I love (hate) that so very many of these language models say things like that when the truth is that the humans who made it, and/or the humans who use it are the actual responsible parties every single time. The models and the "agentic harnesses" that drive them are just software. If software "runs amok", then someone (human) did something wrong/bad somewhere along the way, either accidentally or purposefully. The model trying to take responsibility for human error is hilarious (and a bit sad/scary, because too many people will take it at it's word, despite it being a mindless machine with no actual agency beyond that which the humans provide it in the form of prompting and harness code).
It frustrates me no end that so many folks are so ready and willing to accept the hype and lies about what this technology actually is or can do, when what it actually is and can really do is already amazing enough on it's own even without all the ridiculous AGI/ASI anthropomorphising bullshit. Falling into this ridiculous "machine-god" hype-cult is kinda holding this technology back from it's true full potential, as everyone's all busy doin' stupid stuff it's not really capable of doing well, or designed for instead of focusing on using it for the (many) things it is really really good at doing (various really useful and powerful language, vision, and audio related tasks).
I initially found screenshots of this on Bluesky, figured I'd link to the original source given Threads seems to at least allow us to read without logging in.
With that said, in case Threads isn't available for whatever reason, I've uploaded screenshots of everything (I think?) here: https://imgur.com/a/mceW9WF
Mastodon's dependence on screenshots from other social networks is infuriating, enough that I left Mastodon altogether. Screenshots are text, they can be doctored in 15 seconds. Screenshots in the age of AI are utterly useless.
Yeah the AI part of their post is stupid, but they're right; screenshots can be easily faked (through e.g inspect element), so if your social media feed is mostly screenshots from other social media, you're at a relatively high risk of seeing fake posts. On the other hand you can't fake the content of a post or repost or the quoted content of a quote post, so if your feed contains only non-screenshot content you can be relatively certain that the content you're seeing is genuine.
I've never had that problem on Mastodon, but I can see how it could be annoying if you follow a lot of people who mainly post screenshots of people's posts from other social media sites. Though you could solve that problem by unfollowing those people...
Sure, but Mastodon doesn't embed anything and links from "the outside" usually don' work on other social medium unless you have an account - which leaves people unable to verify. Twitter users don't tend to be inspired by (or present on) other networks, so it goes:
X has FB and Insta screenshots
Insta has FB and X screenshots
FB has Insta and X screenshots
Bluesky has FB, Insta and X screenshots
Masto has Bluesky, FB, Insta and X screenshots
So what I'm saying is on Mastodon, you are exposed to the most screenshots of other networks, because there are no techniques to embed them and people refuse to link to them for the reason they aren't on those networks in the first place.
As for AI, i know toddlers who can prompt but are unaware of what a right click is.
Minor nit - imgur isn't available to UK based people[0] thanks to the Online Safety Act
[0] Which makes browsing Modrinth or Curseforge for Minecraft mods HILARIOUS - "you make a [extra dimension] portal like this: [image not available in your region]" "Gallery: [image not available in your region] [image not available in your region] [image not available in your region]"
It's not due to OSA. Ofcom are the regulator for the OSA, but Imgur was being handed a fine by the ICO in regards to how they handled children's data. It's the same parent company that develops Kik so I'm not super shocked...
I was allowed to read the original post without logging in, but it repeatedly showed me a dismissable pop-up asking me to download the app. Then when I tried to scroll down to read replies it showed me a non-dismissable pop-up asking me to download the app.
Holy shit social media is bad. I've been using Mastodon for so long I didn't realize how bad the others are.
Especially considering the user was mad about the agent doing stuff by itself, so it tried to "fix" this by sending an apology to the buyer without explicit approval of the person "running" the agent, seemingly understanding nothing from the conversation.
Wonder what quantization Meta runs these models on, Q4?
If it uses "memory" files like Claude Code there's a good chance it works.
If it was Claude I'd expect it to now put some form of this into every output even when not really related to the task: "No offers were accepted without consulting you and I haven't shared your address or availability."
Is it guaranteed to work? No.
And obviously it's a terrible idea to set up a chatbot to communicate, negotiate deals, and handle logistics on your behalf.
I don't think it's reasonable to really talk about the fact that in theory, AI can take a user message, save it in some random place, and ingest it later.
It's not in theory, this capability exists in practice, and you have no clue whether this specific instruction will work in 90, 99, or 99.9% of cases.
I think it's unreasonable to pretend that this couldn't possibly work, that you understand to what degree it does work, and that the only thing we should be discussing here was that it isn't deterministic, which every reader already knows.
Your argument exemplified by 'the ai can write to a file and use that information in context later' - implies a 'capability' to do theoretically do something, but not with any consistency at all.
Not in any scenario - even with human oversight - would be 'trust' any system to reliably work on the basis of arbitrary memory.
