Chatgpt - Seriously good potential (or just some Internet fun)

[Mesai]

so the new 'apple intelligence' sounds rather insidious functionality -
Despite apple pleading for respecting security of your data, would you want all of you data emails/photos/searches/purchases/(calls?) 'crunched' by a LLM to allow it to subsequently help you,
like clippy the paperclip, your data also being potentially processed by 3rd party OpenAI

etc https://appleinsider.com/inside/apple-intelligence

Most of it is processed on-device, and the Private Cloud code is open source (whether it's what they're actually using is another story).

You need to enable ChatGPT yourself, so it's exactly the same as anyone using it for anything, except there is some phone integration to make things easier.

 

[Begbie]

That's why responses seriously need to be able to cite courses so you can then go to verify various aspects of the answer

Chat gpt can cite sources?

Unsure how accurate it is mind.

 

[Freefaller]

That's why responses seriously need to be able to cite courses so you can then go to verify various aspects of the answer

I've helped deploy it to search our knowledge base and lessons learned from other major projects. Works really well for our needs it's very basic but at least it's in our own ecosystem and it automatically links to sources (even to external data repositories - all legit that we are connected to).

But any work I agree sources should be cited so some due diligence can.be undertaken. Especially with anything that is related to decision making.

 

[jpaul]

Do the likes of IEEE explore,and professional databases allow their material to be trawled externally ?

At home if google is showing an AI censored set of replies, like for the Chinese citizens, where will you go who is number one.
'Who controls the past controls the future. Who controls the present controls the past.'

 

[Freefaller]

Do the likes of IEEE explore,and professional databases allow their material to be trawled externally ?

At home if google is showing an AI censored set of replies, like for the Chinese citizens, where will you go who is number one.
'Who controls the past controls the future. Who controls the present controls the past.'

Many professional institutions do have a knowledgebase that is accessible (either via APIs etc..). There are also many associations that help to curate knowledge i.e. learning legacy for major projects etc... however it is still scattered all over the place, but the sign posting is getting better.

 

[UberTiger]

Chat gpt can cite sources?

Unsure how accurate it is mind.

Bing Copilot can, here's an example:

 

[dowie]

That's why responses seriously need to be able to cite courses so you can then go to verify various aspects of the answer

That's quite easy to do, just change the prompt slightly, definitely need to verify as you suggested and for sure sometimes models will hallucinate the citations.

IIRC some lawyer got in trouble after thinking he was smart getting ChatGPT to write an argument... not realising it had cited some completely made-up cases.

 

[Minusorange]

Going to suck being any remotely popular youtuber or streamer with people making fake videos of you doing all kinds of weird

 

[SheepBeast]

LLM-based AI is very good at sounding convincing and natural but half the time it is feeding you absolute bull****. Just a waste of time if I have to reach for google to verify everything it says anyway.

 

[jpaul]

Even in the case where chatgpt/copilot could make citations, you'd need to read those to conclude if it precis'd those accurately and you thought they were comprehensive.

for the video fakes I'm not going to pay a cinema ticket to see if bleeding edge tech is good for harrison ford

Indiana Jones and the Dial of Destiny | Official Trailer

Indiana Jones and the Dial of Destiny in theaters June 30.Harrison Ford returns as the legendary hero archaeologist in the highly anticipated fifth installme...

youtu.be

https://www.wired.com/story/indiana-jones-and-the-dial-of-destiny-de-aging-tech/

 

[PlacidCasual]

Our business is using a Microsoft "AI" product in the cloud. The one area where it could be really revolutionary is the unknown knowns. Where our engineers don't know what data or information exists and the "AI" tool can point you in the right direction. Like "how do I maintain this pump?" It then gives you the detail but also provides a link to the original source of information. Being a large old company with poor knowledge management we have so much information but badly managed and we end up having to relearn things or rely on institutional memory. In that environment something that can genuinely look at the full breadth of the digital information could be very useful.

 

[Moobert]

Decided to dabble a bit with this recently and it has went down far too often for it to be relied upon for anything urgent.

 

[jpaul]

a precursor ? would you want your children to experience this

Holocaust survivors to use AI to ‘future-proof’ their stories for UK schools​

Pollock added: “We want young people to understand the individual experiences of people that managed to survive but they do need to understand about the loss and destruction and the nature of antisemitism, the nature of hate that led to this crime against humanity. When the survivors aren’t here and with Holocaust denial … it’s still very much an evolving phenomenon, with an explosion of antisemitism … it’s a critical moment.”

