What the article barely touches upon is that all currently available AIs-as-a-service
- permit themselves to store and use your inputs essentially as they see fit, for essentially any purpose
- have mechanisms designed to "prevent abuse", without defining what that actually means
- are engineered to "keep you safe", without stating clearly what they want to keep you safe from, and without any option to disable those so-called safeguards
- have been carefully tuned to align their outputs with the Upper Middle Class U.S. West Coast Tech Scene political zeitgeist of the day, and that is what you'll get from them, even if it is completely inappropriate in your own cultural environment.
Caveat emptor indeed. LLM-as-a-Service vendors know their customer and this customer’s needs. It’s not you. They’re selling an API to companies for whom safety and governance are part of the value proposition.
As others have mentioned, the data retention aspect is specifically /not/ an issue with OpenAI (and other vendors’) APIs.
I don’t really get the saltiness so many on this site have towards LLM vendors. It just sounds entitled.
It's sort of frustrating to know that OpenAI has a better model that wouldn't cost them anything to release, but they don't do it for PR/whatever reasons.
What “better model”? If you’re referring to whatever LLM GPT-4 was before RLHF, by what measure might this model be better? As I pointed out above, what OpenAI is selling is “best” for their customers.
there is a particular type of person, who, when presented with a marvelous new thing (idea, product, person), focuses on its flaws rather than its unique merits.
we need them to ensure an equitable and sustainable world, but they do not drive progress. i bet they are also as a group far less successful and happy throughout their miserable lives. most will just impotently fart their misery unto others, and only a rare few will actually do something about it…
…at which point they wrap around back to becoming makers of imperfect things for the next round of criticism.
Agree with most of this, but your first point is not true of OpenAI. Their terms of service forbid them from training on API inputs unless you explicitly opt in.
And the prior poster is right, OpenAI API eula/tos clearly state data is not retained by default. I've been involved with legal counsel reviews, on that point there's zero ambiguity.
Whole point of ChatGPT is to collect data to do reinforcement learning, gp is talking about API access which has different terms (and wasn't banned in eu because of that).
> Whole point of ChatGPT is to collect data to do reinforcement learning
No, the point of ChatGPT is to provide a UI to chat with GPT models. Taking the data generated by users and using it for commercial purposes without consent is very likely to be illegal in the EU. Unfortunately they often take years to tackle this kind of stuff.
It does comply with EU GDPR laws -- EU users are allowed to opt out of training. This was after regulators in Italy shut it down. The US FTC is investigating too.
To comply with GDPR it has to be opt in, and choosing not to opt in has to not affect the app's functionality.
Currently, it's opt-out, and doing so destroys your chat history. They did manage to satisfy Italian regulators but I expect they'll get called out on this more thoroughly in the future.
Or maybe they just changed the way this works for Italy?
Nah. Give it 10 years and it'll be the Robinhood fiasco all over again.
Instead of anticipated trades, all these nontechnical power users with novel ideas will make the mistake of telling them to an AI-- and had their ideas stolen and rushed to market by whoever paid for access to mine the logs companies told us they weren't collecting.
We already don't trust cloud providers not to spy on our infrastructure. There is little consequence in lying to us. AI will bring us IP theft at scale.
Don't discuss anything sensitive or groundbreaking with cloud LLMs. No Google Collab, no OpenAI, nothing.
Make do with a local LLM in the form of a Vicuna 13/30B GGML. And even there, you have to watch out for backdoors in Gradio/posthog or weird models that insist on you enabling remote code execution. This domain is unbelievably shady and I can only assume it's because of the low-level access it provides to our collective stream of consciousness.
This is going to turn out to be bad news long term. Don't trust cloud LLMs.
> will always have an alignment, it's trivial to ground them in a different alignment though provided what you want to do isn't so egregious
The problem is that "egregious" is defined with respect to
> the Upper Middle Class U.S. West Coast Tech Scene political zeitgeist of the day […] even if it is completely inappropriate in your own cultural environment
It's easier to see where this goes wrong from the outside, as we all like to think of our social and political zeitgeist as the highest peak of refinement, far better than everyone else including our own past.
So, sexuality.
I live in Berlin; the KitKatClub isn't far from my current workplace, and I pass several billboards (some spinning cubes, some stationary) for Dildo King on the way to work, and what looked like an ad for a brothel on the way to a previous workplace, topless calendars are sold in full public display in normal shopping malls, and there is a neon statue of Leda and the Swan.
The cultural zeitgeist of Silicon Valley is an Overton Window (all zeitgeists are, it's not just voting), and that window includes Apple preventing me from connecting to adult content groups on Telegram and Discord even if I expressly say that is what I want to do.
(In a different direction, when I first visited the USA, a few years before the pandemic, I was shocked to find for sale in the supermarket an entire section of father-daughter valentines' cards, as I'm used to that day being just for couples).
