The AI discussion thread

[Kaido]

I find it significant that some people are so bad at relationships or so poorly matched, that their partner would rather just communicate with AI instead.

tbh I see a place for AI therapy. There's still a strong stigma attached to depression & therapy for men in America. Statistically, the suicide rate among men is nearly 4x that of women. Men represent 80% off all suicides, despite only making up 50% of the population. I've lost a number of friends that way, which has been awful. Having a private conversation system to talk to might end up helping a lot of people! There's a lot of interesting research on things like suicide prevention hotlines & how being empathetic is statistically more successful:

Comparison of the Effects of Telephone Suicide Prevention Help by Volunteers and Professional Paid Staff: Results from Studies in the USA and Quebec, Canada - PMC

Research since the 1960s has consistently found that lay volunteers are better at helping suicidal callers than professionals. Yet, professional degrees are increasingly becoming requirements for helpline workers. In our first study, we conducted ...

pmc.ncbi.nlm.nih.gov

 

[Kaido]

Had Bonnie Tyler on my Youtube playlist today:

Forgot about this gem from the music video:



Playing with an AI re-imagining:



The whole vide is premium-grade meme material lol:



An 80's song deserves the Macross treatment:

 

[Kaido]

The whole vide is premium-grade meme material lol:

View attachment 125593

An 80's song deserves the Macross treatment:

View attachment 125594

Ramastered version would go HARD as an anime re-imagining lol! But alas, I am now out of free video credits haha

(click image if it's not animated)

https://imgur.com/iCMT6iv

 

[Kaido]

Just watched CBS Saturday. They had a segment about developing a relationship with their AI. One guy was asked if he could give up using AI if his partner asked him to. His response was a bit shocking. Watch it if you can.

Just saw it. Man...worth the full watch for sure:

 

[Kaido]

https://twitter.com/x/status/1934098165025935868

 

[Kaido]

Video to Video:

https://twitter.com/x/status/1933563286160879693

https://twitter.com/x/status/1933911763378221341

 

[Kaido]

Video to Video:

https://twitter.com/x/status/1930271923805540816

 

[Kaido]

Video reframing...old animations are going to be so cool!

https://twitter.com/x/status/1928109538487476510

https://twitter.com/x/status/1927741486696730742

 

[Kaido]

RunwayML visual effects for film. How far we've come in the 2 years since I started this thread!

https://twitter.com/x/status/1934312626021949687

Midjourney Video:

https://twitter.com/x/status/1933992009955455193

 

[Kaido]

Midjourney V1 Video Model officially released:

https://www.midjourney.com/updates

 

[Kaido]

New Midjourney style reference system has over 4 billion new style codes:

 

[Kaido]

AI is coming for influencer jobs now!

https://twitter.com/x/status/1935198754933792837

 

[Kaido]

Neurospot CCTV AI Video Analytics:

Neurospot | CCTV AI Video Analytics - Sparknets

Introducing Neurospot, the cutting-edge CCTV AI video analytics solution that redefines surveillance and operational efficiency. With its advanced features

sparknets.com

https://twitter.com/x/status/1935042505470255134

SightBit AI Lifeguard:

Two Very Different Examples of Computer Vision

On another day of scrolling through LinkedIn today I encountered two posts about computer vision that caused me to have very different…

medium.com

 

[[DHT]Osiris]

https://twitter.com/x/status/1934098165025935868

Regarding hallucinations/falsities, it'd be good if in the cases it's able to do so, actually test the results. I'm using my own career in IT as an example, it often muddles command structures together, giving Linux commands for PowerShell requests, or -x arguments where /x arguments are needed, etc. If it could actually perform the request it's authoring to you to verify the command can run without error (and incorporate that into a permanent memory) that would go a very long way.

 

[igor_kavinski]

it often muddles command structures together

You need a 200B+ parameter model for that. Anything less is prone to errors. I have tried everything from 32B to 90B and none of them were able to give me code that compiled successfully except the 231B Deepseek one.

All these free AI services, I doubt they are using such dense models. Only the paid ones would utilize the highest parameter models.

 

[[DHT]Osiris]

You need a 200B+ parameter model for that. Anything less is prone to errors. I have tried everything from 32B to 90B and none of them were able to give me code that compiled successfully except the 231B Deepseek one.

All these free AI services, I doubt they are using such dense models. Only the paid ones would utilize the highest parameter models.

Is it not possible to get the AI to 'try' the mechanism it suggests before suggesting it? Isn't that the basis of AI that trains to walk a robot, play a game, etc?

 

[igor_kavinski]

Is it not possible to get the AI to 'try' the mechanism it suggests before suggesting it? Isn't that the basis of AI that trains to walk a robot, play a game, etc?

I bet that's what the "real" AI companies are doing for themselves and their bigger clients. The little people get the castrated LLMs that only spit out text. It could be possible to do a sort of chained reasoning analysis where the output of one LLM is fed into another LLM for verification and refinement of the results and even being fed into a compiler for finding out the syntax errors and fixing them automatically but it would take a fair bit of Python-fu.

 

[[DHT]Osiris]

I bet that's what the "real" AI companies are doing for themselves and their bigger clients. The little people get the castrated LLMs that only spit out text. It could be possible to do a sort of chained reasoning analysis where the output of one LLM is fed into another LLM for verification and refinement of the results and even being fed into a compiler for finding out the syntax errors and fixing them automatically but it would take a fair bit of Python-fu.

It doesn't even have to be that advanced. Just post it to an online compiler and see if you get error text back.

Try It Online

tio.run