AI - Cheating or just evolution?

[avasopht]

Testing gpt-4o... I have a txt file with a list of movies in Portuguese and English. I ask to separate the English titles from the Portuguese titles and put them in alphabetical order. It gets everything wrong, every time.

You're using it wrong.

It would be best if you used a little prompt engineering so that it works on a single title at a time outputting whether it's English or Portuguese. Tell it to output in JSON (maybe give it an example JSON).

Filtering and sorting shouldn't be done via an LLM. Some filtering can be done with an LLM, but there's no need to.

[avasopht]

I'm here to make my music...
AI can be better at it, but it doesn't matter.
My music is personal to me.

100%.

Besides, even if AI can make better music than me, it can't replace my experience of making it and the joy of listening back to what I created.

[Higor]

You're using it wrong.

It would be best if you used a little prompt engineering so that it works on a single title at a time outputting whether it's English or Portuguese. Tell it to output in JSON (maybe give it an example JSON). Filtering and sorting shouldn't be done via an LLM. Some filtering can be done with an LLM, but there's no need to.

I think the chat should help me resolve this, instead it hallucinates adding titles that are not in the list and omitting others. I copied the list from the txt file and pasted it directly into the chat but the task is still not done correctly. About json, I'm not a programmer so I don't really know what that is.
But, yes the chat does something with that.

[avasopht]

I think the chat should help me resolve this, instead it hallucinates adding titles that are not in the list and omitting others. I copied the list from the txt file and pasted it directly into the chat but the task is still not done correctly. About json, I'm not a programmer so I don't really know what that is.
But, yes the chat does something with that.

You've got to use LLMs according to how they work.

Stating an output format in your prompt gives it constraints that help it produce more appropriate output.

LLMs aren't good for sorting. It's just not something the architecture is good for.



LLMs aren't magic. Use them for what they are good at.



You will undoubtedly find demonstrations somewhere of this capability through the use of an LLM, but this may be achieved through things like LangChain and LangFlow that feed some LLM outputs to code that does the sorting.

[bxbrkrz]

[bxbrkrz]

[Yonatan]

Yes, A.I in the use of physical products are so overhyped and in its infancy at the moment. No standards are in place, no real update plans. So many people will have some AI refrigerator that just behaves strange and company don't give support etc. At the moment, the AI in machines or consumer goods, are still so mediocre.

I mean, even the very advanced robot models in the forefront, is still clumsy and not trained or refined enough to really replace human service workers just yet.

That being said, there will soon be a breakthrough when the bridge of neural network computing and machine learning will mature and leap into actual really smart products. We are not really there yet, a few "wow" can be seen at display but it is in no way any affordable intelligent consumer goods. Give it 5-7 years and the boom is real. By 2030, I think things will have start taken off in a way that makes us have to change how we structured our societies. Education, art, manufacturing, economics, medical fields. All the field of HR-type-of-jobs and office tasks now being seen as high educational labour, will be taken over by AI. Rightly so because most office workers are bored by their jobs anyway but they get decent pay and seen as smart because they went to some university.

AI will also help societies to filter off some % who today run to the doctor for minor things that actually specialized AI will do a better job taking care of. Today, both psychologists and medical system is overloaded of quite mundane things and also the medical profession complain about the load of New Public Management tasks of documenting everything, so that will AI take care of and will free us to focus on the more human interaction. Many inbetween middle bosses will be redundant too. And overall we could lessen our labour hours. But it will not happen without some struggle, old ways of thinking will still try persist. And people will at beginning lose their jobs, lose hope and wonder what the heck life is for, and the people with powerful positions will continue to collect the fruits and become wealthier and wealthier. The tension will create a bit of a havoc before the new structures are in place.

[Aosta]

For those that lived through the digital revolution it was the same thing, everything was 'digital' including breakfast cereals. You are always going to get snake oilers, scammers and liars out for a quick buck when an innovation evolves or a crisis happens. Remember Y2K? The amount of money spent on 'Y2K ready' PCs and software CDs to stop the virus from destroying your system, turned out to be a whole load of horseshit but some made millions from it.

[avasopht]

Remember Y2K? The amount of money spent on 'Y2K ready' PCs and software CDs to stop the virus from destroying your system, turned out to be a whole load of horseshit but some made millions from it.

The Y2K bug was a genuine problem.

Many critical systems (hardware and software) that were no longer maintained didn't support dates past 1 January 2000. Some also risked a glitch on 9th September 1999 (sometimes used to represent an unknown date).

Many developers had to work tirelessly to patch older systems.

Microsoft Exchange Server suffered a similar problem in 2022 when emails got stuck on the transport queue because they used 32-bit integers to represent dates.

As digital calendars around the world changed to 01.01.2022, Microsoft customers found their Exchange servers stopped processing emails.

Microsoft says it is aware of a programming flaw which saw some customers' Exchange servers stop processing emails just as the clock struck midnight on New Year's Eve.

System administrators, who are sharing workarounds on social media, have dubbed the bug Y2K22 - in the style of the Y2K bug which affected some computers at exactly the same time 22 years earlier.

This was only 2 years ago.


Below a developer of the time explains the Y2K bug and how they fixed it:

Imagine that you make mobile phones. If I came to you now and told you that your phone was going to fail in 2050, would you worry? Probably not. There’s almost no chance that anyone would still be using it by then.

Imagine that you’re a lowly programmer working in a bank and it’s the year 1970.

You’ve got these big reels of tape that are supposed to hold your customer accounts. It’s a massive amount of data. There’s no memory or even disk that’s big enough to hold them. We’re talking about millions of accounts and every one of them could be a few hundred characters (bytes).

Those tapes are limited in size and they take time to read and write. The smaller you can make an individual customer account record, the more you can fit on a tape and the faster it will be to process.

Ten years before, your older colleagues were similarly limited with an 80-character punch card.

So you abbreviate as much data as you can. Dates become six digits: DDMMYY.

Do you worry about the year 2000? You might, but you might not. Memory will be cheaper in 30 years. You’ll be using a newer, faster computer and this program will probably have been forgotten. You need to compress this data now. There’s plenty of time to worry about the two digit years later.

...

Most of that checking doesn’t find any problems. But some of it does. Someone needs to reformat all those customer records and adjust the programs to use four-digit years. Then they need to extensively test because corrupting your entire customer database is a bad thing. People get irritated if you lose their life-savings.

We did that work. Some people were in the right job at the right time to make a lot of money checking every line of every program. They found a lot of bugs. They fixed almost every one of them.

Was the Y2K bug over-hyped? Probably not. We probably had to have that level of awareness just to get sufficient buy-in. Without it there would have been companies who never even thought they might have a problem.

...

So January 1st 2000 happened and nothing much went wrong. The people that needed to do the work had done it. No aeroplanes fell out of the sky. No power stations exploded. There were a few failures but they were not enough to bring down civilization, or even to notice (unless you were affected.) There were probably many more — maybe that plumbing company didn’t take the issue seriously and had a difficult year — but nobody got to hear of them.

So then people started looking back at all the hype and the millions of pounds that was spent and assumed that it was all a big scam. It wasn’t; that money needed to be spent, the work needed to be done and it was done.

Did we learn anything? I hope so. I hope we learned that software lasts longer than we think and that data lasts even longer. But I can still imagine some millennial, fresh out of college, sitting down in an office for the first time and thinking “what’s the chances that this stupid program is still going to be here in 2099?”

[bxbrkrz]

[Yonatan]