Walls of text are long for a reason. Yes, many of them might have filler. But sometimes there is nuance in the filler that a summarizer will happily throw away.
I don't want long form to be Instagrammified --- condensed into a ten-second paragraphs with 40% of the details that perfectly fits an endless consumption model that's perfect for advertising.
Well written articles deserve to be long form and not summarized as you note. But the large majority of web articles belong to the former category.
Still, things in the Siri/Alexa class have to have some kind of language model.
Of course, there are a lot of other considerations linked to topics like accountability, etc.. but take autonomous driving as an example: we don't need self-driving cars to be perfect and do exactly 0 incidents to be a better solution than human driving. They just need to be significantly better than us to be a very useful technology for mankind and let's hope this happens soon.
I've seen business people write prompts. They don't understand the limitations or strengths. They tell the model it has a degree in something, they tell it not to hallucinate, they anthropomorphize it so far that they greet it in the morning "to get into the mindset that you have a virtual colleague".
All the people who fearmonger about AGI really don't help, because it implies our LLMs are anywhere close to sentience. They're not.
Maybe this way Vivaldi can finally allocate resources to workspaces sync across devices.
Only 2 browsers do that: Edge and Arc. And Arc doesn’t work in Linux unfortunately.
Right now, I'm doing this very manually, but am actually thinking of switching my browser just to have this capability baked in.
EDIT: Since everyone is asking why and how I use summaries, here's an example: https://news.ycombinator.com/item?id=40881021
I just don't read it in the first place.
I fucking read, like absolutely everyone and their dog did two years ago before this pest stole everyone's attention.
Sure, and 10% of the time you get something that's lies which you don't get from actual reading.
But for a lot of non-fiction content, I think it's perfectly legitimate and common to have some specific piece of information you're looking to extract, or want a rundown of key highlights.
For both, I'd say it could also help in choosing which content to read. Books already typically have a form of summary on the back. If you're researching something, you're going to have to determine which articles are relevant before reading every article top to bottom - and a summary seems helpful for that.
I either read articles that are 1) fun to read and silly, or 2) fun to read and informative. If they're not fun to read, it's unlikely that the content is any good. Good educators often know how to write in an engaging way. If the content is difficult to read, or super lengthy, it's probably shit content.
How do you judge quality of the information you're given when you have it summarized? How do you make sure the LLM didn't misunderstand something, like sarcasm, or a quote, etc.?
Gruber has always been loquacious, but this one seemed excessive to me. Still, I was curious for his overall impressions on Apple Intelligence.
The summary sounded interesting to me, so I then dedicated time to read the source. To your credit, ChatGPT's summary was much more critical than the actual article.
Most of the time, I don't want a summary of an article. If the article is really good and I want to know the most important details, then I should read the article. Maybe I can write a summary myself, so that I can use it for future reference.
If an article is bad, then why would I care about a summary? I wouldn't read it either way.
There is a a possibility that an article is actually really good, but it's badly written, or way too verbose, or I don't understand the language/terminology/topic. In these cases, a simpler summary written by someone else (or an LLM, I guess...) would be useful. Domain-specific articles or scientific papers would fall into this category for me. Even for this subset of articles, if I don't really understand the topic being discussed, would a summary really give me any benefits? I still don't understand the topic well enough for it to be useful to me. I would get more out of actually trying to read the article and studying the material being discussed.
This is active reading, not some robo-moron digesting an article, shitting it out and then spoon-feeding the contents of the toilet-bowl back to you.
Then why are you reading it? You’re conditioning yourself to have a mental corpus of uninteresting, unimportant drivel. At best, that’s a waste of time.
Will you miss some cool stuff? Yes
Will you still encounter time wasters? Yes
We are consuming too much and not doing enough productively ourselves.
Their reasoning is shoddy. Misunderstands LLM's and plagiarism is of no practical concern to a browser, may as well remove the copy function from the browser if that's the view, theres also use cases where hallucinations don't matter such as the AI tab sorting and semantic search. He should've just said they don't see the value in it and left it at that.
Could be the copyright-LLM-bad crowd or luddites, who knows. /s
I think the article is excellent in its analysis and conclusion.
As much as I like the recent advancements in Generative AI, I don't think it should be hamfisted in every software, and I also don't think they come nearly close to all the hype surrounding it. It is a useful tool when you understand its shortcomings, and there is a long way to go until they are ready for more widespread use. Good on Vivaldi for taking this stance.
As a long time Firefox user, I replaced FF with Vivaldi on Android and haven't looked back, since FF has failed to address the battery drain issues that so many people are experiencing:
https://connect.mozilla.org/t5/discussions/firefox-for-andro...
No way bundling a LLM is going to help that matter.
I am pleasantly surprised how snappy Vivaldi feels.
Firefox is my backup browser. I really like it on desktop, but I agree that on mobile it feels clunky.
I see no reason to hamfist LLM into the browser itself yet. A lot of the ideas surrounding LLM seem to be more on the style of devs trying to latch on the latest trend without any proper thought of how useful/desirable LLMs can be on their product.
Also, while I don't share the ethical concerns surrounding how LLMs highjacked intellectual property, I think those are legitimate concerns. I have no issues with a dev taking a stance based on it, especially on a product I use for free.
unlike humans, of course.
An irreducible and unavoidable tendency towards committing copyright infringement and plagiarism are not features one generally wants in either a human author or a piece of software.
Should the world refuse to use LLMs because of that?
Still, there is some marketing damage control dismissal of LLMs in the note that could have been avoided.
Why Vivaldi won't follow the current AI trend(.) ---> the article explains why Vivaldi will not follow the trend.
Why (won't) Vivaldi follow the current AI trend(?) ---> someone's wondering and asking why Vivaldi will not follow the trend.
/s, because this is the internets.
The situation is analogous to war reporting. War hasn’t become more brutal (holding scale constant). We just see it more clearly. Similarly, I don’t see evidence we have more morons. They’re just given more attention.
Ok. But in browsers ?
Search datasets still are not really semantic search, and I'd wager many require online search if they required a well structured dataset, or a dumbed down query.
You only have to look to the relative popularity of something like Arc browser to determine there is a market for such things.
< 0.01% is not demonstrating a market
"underpaid".
I wonder what this has to do with the discussion on how LLMs work.
I'd like to know more why does Vivaldi team thinks these people were underpaid, how do they know how much per hour they were paid and what would be a fair price for the job according to Vivaldi's evaluation of the scope.
I'm really interested to know how did the Vivaldi team conclude that OpenAI forced unwilling people to do the work while paying less than what these people expected to receive for their time.
And most importantly. What does this have to do with LLM integration in browser?
No, the phrase they have written specifically talks about the RLHF workers and they do evaluate the pay of these people and somehow bring this issue into the discussion of how LLMs are useful to browsers.
And because I'm being downvoted for asking a simple logical question whether this has anything to do with LLM features in browser, I logically conclude that there are some things in current American culture that I literally (literally!) don't understand.
Just because we can doesn’t mean we should.
We're way beyond the days when every program you run has equal access to all your data. These days, different programs have different levels of data access.
https://www.zdnet.com/article/ethical-hacker-says-his-window...
You're basically exporting surveillance as an externality with recording technologies that by design collect everything around them. These technologies were basically multi-modal keyloggers.