I can. None of this tech will maintain Google as the chokepoint of internet and make billions for them. You can download and run a LLaMA or Mistral but can't download a Google. On-topic information has been commoditised.
Even OpenAI is in a bad spot. They owned the whole LLM mountain in 2020, now they only own the peaks. Almost the whole mountain has been conquered by open models. Their area of supremacy is shrinking by the day.
The two factors at play here are 1. ability to run LLMs locally and 2. ability to improve local LLMs with data exfiltrated from SOTA LLMs.
I foresee an era of increased privacy as a consequence of this. We will be filtering our browsers with LLMs to remove junk and override Google, Twitter and FB. Bad news for ad providers, they got to move into some other profitable line of business now. Nobody got time to read garbage and ads.
Right now GPT4 is useful/powerful and probably anything else sucks. That may change but the mountain has real value at the top only. At least for
chat used for either human assist or decision making or code generation. Categorisation, yeah you can use lesser LLMs.
Not necessarily true anymore. As a counterexample, consider CodeLlama 33b, which is quite good (and which has replaced GPT-4 for my coding assistant needs).
OpenAI's models are likely to remain the best, but I see open models catching up and becoming "good enough." Why pay for GPT-4 when I can run a model locally for free? (Barring the initial capital cost of a GPU; and not even this if you're, say, using a Macbook)
Like yes, as long as GPT-4 remains the most useful (and now they have multi-modality in Plus so they've just extended their lead) nobody will adopt anything else.
So grandparent's argument makes even less sense as search should also be fungible but it isn't when Google still provides the best search product and has since it started. If OpenAI can maintain that lead and continue to ship improvements that continue to push their peak higher, most people will not pay for anything less even when that price is free.
I don't work in tech and I use LLaMa. It is great for giving me ideas, writing overviews which I use to make sure I have most of my bases covered (and helping me see which of my ideas are more original), getting more examples, along with other creative tasks.
Could I get my employer to pay for it if all free options disappeared? Probably, but I don't have to while they exist.
> We will be filtering our browsers with LLMs to remove junk and override Google, Twitter and FB ... Nobody got time to read garbage and ads.
Exactly, there is a paradox that is getting more extreme by the day - that social media (and the broader web) is a wellspring of knowledge, yet also a vortex of addiction, filter bubbles & lost productivity. I want one without the other.
This pushed me to start building open source at OpenLocus, contributions & feedback are welcome - details in other comment.
Does anyone know vision models that detect ads? I wasn't able to find any, which is kind of surprising... That sounds like an armsrace that would exist.
I tried something not dissimilar, but without AI models: having webpages rendered in a real browser, but "headless". But there was really no way for the webpage to detect it was "headless": I'd render in a real browser, in an actual X session, under Xephyr. And I wouldn't be showing that Xephyr X server (so I wouldn't see what that real browser was rendering).
I'd then put gray rectangles on what where ads and only show visually that, with the grey rectangles, after having covered the ads with gray rectangles.
I just did it as some quick proof-of-concept: there are plenty different ways to do this but I liked that one. It's not dissimilar to services that renders a webpage on x different devices, without you needing to open that webpage on all these devices.
But the issue is that while it's relatively easy to get rid of ads, it's near impossible to get rid of submarine articles/blogs and it's getting harder and harder by the day to get rid of all the pointless webpages generated by ChatGPT or other LLMs that are flooding the web.
Meanwhile sticking to the sites I know (Wikipedia / HN / used car sales websites / a few forums I frequent etc.) and running a DNS locally that blocks hundreds of thousands of domains (and entire countries) is quite effective (I run unbound, which I like a lot for it's got many features and can block domains/subdomains using wildcards).
I'm pretty sure detecting and covering ads before displaying a webpage can be done but I'd say the bigger problem is submarines and overall terribly poor quality LLM generated webpages.
So basically: is it even worth it to detect ads while the web has now got a much bigger problem than ads?
Absolutely its still worth blocking ads. They abuse my compute rendering client side. Off with their heads. Presumably you could use ai to sniff out putative ai content in text and flag or block that too.
It won't work as long as you have HTTPS preventing any large scale interception and analysis of traffic. This is why Google was so quick to promote and switch to HTTPS. I bet you my left nut right now there is some random ex Googler with an NDA and interesting story to tell about this topic.
It pretty much means you have to detect the ad locally. And by then you've already lost and transferred it down at least.
I agree heartily with you. OpenAI may be the best, but the open models are getting good.
A weak analogy: as you might take a hybrid cloud/local compute strategy, I think it makes sense to be very flexible and use LLMs from different sources for diversity and not getting locked in. I mostly use OpenAI, but I am constantly experimenting with options. Local options are most exciting to me right now.
I usually use APIs, but OpenAI’s app that supports multi modal input images and voice conversation is impressive and points to a future mode of human computer interactions.
> We will be filtering our browsers with LLMs to remove junk and override Google, Twitter and FB
I cannot wait for the day where I can create a list of blocked terms "Musk, Trump, American politics, ..." and have an intelligent LLM filter so that it doesn't matter what site I visit, whether The Verge, or The Guardian, or Reddit, all articles or posts related to these terms will be gone.
I think this has a chance to have a far higher effect than just ad tech. It's going to effect all media publishers. But I also expect them, led by Google and Facebook, to fight this tooth and nail with every dirty trick they can.
Great start. Please develop a plugin I can secretly add to my Mother's browser to dial down the outrage she reads on the fly! She'll notice 'redacted' ;)
The main issue is that its very slow and expensive to browse the internet like this. The LLM will only perform well if you have it do chain of thought reasoning, and that has a latency hit because of a longer generation.
