Maybe someone with more time at hands could look up what Google said with respect to ads and what happened later.
This is one of the rare instances where it's very easy to predict the future: the prompt auction market will look similar to the existing online ad market, financial firms will pay for prompt streams for sentiment analysis, companies and interest groups will pay to have their products or agenda included favorably in the training data for future open weights models... any way you can think of that LLMs can be monetized, you will see it happen. And fast. The financial pressure is way too high for there to be too long of a honeymoon phase like we had with web 2.0
And how much trust are you going to have with your model results that they haven't been transformed and adjusted by advertising priorities?
search engine results do this all the time, reordering output by advertiser input. its a pretty small jump from that to rewriting output from models, and even better where its all a black box.
oh absolutely, its been a progression though and search order rewriting was implemented very early on as part of ad integration. its common in search relevancy/tuning circles
your example with google isn't necessarily applicable now because they've shown a roadmap that can be done and squeezed down tightly between the "hey, we're good folks" to "you're our captive cattle, we can do whatever the fuck we want. there's nothing you can do, since all our competitors will be doing the exact same thing shortly"
In what way would that be securities fraud? I guess you could get nailed under Section 17(a), but really hard to make a case they're defrauding investors by representing they were going to make ads worse performing than they ended up making them.
In order for it to be securities fraud it has to be tied to a securities transaction and the misstatement has to be material to a reasonable investor's decision.
I think they said the ad vendors wouldn't but the matching algorithm would still be aware of it. Which IMO is the bare requirement to have ads be anything but magazine style ads.
I mean, the ad doesn’t necessarily have to be made aware of the exact prompt context, just that the ad itself was relevant. You can basically have the ads prequalified for areas and serve them when relevant. Now that does show the user is talking about something relevant most likely, and depending on how they decide to serve them or provide referring, it may traceable to a profile/identity built for that user externally.
I’d be more concerned as to how this ends up in agent platforms using the LLMs, when you don’t have a fairly autonomous agent based system using these the entire point is that a human isn’t involved, so who are you serving ads to and where are you injecting them.
Moreover, if you are injecting them everywhere, does that survive stare for subsequent steps, meaning from the first set of results I get, does that loop back in again with the ad injected into the context. Because now, we have yet another dangerous way of injecting instructions into an already issue prone surface area.
I’m guessing they’re going to have special APIs that don’t include ads, and those are going to cost more, especially for non embedded agents (processes that already exist inside ChatGPT that kick off transparently from prompts, like asking it to work with an office document). After all the customers using agents aside from developers are mostly businesses, so it’s where the money is. The ads will exist for the poor to subsidize their use, and probably create even more barriers for agentic use like I described. Just my thoughts.
And good luck litigating against any business in this administration. Unless they explicitly tick off certain people or refuse to kiss the ring, they can get away with almost anything right now and there’s little risk of doing it or not because ticking off this admin will raise illegitimate prosecution even if you’re perfectly legal, almost the same level of if you’re not. It’s the ideal playground for doing all sorts of manipulation, just kiss the ring and you’ll be fine.
I don’t know that there were any promises anyway. But if there were, then an investor could have plausibly believed that that was a better long-term business model.
It’s early days for these LLM hosts, maybe investors could be worried about taking the really annoying business notes before users are properly addicted.
It would also be a huge security risk. But I can't think of any fundamental difference with Google queries, other than the sheer entropy of user data involved.
And I'm not a tinfoil internet anarchist, but just because Google only leaks user data in aggregated form to advertisers, doesn't mean that they don't leak their user data, it's just that they did so in a legal and responsible manner.
Maybe considering the difference in data volume and intimacy between queries and AI conversations, the privacy implications of advertising merit a difference in treatment, but I wouldn't be surprised if that is lost to a more simple 'Google did this so we can do it too' momentum.
you can use chatgpt without an account, just not all of it
and you can't make full use of Google without an account. for example, you need an account to upload to YouTube, manage your website in search, place ads, opt out of data usage. the list goes on
And you can also search on google with an account, and your queries are stored for you to see right? I'm pretty sure I can see a history of my searches.
This is a classic example highlighting the upside of local llms.
However the local llms I can run on reasonable hardware are so dumb compared to opus, and even if I shelled out five figures of hardware to run the largest/smartest open model it still will be noticeably worse.
Right now the remote models are just so much smarter and more affordable under most usage patterns.
I find myself wondering what people would think if we swapped "Israeli" for "Chinese" in your reply. And why this double standard exists in all our minds.
Clearly you can't run a country when your elites owe their first allegiance to somewhere else.
> No internet open browsing or open chat features.
> AI toys shouldn’t need to go online or talk to strangers to work. Offline AI keeps playtime private and focused on creativity.
> No recording or long-term data storage.
> If it’s recording, it should be clear and temporary. Kids deserve creative freedom without hidden mics or mystery data trails.
> No eavesdropping or “always-on” listening
> Devices designed for kids should never listen all the time. AI should wake up only when it’s invited to.
> Clear parental visibility and control.
> Parents should easily see what the toy does, no confusing settings, no buried permissions.
> Built-in content filters and guardrails.
> AI should automatically block or reword inappropriate prompts and make sure results stay age-appropriate and kind."
Obviously the thing users here know, and "kid-safe" product after product has proven, is that safety filters for LLMs are generally fake. Perhaps they can exist some day, but a breakthrough like that isn't gonna come from an application-layer startup like this. Trillion dollar companies have been trying and failing for years.
All the other guardrails are fine but basically pointless if your model has any social media data in its dataset.
I'm sure you are correct about being able to do some clever prompting or tricks to get it to print inappropriate stickers, but I believe in this case it may be OK.
If you consider a threat model where the threat is printing inappropriate stickers, who are the threat actors? Children who are attempting to circumvent the controls and print inappropriate stickers? If they already know about topics that they shouldn't be printing and are trying to get it to print, I think they probably don't truly _Need_ the guardrails at that point.
In the same way many small businesses don't (most likely can't even afford to) opt to put security controls in place that are only relevant to blocking nation state attackers, this device really only needs enough controls in place to prevent a child from accidentally getting an inappropriate output.
It's just a toy for kids to print stickers with, and as soon as the user is old enough to know or want to see more adult content they can just go get it on a computer.
ChatGPT allegedly has similar guardrails in place, and now has allegedly encouraged minors to commit self-harm. There is no threat actor, it's not a security issue. It's an unsolved, and as far as we know intrinsic problem with LLMs themselves.
The word "accidentally" is slippery, our understanding of how accidents can happen with software systems is not applicable to LLMs.
North Korean defectors are well known for being unreliable sources. They rarely have skills nor social connections and thus are massively incentivized to join the existing markets for anti-NK propaganda in both the West and India.
The kind of obvious propaganda like "it's a crime to have the same haircut as Kim Jong-Un" (or to not have it, depending on the source). No one is saying life there is great, but there is a track record of fantastically untrue stories.
If they rarely have skills or social connections, how were they able to achieve the feat of escaping from the Hermit Kingdom in the first place? It seems to me that this high bar selects for people who can thrive in adverse circumstances.
reply