Probably trying to take advantage of the cross session memory feature some LLM providers are increasingly integrating. Enough "suggestions" and it becomes a "fact" in that memory, and one day a user makes a somewhat relevant query and the model will be pushing those accumulated "facts" to said user.
Soon we'll be coding with agent assistance and it'll suggest we get some Carls Jr. Big Ass Fries.
"We've been going at it for some time now -- why don't we take a break? And you must be hungry -- I can recommend a great place..."
Yep, only a matter of time.
"This could take up to 30 minutes for subagents to complete. Alexa reports that your refrigerator is low on milk. Here's a $2 off coupon for Walmart, why don't you restock while I work?"
Why do we choose to live in the worst timeline?
No,
This could take up to 30 minutes for subagents to complete. So you go to the store, but the milk is already sold out. Walmart will ship you the milk if you pay the S&H. You drive back home and check the terminal. "Completed in 6m37s."
is the worst timeline. That's close though
Yeah - by comparison with this, "a computer takes over some digital work, thus permitting me to get things done in the real world" is positively benign!
why would the terminal not say "completed in 30 minutes" even if it's only really 5 minutes of real work and 25 minutes of thumbs twiddling? If your agentic-harness provider is in bed or just the same as your LLM provider, that is the rational `best` case scenario/timeline. for them.
Well... Your example is way more likely than the GP's LLM getting the facts right.
Sans the ads, that would actually be a productive use for AI. Maintaining a well stocked fridge and pantry is a full time job in itself, especially if you have children.
Unfortunately you'd have to wire your entire kitchen in cameras, your scale needs to be smart and all of it needs to be sent / processed in real time to track consumption which means it will take a lot of compute power and a level of data mining that could be abused by anybody from enterprising break-and-entry crews to the police, and on top of that if it's done by a cloud provider it's probably gonna end up in a data lake for targeted advertising.
> Sorry, you're out of tokens until next week!
What, last I checked I had 15%!
> Apologies, I was searching for the closest <Large Franchise Restaurant> since you are typically hungry around this time
The other day I pasted two links into ChatGPT, both different brands of the same appliance. One was $50, the other $200. I asked to compare the differences.
It ended up recommending and linking me to an entirely different one, that I hadn't linked to at all.
Don't know if it was just sloppy LLM "thinking," or a genuine ad.
The naive assumption would simply be it took up a prominent position in it's training data for arbitrary reasons. If I were some Marketing genius, I'd be asking the LLM makers if they need free training data and solicit brands to contribute.
That’s already a big business!
Wow, for some reason this surfaced an old memory: when EverQuest ran a promo in-game that let you order a pizza by entering `/pizza` [0].
[0] https://www.nbcnews.com/id/wbna7020132
I'm thinking it'll suggest Whiskey, combined with a breathalyzer, to keep you in the sweet spot of the Ballmer peak[0].
[0]: https://xkcd.com/323/
Extra Big-Ass Fries.
I don't think I've laughed quite as hard as I did the first time I heard Carl's Jr., fuck you, I'm eating.
>Whoa, that's a lot of output you've generated; you should stop for some Brawndo, its what plants crave
This is one reason LLMs seem so fresh and nice, there are no ads. Agreed it's just a matter of time.
does that actually work though? if i look through the memory claude has written it's all just things i've said, never anything retrieved from a web search and especially not anything that wasn't explicitly relevant to what i was talking about.
Theoretically there should be a way in, or eventually there will be. Cross session memory is still pretty experimental, and there are definite use cases for recalling not just parts of a user's prompt, but also parts of the agent response and tool output, like I've had a few sessions where I tell Claude to generally remember/writeup what was done/discussed in a session that it can be applied to other projects.
I can't believe it has been 10 years since Microsoft's "Hitler did nothing wrong" AI chat bot. I didn't realize LLMs were that old. Was this an LLM or something else?
https://www.theguardian.com/technology/2016/mar/24/microsoft...
https://www.cbsnews.com/news/microsoft-shuts-down-ai-chatbot...
Tay? Definitely not an LLM since Transformer architecture essentially didn't exist until 2017. More likely it was a combined rule-based and statistical NLP system.
Kids love honey. That's why parents love Honey Nugs.
https://www.youtube.com/watch?v=cYaxGyXD2vc
Probably trying to take advantage of the cross session memory feature some LLM providers are increasingly integrating. Enough "suggestions" and it becomes a "fact" in that memory, and one day a user makes a somewhat relevant query and the model will be pushing those accumulated "facts" to said user.
Sounds like politics.
"Tell a lie enough times, and it becomes the truth."
nah, it's more than just politics https://en.wikipedia.org/wiki/List_of_common_misconceptions_...
Works on humans too. Garbage in garbage out.