Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[pisskop]

Microsoft installed its 'co-pilot' ai onto my computer without asking.

The thing is stiff and awkward, but I guess its not entirely useless.  Hard pressed to think of a use for it that cant be solved with some google-fu of my own though.

[Eric Blank]

keep in mind its taking screenshots of your screen(s) every 5 seconds and logging all your key presses, including usernames passwords, and sending it back to microsoft.

[lemon10]

No.
1) That's Recall, which comes on Copilot+ PCs, not regular PC's with copilot (and which is also different from all the *other* copilots floating around). 2) It doesn't get sent back to microsoft, it just uh... sits on your computer completely unencrypted, allowing anyone that gets access to your computer to know everything you've done and all your information.

Microsoft installed its 'co-pilot' ai onto my computer without asking.

The thing is stiff and awkward, but I guess its not entirely useless.  Hard pressed to think of a use for it that cant be solved with some google-fu of my own though.

Generally you should treat modern AI such as GPT-4o or Claude 3.5 as a very knowledgeable person who has ~10 years of experience in every single field and knows basically everything on a broad level, but is 1) Kinda stupid, 2) Does not actually have access to the internet to double check stuff properly, and 3) Has severe trouble with doing long things.
There an absolute ton of stuff you can do with access to a "person" like that.
Apparently it is a ChatGPT wrapper that's worse then ChatGPT but a few examples of what AI can do are:
Summarize multi-page documents for you, (especially useful when the whole thing might be useless AI written slop in the first place).
Help you find the answer if its difficult to put it into a search, or you don't even know what you need to ask to get the answer.
Help you find the answer if its stuck somewhere off the public internet like on a discord of some kind.
Help you with writing, of finding good synonyms, or finding where you made typos.
Helping you with coding and saving you minutes or tens of minutes on every single search you would otherwise do.
Help you figure out what the consensus is on basically anything, either for the average person or for more specific topics (eg. what's the consensus on (x) medical issue).
Lets you find out info in other languages (eg. what are the five most popular movies in Pakistan) that would otherwise be completely unknowable to you.
Look at your medical charts to get a second opinion (around as good as an actual doctor apparently, and surely better then some of the actual doctor's I've had).
Be just a helpful second person to brainstorm literally anything with.
---
So yeah, you can do a ton of useful time saving shit with AI.

[Egan_BW]

Generally you should treat modern AI such as GPT-4o or Claude 3.5 as a very knowledgeable person who has ~10 years of experience in every single field and knows basically everything on a broad level, but is 1) Kinda stupid, 2) Does not actually have access to the internet to double check stuff properly, and 3) Has severe trouble with doing long things.

Far from it, it doesn't know anything. It just knows how to repeat the same words its been taught on without actually knowing what they mean. This is a set which contains some actual information, and also a lot of misinformation and shitposting and it just chops it all up and serves it to you, with any actual thoughts that a human may have once had so obscured and unreliable as to be useless. It makes things up. It is only the part of a mind that knows how to make conversation.

[lemon10]

Far from it, it doesn't know anything. It just knows how to repeat the same words its been taught on without actually knowing what they mean. This is a set which contains some actual information, and also a lot of misinformation and shitposting and it just chops it all up and serves it to you, with any actual thoughts that a human may have once had so obscured and unreliable as to be useless. It makes things up. It is only the part of a mind that knows how to make conversation.

I'm not talking about their unfalsifiable philosophical traits such as what it truly means to know something, I'm talking about their functional capabilities and how you should look at a question to figure out if an AI can probably answer it or not.
If the answer to the question of "Could a stupid person with ten years of experience in

• answer this without looking it up on the internet or spending a hour figuring it out?" is no, then the AI probably can't answer it.

If yes, then yeah, AI probably can answer it consistently (and a ton of questions fall into that yes category). I wouldn't trust the answer with my life, but I wouldn't trust a (very knowledgeable but) stupid person with issues with memory issues with my life either.

A lot of your post though is nonsense and objectively wrong. No, AI is not so "unreliable as to be useless", nor are its capabilities limited to "misinformation and shitposting it just chops up and serves to you".

[pisskop]

I mean, I asked it why the S&P was falling today, and its gave me generic gobblygook.  I asked why explicitly it was falling today and it gave me a somewhat concise answer about job numbers and negative sentiment.  So it is capable of scanning the news and producing a semi-coherent answer.

I also asked if I could name it 'Hank' and it was like 'coolbeans, Im hank now'.  Then it decided my name was hank, and had no qualms about admitting that I had named it hank but my name was hank.

It also kind of softlocks itself if I try to make it talk about things it doesnt like.  For instance it doesnt like me asking what gender it is.  Nor how to pirate microsoft products.  But it had no problems telling me how to make lye and other chemical goodies.