This is very interesting desire to me.. Wont you end up with a weird, contrived context for everything else you do read though? Like you would be reading an article about the UN and it would say: "and then, some country put forward a new resolution." Wouldn't that, like, kinda drive you crazy?
Like I know the state of journalism is less than stellar, but patching it after the fact for each reader seems like the wrong direction. The implicit conception of "the news" in this desire reifies it into a weird kind of commodity for your personal entertainment/edification; which is precisely the conception operating today which makes it so bad!
Like, maybe, if you have psychological considerations where certain triggers are very damaging, I can kinda understand this. But if that is really the case, then just why read the news anyway? Of course you gotta read some sometimes, but in general you can read other things. There is a lifetime and a half of fiction and nonfiction to read, no GPU required!
I don't want to remove all mentions. If I'm reading an article about something else and it happens to mention something on my list, that's fine. I just want to remove all the top level articles and posts about these things.
It's not about being triggered by anything or trying to hide from anything. I also don't need (or even want) 100% efficacy. It's about cleaning up noise. For example, at this point I'm fully aware that Musk has turned Twitter into even more of a cesspool. I don't need any more information about that. And yet I get it, all the latest "juicy Musk gossip" any time I go near any tech sites. And it's just noise to me at this point.
Same with American politics. I'm not from the US so I'd be happy with a short monthly synopsis on what's happening there. But on the English speaking web, American politics is everywhere. It's exhausting. I want to filter it, reclaim the attention it steals from me while still being engaged with online society to a degree that I choose. And I believe that reclaiming this attention, energy, and time would allow me to engage more with subjects I do care about.
I see, that totally makes sense, but in that case it doesn't seem like you need a fancy AI for this, or at least it feels way overkill for what you actually want here.
patching it after the fact is definitely needed. same way I use a plugin to filter out curse words because when it comes to your brain, garbage in = garbage out. patching "news" stories after the fact will never be perfect though because half the problem with "journalism" is omission of facts, not just biased half-truths or opinions.
Explains why Google tried (or is trying) so hard to push WEI though. If they can't programmatically evade adblockers, might as well programmatically prevent them.
I do use a feed reader already but it's not good for discovering new things. The problem is when I venture out to discover what's going on in the world I'm deluged with Musk spam. A couple of years ago it was Trump spam.
Agreed. I use
- A feed reader with multiple sources
- On X, I started by muting words and accounts. I now only watch one highly curated list
- Here
- podcasts
- Engaging with relevant people directly
I'm sometimes a bit late on new trends, but I spend more time on long forms and reinforce a meaningful network in the process.
We started work on an open-source LLM-powered media filter at OpenLocus (https://github.com/openlocus), aiming for an alpha release by late 2023.
In the more medium term we are collaborating on improving information overload, filter bubbles & misinformation with labs at AllenAI, CMU, UPenn & Utah.
Contributions and feedback are welcome! Feel free to hit me up for an early access - email in the profile.
But how can we possibly be this optimistic though? Like, time and time again there are these little glimmers of fantasy that some technology alone can alleviate the heinous/annoying parts of a mature capitalism, and every time the technology has ended up, after all is said and done, on the other side (so to speak).
Like, go back and read old mailing lists. See how everyone was so assured that computers themselves, then cryptography, then the modern internet, would create radical changes in the way the economy works, and enable avenues for greater self-determination and happiness. They all seem so naive in retrospect, knowing what we know and how it all played out.
I love that so many people are these days inspired by the llm tech to imagine a better world with them, and its important in itself to inspired like this. But precedence does not favor putting all your hopes in solely the technology itself for a better world. Hope I am wrong though!
It will undoubtedly be a turbulent time. All the skills and knowledge expressed in text or images is compressed into models. And these models offer back all they know to each of us according to our needs. Learning from everyone indiscriminately, serving everyone specifically, open sourcing skills and knowledge to a much deeper degree than open source had on software.
This will have social implications, it will speed up the learning process of any new technology. For example by upgrading Copilot, Github can upgrade coding practices everywhere (where it's used) to new coding standards instead of taking years, that's how fast it can go.
On another line of thinking, an image-language-action model can be used to control robots, and robot hardware is getting accessible. There is a chance for inventing self reliance automation for people, a possible solution to job loss.
Right, there are lots of things it could do, but the point is that these capabilities, or rather this potential, is merely necessary but not sufficient for actually following through with those nice things.
It doesn't matter what it can do when there is a truly scary amount of power and interest directed towards, e.g., making sure job loss (or unemployment in general) doesn't change too much. They'll make these things illegal before they'd let it get even close to messing something like that up!
There are already so many innovations and pieces of technology that could be helping instead of hurting. It should seem clear that the sheer capability of something could never be enough to change the state of affairs alone.
Sure the printing press was critical, but there was also a whole lot of blood shed and tumultuous times before the enlightened subject of the printing press could enjoy their new class and literacy.
I can. None of this tech will maintain Google as the chokepoint of internet and make billions for them. You can download and run a LLaMA or Mistral but can't download a Google. On-topic information has been commoditised.
Even OpenAI is in a bad spot. They owned the whole LLM mountain in 2020, now they only own the peaks. Almost the whole mountain has been conquered by open models. Their area of supremacy is shrinking by the day.
The two factors at play here are 1. ability to run LLMs locally and 2. ability to improve local LLMs with data exfiltrated from SOTA LLMs.
I foresee an era of increased privacy as a consequence of this. We will be filtering our browsers with LLMs to remove junk and override Google, Twitter and FB. Bad news for ad providers, they got to move into some other profitable line of business now. Nobody got time to read garbage and ads.