Or consider alcohol. The US doesn't have a single attitude to alcohol (see map on https://en.wikipedia.org/wiki/Dry_county), but I can say my workplace kitchen has beer, rum, vodka, and sake, while some places in the world consider all alcohol to be illegal.
Or beef: steaks are part of my stereotype of American culture, eating it will get you lynched in parts of India.
This kind of thing happens often, with a great many taboos in one place being common in another.
And, barring an AI taking over and crystallising culture into only one form, it will happen again — many things currently forbidden will be allowed, and many things currently allowed will be forbidden, when Gen Z become the most common grandparents.
Pretty good examples and simple explanations. I didn't realize Claude 2 was so good at working with PDFs natively. I wonder if they're doing anything special? Is this just due to larger context length they have?
Also, biased opinion on my part: I'm especially interested in watching how these things affect data science and data literacy as a whole. Code interpreter is a game changer in my opinion, the most powerful tool that I think deserves all the press it is getting. Also: I released an open source code-interpreter for data (https://github.com/approximatelabs/datadm) and even though I know how to code and use Jupyter daily, I still find myself doing analysis with it instead.
All in all, it does seem like the different models and agents are gaining "specialization" skill is actually good for the user (rather than just using a single jack of all trades super chat model). Even though GPT-4 takes the language model crown, there's still specialization that matters and improves quality for different tasks as discussed here.
I wonder if in 2-5 years we'll all use "a single" AI chat interface for everything, or every specialization continues to "win at its own vertical" and we just have AI embedded inside of every app
I think they necessarily need to specialize, as certain information is only available in the context of the domain. I think bigger context windows will hit a limit, and you'll need to have actually trained and guided the AI on specifics of the domain to be useful.
At the moment it's only the fact that public documentation is available for so many tools that it's proving useful for so many things. But what about massive, closed source, boutique enterprise systems? You can feed it docs as context, but it would be better if it were trained on docs, support tickets and internal forums then properly aligned.
This will create an excellent search engine but a terrible reasoning machine.
There are a lot of ways to search through docs and support tickets now. The ability of an LLM to draw inferences and summarize all of that information comes from being trained on a very large amount of data with billions of parameters. The data can be highly specialized. There just needs to be several thousand gigs of it for the AI to do things that are rare and useful.
One other use case is — generating large amount of data on a scale not really possible before. I built https://meoweler.com - a travel site with all of the content generated by AI.
Nice article. I've been trying to stress to my coworkers, the bar for AI right now is not "it literally does my entire job for me with zero assistance or additional context". It's just a tool that if you learn to use correctly can dramatically speed up some tasks.
As an aside, Claude 100K looks very cool, but how many people even have access? Our CTO reached out to Anthropic directly they wouldn't even give him the time of day. It seems like if you aren't planning on spending 5 figures monthly on it it's a lost cause. I get it, but, lame.
One usecase that isn't mentioned here is call transcription. There are a lot of usecases where recording your own call and having some automated way to turn it into a transcript, and possibly generating some structured summary is a big time-saver.
OpenAi's Whisper is the most accurate transcription model that I know of. The weights are open-sources, so it can be self-hosted, or you use the API. The downside is that you have to roll your own diarization (seperating the text between speaker A/B). I used pynote audio.
Paid services like fireflies, and transcription tools built into your call software, are much easier to use, but lead to some transcription quality dropoff.
Maybe Phind is missing, at least for programming and in some cases obtaining technical information about something related to programming is really useful.
It’s not a wall, more like a gate. You can just click to continue reading. It’s a bit frustrating that it uses a dark pattern to make it seem like you are required to log in, but you can continue reading without logging in or subscribing or anything.
I'm not sure this is really "dark pattern" – it's a full screen modal, sure, but it's not like the close button is hidden or anything (and you can close with ESC as others mentioned).
It’s not “hidden” but the modal pops up after scrolling far enough down that you’re absorbed in the content, the “continue reading” is not styled as a button, and is accompanied by a chevron to the right which implies navigation away from the current content (I want to scroll down, not go right), which I think pushes it over the edge into dark pattern territory.
It's arguably in the user's interest if you look at the blogger and reader as participants in an ecosystem.
The current state of the ecosystem is such that getting someone to subscribe is extremely important for ongoing engagement, and ongoing engagement is often the prerequisite for continued writing.
If you're a reader interested in the kind of content the author is writing, and if you want to find more of this kind of content going forward, an easily escapable call to action is in the user's interest.
Is it annoying? Also yes. But in a world where everyone runs an ad blocker and social aggregators are fragmenting, it's better than a fully erected paywall, and better than nothing at all.
> It's arguably in the user's interest if you look at the blogger and reader as participants in an ecosystem.
I don't want to be part of the ecosystem. I want to read the article, leave, and never come back. It's not in this user's best interest to be bothered by a popup.