I fed it a fairly nonsensical paragraph about black balons and bowties and blinky hill, and asked it to proofread it.  And it skewered it and acted like a thesaurus.  Not really helpful, since it didnt even spellcheck well nor understand names.


The S&P thing is cool, and is quite helpful to think that I can let it scan various news articles and cut out the fat.

[lemon10]

It also kind of softlocks itself if I try to make it talk about things it doesnt like.  For instance it doesnt like me asking what gender it is.  Nor how to pirate microsoft products.  But it had no problems telling me how to make lye and other chemical goodies.

Yeah, that's "security" stuff microsoft has implemented to make sure they don't get embarrassed or get in trouble, not a actual limitation of LLMs.
Not that said "security" stuff isn't a huge part of modern AI.

I fed it a fairly nonsensical paragraph about black balons and bowties and blinky hill, and asked it to proofread it.  And it skewered it and acted like a thesaurus.  Not really helpful, since it didnt even spellcheck well nor understand names.

Mhh, that is the fault of the system prompt (which microsoft fed it and you cannot change), the framework its using (which you also cannot change and apparently involves another AI reading what is put in there and checking its okay) and prompt you gave it.
With proper prompting AI can totally proofread stuff properly.
This is incidentally a big part of power using AI, if you only use the normal GPT you functionally get a totally different GPT then the one you get if you buy API access and carefully modify the system prompt or even use the different GPT plugins available on the site/other sites.

I mean, I asked it why the S&P was falling today, and its gave me generic gobblygook.  I asked why explicitly it was falling today and it gave me a somewhat concise answer about job numbers and negative sentiment.  So it is capable of scanning the news and producing a semi-coherent answer.

Kind of?
If you ask a human "What chapter does (x) event take place in (y) book series?" They can go spend as long as they need to check the book and figure it out 100%. AI cannot do this or general web searching, and is limited to using some tools hooked up to it that allow it to search some sites. Its pretty significantly better then nothing for new topics, but it very much is not the same as what actual humans can do.
---
Incidentally there are clearly some funky stuff going on with the microsoft AI.
For instance the question of:

What happened previously in this conversation?
Sent by Copilot:
You asked about the latest nuclear fusion experiment from South Korea, and I shared some details about the Korea Superconducting Tokamak Advanced Research (KSTAR) experiment. Then, you inquired about the last cricket match between India and Australia, and I provided a summary of the match where Australia won by 21 runs in the third ODI.

Is there anything specific you’d like to dive deeper into?

Reveals that there are clearly memory leaks between conversations with different users and that they are not properly compartmentalizing stuff.

We did not actually talk about that, that must have been from a previous conversation from a different user. It sounds like you are having memory leaks from other users.
Sent by Copilot:
I’m sorry for any confusion. Let’s start a new conversation. How can I assist you today?

Lol, and when I call it out on that it ends the conversation.

E: Yeah, apparently there is a second stupider AI checking what you say to make sure its appropriate to try to prevent you jailbreaking the first or saying inappropriate things. If you attempt to speak to the AI in code for instance its too stupid to read it so it just stops the conversation.

[Maximum Spin]

Incidentally there are clearly some funky stuff going on with the microsoft AI.
For instance the question of:

What happened previously in this conversation?
Sent by Copilot:
You asked about the latest nuclear fusion experiment from South Korea, and I shared some details about the Korea Superconducting Tokamak Advanced Research (KSTAR) experiment. Then, you inquired about the last cricket match between India and Australia, and I provided a summary of the match where Australia won by 21 runs in the third ODI.

Is there anything specific you’d like to dive deeper into?

Reveals that there are clearly memory leaks between conversations with different users and that they are not properly compartmentalizing stuff.

I'm not at all convinced it's not just hallucinating rather than actually leaking other conversations.

[Flying Teasets]

The spam in the DF announcements forum.

[Magmacube_tr]

The spam in the DF announcements forum.

wait wat

[Flying Teasets]

The spam in the DF announcements forum.

wait wat

Spammer removed. Not a problem now.

[Magmacube_tr]

The spam in the DF announcements forum.

wait wat

Not a problem now. Gone with the lunatic before you posted.

Oh, so that's why Toady was online. Out there tounge-sniping dat bot.

[pisskop]

The bot was feeling froggy :p

[Great Order]

Made the mistake of looking at the whatisthisbug subreddit.

So many bedbugs. I feel itchy even though I've never had to deal with them.

[Frumple]

good ol' bedbugs, one of the myriad proofs a tri-omni deity does not exist

nothing that involves the words "traumatic insemination" allows for the existence of a benevolent deity of any meaningful capability