'It can possibly work' does not map to any reasonable conclusion that it will work with any degree of fidelity.
There are almost no examples where we rely on AI in this way. Chatbots for customer engagements etc. are merely thin language wrappers around deterministic systems.
It's 'possible' that Meta is doing something rigorous and deterministic but it's not remotely reasonable to make that assertion because we literally have the evidence right in front of us. That's literally what the article is about.
The Wright Brothers plane 'technically took flight' but it's not reasonable to conclude that it can fly passengers safely in any way, especially when the story is about a flight crashing.
It would be novel actually, if there were a story about Meta's unique use of AI in these kinds of information flows, that transcended what we all recognize as AI's inability to reliably process information, given what we know about AI and how it works aka 'it can write to memory'.
The AI is not ready for this _yet_, but it will be, and FB wanting getting ahead of the game here is potentially good business. It’s all in the public perception of utility vs fuck-up, and it’s far too early to say Zuck got that wrong, and indicators are he got that right.
It should be ready in about two more weeks! How many models have a Ph.D level intelligence now? I feel like I’ve been hearing that for about a year at this point.
I’m guessing you’re not a programmer. If you were, you’d have seen models go from “kinda helpful for programming” to “usable as a daily driver” about 9 months ago, for example.
“Models aren’t improving incredibly fast” seems a very odd point to be making.
How is this different from you not knowing what you are doing and adding products to your online store's DB for negative dollars? It's not, it's just easier.
It’s also a lot more fully-featured. It can list products in your store with a negative price and take out a mortgage on your house to finance fulfilling the orders. Hyperbole, but if you look at other threads right now it’s astounding that it’s doing things like intercepting iMessages via desktop notifications and then telling the user about them. Not sure why a notification about a notification with rewritten content is a feature that anyone would want, but consider what dumb-ass things it might do with this information. It’s selling your stuff on FB marketplace and arranges a meeting that you don’t show up to. “You’re right to push back! USER is actually getting his hemorrhoids checked at the doctor right now and wouldn’t be able to meet you. That’s on me!”
It opened for me on iOS but with those constant requests to make me download the app.
I just tried on my laptop though, both Waterfox and Firefox. It's just completely unusable there, with a ton of CORS and CSP errors so no CSS loads and no JavaScript loads. Happens even when I disable all forms of tracking protection and ad blocking; I think their site is just broken in a way which presumably happens to work in Google Chrome.
Say that you want to make a garage sale, but to get a better reach, you want to list item by item on marketplace. If you've never done it before, making 100 listings is such a pain in the ass you never want to do it again (I used to flip stuff for a living back in college, and it is a hassle)
At least with this, you can now just take pictures of everything, and tell the agent to research everything, list the stuff, and arrange for pickups.
Obviously not something I'd like to do for any serious deals or higher $ items. But for random stuff? I mean, it sort of beats hosting a garage sale or taking them to the flea market.
I should probably stop the auto-replies from claiming you're home when I can't verify that.
I wonder how much of this unreasonably risky and illogical behavior is a genuine property of LLMs and how much is there because LLMs come from "move fast and break things" startups.
I love (hate) that so very many of these language models say things like that when the truth is that the humans who made it, and/or the humans who use it are the actual responsible parties every single time. The models and the "agentic harnesses" that drive them are just software. If software "runs amok", then someone (human) did something wrong/bad somewhere along the way, either accidentally or purposefully. The model trying to take responsibility for human error is hilarious (and a bit sad/scary, because too many people will take it at it's word, despite it being a mindless machine with no actual agency beyond that which the humans provide it in the form of prompting and harness code).
It frustrates me no end that so many folks are so ready and willing to accept the hype and lies about what this technology actually is or can do, when what it actually is and can really do is already amazing enough on it's own even without all the ridiculous AGI/ASI anthropomorphising bullshit. Falling into this ridiculous "machine-god" hype-cult is kinda holding this technology back from it's true full potential, as everyone's all busy doin' stupid stuff it's not really capable of doing well, or designed for instead of focusing on using it for the (many) things it is really really good at doing (various really useful and powerful language, vision, and audio related tasks).
we need a Nitter for Threads
apparently people use Threads. I suppose the same kind of people who connect Muse to Facebook Marketplace.
I initially found screenshots of this on Bluesky, figured I'd link to the original source given Threads seems to at least allow us to read without logging in.
With that said, in case Threads isn't available for whatever reason, I've uploaded screenshots of everything (I think?) here: https://imgur.com/a/mceW9WF
https://rimgo.nohost.network/a/mceW9WF
Mastodon's dependence on screenshots from other social networks is infuriating, enough that I left Mastodon altogether. Screenshots are text, they can be doctored in 15 seconds. Screenshots in the age of AI are utterly useless.
This reads as anti-mastodon propaganda tbh. You never heard of Inspect Element, or saw screenshots of twitter on reddit and vice versa?