In fact, I believe it's actually the best interests of the author that's being looked after, not the users or readers, by using a modal popup to interrupt somebody's reading. It's as rude as walking up to someone while they're reading a book and waving your hands between their eyes and the book they're reading to get their attention if you noticed they were reading a book you personally wrote in an effort to sell them more books or to ask them for their email so you can send them special offers.
I suppose I could stop going to the park where authors think this is acceptable to do and limit my reading to parks where "ecosystem" authors avoid, but eventually other people start using their annoying tactics and you can't escape it no matter where you go.
We're part of that ecosystem whether we want to be or not.
As the beneficiaries of free content, it seems like a complete non-issue to just say "no thanks" when alternatives include: no content, or fully paywalled content. If you're just expecting free content that caters to you in every way possible way, I'm curious how this is sustainable for any author, or why authors should be expected to work this way.
> it's actually the best interests of the author that's being looked after, not the users or readers
There are no users/readers if there is no content. There is no content if there are no engaged users/readers. My point is that actively building an audience (good for the author) is actively good for the reader, because it makes continuing to write a viable thing for the author to spend their time on.
If you're just coming for a single article and you'll never return again, that's understandable and your prerogative, but you're now a double beneficiary: of the author, and of the readers who do return.
> It's as rude as walking up to someone while they're reading a book...
I couldn't disagree more. Perhaps the moment you pay for the blog post you'd have more standing to complain about the conditions surrounding its presentation.
And I'm also not saying the state of the ecosystem is good, or that I like it. I'm also not saying that the ecosystem can't or shouldn't change. But I think it's unreasonable to expect writers not to have self interests, while taking a stance that is wholly self interested.
> If you're just expecting free content that caters to you in every way possible way, I'm curious how this is sustainable for any author, or why authors should be expected to work this way.
speaking for myself, i mostly look for content that the author wanted to produce (just by/for himself). not for me or any engagement metrics.
> There is no content if there are no engaged users/readers.
this is plainly false
> If you're just coming for a single article and you'll never return again, that's understandable and your prerogative
not only is it my prerogative, it's the norm.
> Perhaps the moment you pay for the blog post you'd have more standing to complain about the conditions surrounding its presentation.
thou criticism is still allowed for the unwashed masses, not only because it's popular.
This really depends on the blog. If it's a person who is making their living off of a newsletter, then readers = content.
I'm not saying that there are no blogs without readers, and somewhat awkwardly was trying to make the point that an author taking steps to create an engaged audience may be doing so as a condition of continuing to write. This is obviously not universal. My personal blog has not many readers. I still write. I don't ask them to subscribe.
>> If you're just coming for a single article and you'll never return again, that's understandable and your prerogative
> not only is it my prerogative, it's the norm.
This quote cuts out the only point of that sentence: even if someone is visiting a blog with the sole purpose of leaving and never returning, there is no standing to be upset when the author providing them with content at no cost simply asks if they'd like to subscribe. This violates basic notions of reciprocity and seems rather...childish, frankly.
This is not to imply that there's any expectation that someone should subscribe. I rarely do.
I agree. But there are a few things which I think make this a dark pattern: the modal pops up after scrolling far enough down that you’re absorbed in the content, the “continue reading” is not styled as a button, and is accompanied by a chevron to the right which implies navigation away from the current content (I want to scroll down, not go right).
I've responded to other similar comments, but there are a few things which I think make this a dark pattern: the modal pops up after scrolling far enough down that you’re absorbed in the content, the “continue reading” is not styled as a button, and is accompanied by a chevron to the right which implies navigation away from the current content (I want to scroll down, not go right).
I agree there's a bunch of people who call anything they don't like a 'dark pattern', but I think this crosses the threshold for 'dark pattern' to be a reasonable description.
If someone's self hosted blog doesn't have a media query for dark mode, that's fine, but a platform that sells a itself as the authoring platform should make that minimal effort to prove a second colour palette.
[1] Maybe this has changed since I last read a medium article.
Same and it pops up only when you are in the middle of reading (on scrolling down) which was so distracting I didn't even continue reading.
Probably they know or measured that immediately showing pop up on a page load gets ignored and closed so they sneaked in the middle of reading, together with the small print dismission text they look pretty UX evil.
> These systems are built around models that have built-in biases [...] (if you ask it to create a picture of an entrepreneur, for example, you will likely see more pictures featuring men than women
- permit themselves to store and use your inputs essentially as they see fit, for essentially any purpose
- have mechanisms designed to "prevent abuse", without defining what that actually means
- are engineered to "keep you safe", without stating clearly what they want to keep you safe from, and without any option to disable those so-called safeguards
- have been carefully tuned to align their outputs with the Upper Middle Class U.S. West Coast Tech Scene political zeitgeist of the day, and that is what you'll get from them, even if it is completely inappropriate in your own cultural environment.
Caveat emptor.