Manfred Goldberg, who survived the Stutthof concentration camp and a death march, is the first survivor to have had his story preserved using this technology for UK students. For the programme, Goldberg answered more than 1,000 potential questions and filmed for more than 20 hours across five days. Through AI, the survivor is able to answer a variety of questions ranging from his own experiences of the Holocaust, moving to the UK and even his favourite football club.

radio article suggest it will (in theory) only answer questions for which answers were recorded,
so we will repeat this for testimonies of other historical events ww2 veterans ....

 

[413x]

Damn.
Chat-gpt 4o is very good.

Im certainly scared for my Job.

Wonder if Google will end up losing market share with this.

I feel like in coding type jobs you almost need to be using it.
It can really help where you get stuck especially.

Today I got it to write a bit of DAX I couldn't figure out. Previously I would have trawled Google + worked it out. No I just need to learn how to ask it better what I need!

 

[GravyMonster]

Im certainly scared for my Job.

I wouldn't be too worried...

Today I got it to write a bit of DAX I couldn't figure out. Previously I would have trawled Google + worked it out. No I just need to learn how to ask it better what I need!

Did you learn how to achieve the outcome yourself when the need arises in the future? This is the discussion I have with devs I work with who rely on AI perhaps a little too heavily. Have you learned any useful skill other than proompting? What's the benefit to you? Unless you took the time to study the answer it gave you and evaluate if it was accurate, correct, secure etc. then in effect all AI has done is speed up the process by which you can plagiarise Stack Overflow.

 

[D3K]

Damn.
Chat-gpt 4o is very good.

Im certainly scared for my Job.

Wonder if Google will end up losing market share with this.

I feel like in coding type jobs you almost need to be using it.
It can really help where you get stuck especially.

Today I got it to write a bit of DAX I couldn't figure out. Previously I would have trawled Google + worked it out. No I just need to learn how to ask it better what I need!

Claude 3.5 Sonnet is wiping the floor with 4o for coding.

Which means openAI are gonna have to release 5 sooner than they wanted, I think.

This arms race is great. Everyone trying to out do each other keeps driving it forward. And opensource local models keep getting better which further piles on the pressure of the for-profits.

On a bum note, looks like the 4o real-time speech is still a way out. Closed testing groups end of July. Not quite the "few weeks" timescale they promised in last showcase.

 

[413x]

I wouldn't be too worried...

Did you learn how to achieve the outcome yourself when the need arises in the future? This is the discussion I have with devs I work with who rely on AI perhaps a little too heavily. Have you learned any useful skill other than proompting? What's the benefit to you? Unless you took the time to study the answer it gave you and evaluate if it was accurate, correct, secure etc. then in effect all AI has done is speed up the process by which you can plagiarise Stack Overflow.

In this case I was under tone pressure. So unfortunately I just needed to get it done. It's not great.. But needs must

 

[413x]

Claude 3.5 Sonnet is wiping the floor with 4o for coding.

Which means openAI are gonna have to release 5 sooner than they wanted, I think.

This arms race is great. Everyone trying to out do each other keeps driving it forward. And opensource local models keep getting better which further piles on the pressure of the for-profits.

On a bum note, looks like the 4o real-time speech is still a way out. Closed testing groups end of July. Not quite the "few weeks" timescale they promised in last showcase.

Thanks. I will check this out tomorrow. Not actually heard about that. I'm only just starting even trying to use these tools to be fair

 

[Tykey]

I found Claude 3 Opus to be no better than GPT4, they both have new models out (GPT 4o and iirc Claude 3.5), I'm finding 4o to do most things I want very quickly, I'm not risking and possibly wasting another months £ on Claude. I take the benchmarks with a pinch of salt because for some tech reasons I read which I forget now, and when I tried them, well as I said, no different.

 

[dowie]

This is really good... a nice reference to a 00s internet joke in here.

https://twitter.com/x/status/1806822229793177624

Claude 3.5 Sonnet is wiping the floor with 4o for coding.

Yup, I think GPT4 is often better than 4o too. The reason for the hype with 4o is that it's multimodal, you don't need additional text-to-speech and speech-to-text and vision neural nets to chain together so you can show it things and talk to it with much lower latency. But it's ability to resolve your text queries, help with code etc.. isn't necessarily better than their GTP4 model.

I think people assume 4o was supposed to be the better model in that sense simply because it came later and has that version number but really it's just supposed to be an equivalent model to GPT4 with the multimodal capabilities (and it's cheaper to run inference on IIRC) and in reality if you're not using those capabilities, or serving a product to customers on top of it then GPT4 may well be better for some tasks (Or indeed models like Claude from rival firms.)