Yeah the AI part of their post is stupid, but they're right; screenshots can be easily faked (through e.g inspect element), so if your social media feed is mostly screenshots from other social media, you're at a relatively high risk of seeing fake posts. On the other hand you can't fake the content of a post or repost or the quoted content of a quote post, so if your feed contains only non-screenshot content you can be relatively certain that the content you're seeing is genuine.
I've never had that problem on Mastodon, but I can see how it could be annoying if you follow a lot of people who mainly post screenshots of people's posts from other social media sites. Though you could solve that problem by unfollowing those people...
Sure, but Mastodon doesn't embed anything and links from "the outside" usually don' work on other social medium unless you have an account - which leaves people unable to verify. Twitter users don't tend to be inspired by (or present on) other networks, so it goes:
X has FB and Insta screenshots
Insta has FB and X screenshots
FB has Insta and X screenshots
Bluesky has FB, Insta and X screenshots
Masto has Bluesky, FB, Insta and X screenshots
So what I'm saying is on Mastodon, you are exposed to the most screenshots of other networks, because there are no techniques to embed them and people refuse to link to them for the reason they aren't on those networks in the first place.
As for AI, i know toddlers who can prompt but are unaware of what a right click is.
Minor nit - imgur isn't available to UK based people[0] thanks to the Online Safety Act
[0] Which makes browsing Modrinth or Curseforge for Minecraft mods HILARIOUS - "you make a [extra dimension] portal like this: [image not available in your region]" "Gallery: [image not available in your region] [image not available in your region] [image not available in your region]"
Sounds like a local government problem for y'all to solve, not a problem for the rest of the world to work around
It's not due to OSA. Ofcom are the regulator for the OSA, but Imgur was being handed a fine by the ICO in regards to how they handled children's data. It's the same parent company that develops Kik so I'm not super shocked...
Oh, well that changes things.
I was allowed to read the original post without logging in, but it repeatedly showed me a dismissable pop-up asking me to download the app. Then when I tried to scroll down to read replies it showed me a non-dismissable pop-up asking me to download the app.
Holy shit social media is bad. I've been using Mastodon for so long I didn't realize how bad the others are.
and this is the least worst. X only shows you one post and Reddit won't show you anything.
It will be chaos if people let AI agents run amok with their accounts.
Many more such cases are to come.
Bonus score if your Meta AI Agent does something with your account that the Meta AI Moderator deems inappropriate and bans you.
"that's on me"
Nice touch by the mechanical parrot, to worthlessly owning it.
Especially considering the user was mad about the agent doing stuff by itself, so it tried to "fix" this by sending an apology to the buyer without explicit approval of the person "running" the agent, seemingly understanding nothing from the conversation.
Wonder what quantization Meta runs these models on, Q4?
That jumped out at me too. It’s hilarious to see that response so often when AI makes a mistake
AI has seen how much humans lap up content that is #relatable and started acting like that
The AI is not remotely ready for this - this is an absolute delusion they are selling.
"By the way don't do this again" <- as if the AI has the ability to ingest and systemically diffuse this.
I think Zuckerberg himself is deeply into the Koolaid, and is likely himself unaware of the limits of this tech.
He's probably surrounded by enablers.
If it uses "memory" files like Claude Code there's a good chance it works.
If it was Claude I'd expect it to now put some form of this into every output even when not really related to the task: "No offers were accepted without consulting you and I haven't shared your address or availability."
Is it guaranteed to work? No.
And obviously it's a terrible idea to set up a chatbot to communicate, negotiate deals, and handle logistics on your behalf.
"If it uses "memory" files like Claude Code there's a good chance it works."
You have explained literally why it would not work.
The AI can absolutely not depend on 'arbitrary statements in some file' as operational policy.
For a very, very narrow scope of work, when it's well defined, when the information is rigorously applied, sure ...
But they don't have that.
The are throwing agents out there like they can handle this degree of complexity and nuance, when they cannot.
100% failure rate over any period of time.
I didn't say it would work with high reliability, or that it would fix this clearly unsuitable use case.
However, it is untrue that it doesn't have the capability to memorize an instruction and diffuse it to new sessions.
Simply saying "do not ever do this again" can result in the behavior not reoccurring with any likelihood.
You'd have to benchmark whether with the instruction in place it would violate it, and in how many cases, so that you can understand the risk better.
I don't think it's reasonable to really talk about the fact that in theory, AI can take a user message, save it in some random place, and ingest it later.
I think we get that.
It's completley unreliable, which is the issue.
It's not in theory, this capability exists in practice, and you have no clue whether this specific instruction will work in 90, 99, or 99.9% of cases.
I think it's unreasonable to pretend that this couldn't possibly work, that you understand to what degree it does work, and that the only thing we should be discussing here was that it isn't deterministic, which every reader already knows.
No, this is upside down.
Your argument exemplified by 'the ai can write to a file and use that information in context later' - implies a 'capability' to do theoretically do something, but not with any consistency at all.
Not in any scenario - even with human oversight - would be 'trust' any system to reliably work on the basis of arbitrary memory.
'It can possibly work' does not map to any reasonable conclusion that it will work with any degree of fidelity.
There are almost no examples where we rely on AI in this way. Chatbots for customer engagements etc. are merely thin language wrappers around deterministic systems.
It's 'possible' that Meta is doing something rigorous and deterministic but it's not remotely reasonable to make that assertion because we literally have the evidence right in front of us. That's literally what the article is about.
The Wright Brothers plane 'technically took flight' but it's not reasonable to conclude that it can fly passengers safely in any way, especially when the story is about a flight crashing.
It would be novel actually, if there were a story about Meta's unique use of AI in these kinds of information flows, that transcended what we all recognize as AI's inability to reliably process information, given what we know about AI and how it works aka 'it can write to memory'.
Had that problem here. Our product lead thinks it’s infallible and is surrounded by yes men.
Day one our agentic platform goes live it causes a reportable compliance issue. Massive clean up. Reputational impact. Turned off.
No one held accountable still. Assuming that adding more guardrails will fix everything.
You should write about that, because it's the primary systems failure of AI right now.
'Managers Delusion' - which includes techies as well, to be fair, at least they have the excuse they are one step removed.
I'd rather everyone burned for it than heeded a warning.
They are morally bankrupt and will try and do it again and again if it's stopped too early.
They're building for future capabilities. It's a logical approach to leveraging the data they have.
The AI is not ready for this _yet_, but it will be, and FB wanting getting ahead of the game here is potentially good business. It’s all in the public perception of utility vs fuck-up, and it’s far too early to say Zuck got that wrong, and indicators are he got that right.
It should be ready in about two more weeks! How many models have a Ph.D level intelligence now? I feel like I’ve been hearing that for about a year at this point.
I’m guessing you’re not a programmer. If you were, you’d have seen models go from “kinda helpful for programming” to “usable as a daily driver” about 9 months ago, for example.
“Models aren’t improving incredibly fast” seems a very odd point to be making.
Fake surprise - it's obvious that would happen, he did it deliberately for clicks and engagement.
It's plausible, but there's every reason to believe that someone would trust the technology handed to them by a megacorp to do as advetised.
The 'deliberate failure' is on Meta here.
Or he just wanted to test if it works for clicks and engagements, and he got a result he didn't anticipate.
Classic case of 'AI gonna AI.' Always review your automated systems' outputs, especially before they hit public channels.
How is this different from you not knowing what you are doing and adding products to your online store's DB for negative dollars? It's not, it's just easier.
It’s also a lot more fully-featured. It can list products in your store with a negative price and take out a mortgage on your house to finance fulfilling the orders. Hyperbole, but if you look at other threads right now it’s astounding that it’s doing things like intercepting iMessages via desktop notifications and then telling the user about them. Not sure why a notification about a notification with rewritten content is a feature that anyone would want, but consider what dumb-ass things it might do with this information. It’s selling your stuff on FB marketplace and arranges a meeting that you don’t show up to. “You’re right to push back! USER is actually getting his hemorrhoids checked at the doctor right now and wouldn’t be able to meet you. That’s on me!”
Wow that Threads website is absolutely unusable on mobile
I don't much care for it on desktop either
Not sure how many weeks? or months? since I last visited it, but Firefox with UBO just displayed this perfectly. A low bar, I know.
I don't know about you, but it literally throws me back to HN when I click on the link. With a prompt to download the app.
It opened for me on iOS but with those constant requests to make me download the app.
I just tried on my laptop though, both Waterfox and Firefox. It's just completely unusable there, with a ton of CORS and CSP errors so no CSS loads and no JavaScript loads. Happens even when I disable all forms of tracking protection and ad blocking; I think their site is just broken in a way which presumably happens to work in Google Chrome.
I can see some use for it, in the future.
Say that you want to make a garage sale, but to get a better reach, you want to list item by item on marketplace. If you've never done it before, making 100 listings is such a pain in the ass you never want to do it again (I used to flip stuff for a living back in college, and it is a hassle)
At least with this, you can now just take pictures of everything, and tell the agent to research everything, list the stuff, and arrange for pickups.
Obviously not something I'd like to do for any serious deals or higher $ items. But for random stuff? I mean, it sort of beats hosting a garage sale or taking them to the flea market.
I'm not all pessimistic about this.
Not sure if this is just rage baiting or people actually think this may be a good idea.
I wonder how much of this unreasonably risky and illogical behavior is a genuine property of LLMs and how much is there because LLMs come from "move fast and break things" startups.