All things AI

[psydii]

I mean given a long enough timeline we'll get there. But IMO the "flaws" that cause the limitations in the current "AI" system are fundamental to the maths behind them. To brute force a solution (to compensate for these fundamental flaws) requires greater than exponential increases in power. In decades prior we could rely on getting better at making semiconductors. But now we are approaching the physical limits for how small things can be made, so the only things we can do (in the long run) is use more power,  or change the fundamental maths.

 

[mavhc]

LLMs are so new we don't know what improvements can be made. a) just use better data, not just the entire internet, b) quantization, c) split up tasks, now you have multiple context windows, d) when running out of context window, summarize the previous output, e) multimodal, f) design chips specifically for AI, not just repurposed GPUs. Or a combination of some/all of the above.

 

Given a modern LLM with 1 trillion parameters, that's only 100x less than the number of human brain synapses. We're up to the level of a cat or bird. 10 trillion is a baboon

[psydii]

23 minutes ago, mavhc said:

LLMs are so new we don't know what improvements can be made.

 We're just using more power to brute force maths that was known for decades but impractical to use computationally. All of the research (and now products) of the last decade are as a result of the maths that GPUs are really really good at, at higher and higher cost.  Any chip designed specifically for AI would actually be a damn good GPU because the maths is the same.

The sort of improvement required for AIs to be more than job destroying autocomplete machines, requires fundamentally different maths running on fundamentally different hardware, and that is (at least) decades away.

Edited November 7, 2025 by psydii

[sigma]

The last thing we did before Covid in 2020 was to go to a talk at the Ri by Vint Cerf on “Will computers ever think like human beings?” 
There’s a full recording here: https://m.vk.com/video-220369294_456239141

[Ditto]

The biggest problem with AI is the 'I'. Thinking that the goal is intelligence in the way that humans exhibit intelligence is the wrong way to think about it's use. Humans are very good at being human, from the extreme socially effective, compassionate being, to the most horrific being on the planet. We could do with a better term than AI. However, what is currently referred to as AI is extremely helpful to allow humans to make rapid steps forwards in a variety directions, but can be used to the detriment of humans too. We don't need technology that thinks like humans, but we can benefit from technology that works better and faster than things we are not so good at.

[Roberto]

18 hours ago, Ditto said:

The biggest problem with AI is the 'I'. Thinking that the goal is intelligence in the way that humans exhibit intelligence is the wrong way to think about it's use. 

 

I think if I were inventing and naming it, it would have been called something like IA - Intuative Automation.

[6Foot2]

3 minutes ago, Roberto said:

I think if I were inventing and naming it, it would have been called something like IA - Intuative Automation.

 

IDEA: Intuitive Data Experience Automation

[mavhc]

On 07/11/2025 at 15:54, psydii said:

 We're just using more power to brute force maths that was known for decades but impractical to use computationally. All of the research (and now products) of the last decade are as a result of the maths that GPUs are really really good at, at higher and higher cost.  Any chip designed specifically for AI would actually be a damn good GPU because the maths is the same.

The sort of improvement required for AIs to be more than job destroying autocomplete machines, requires fundamentally different maths running on fundamentally different hardware, and that is (at least) decades away.

Not "just", there's many papers on the improvements made of the past decade since deep learning, alphago, "Attention is all you need", LSTM, BERT, FlashAttention, Mixture of Experts, Diffusion, RLHF, etc

 

How do you know we need different maths?

[6Foot2]

Link: ChatGPT is down worldwide, conversations disappeared for users

 

Quote

OpenAI's AI-powered ChatGPT is down worldwide with users receiving errors when attempting to access chats, with no reasons currently given.

 

If you are affected, you will see "something seems to have gone wrong," errors, with ChatGPT adding that "There was an error generating a response" to their queries.

 

In our tests, BleepingComputer observed that GPT keeps loading, and the response never comes.

 

In addition, some users reported that conversations disappear and new messages keep loading...

 

[6Foot2]

Link: AI To Be ‘Embraced’ by Famous Stop-Motion Animation Studio

 

Quote

... Wallace and Gromit creator Nick Park has opened up about how AI will influence the future productions of Aardman Animations. The co-founder stated that the studio will explore how the technology could fit into its workflow.

 

Speaking to Radio Times, Park reflected on earlier technological shifts that once raised concerns in the animation landscape. He recalled the release of Toy Story in 1995, saying, “When Toy Story came out, we thought, ‘How long have we got?’ But we’ve managed to survive CGI.”

 

He added that interest in handcrafted stop-motion has grown over the years, noting, “In fact, there’s been a resurgence of interest over the years in our stop motion animation.” Although the studio incorporates CGI when needed, Park noted that AI “is a whole new thing.” He added, “Obviously, a lot of people will be fearing for their jobs.”

 

Park explained the studio’s current stance, noting, “We want to embrace the technology and find in what ways it’s going to be useful to us, maybe to do animation a bit quicker, but we’re going to be very cautious not to lose our values.” He stressed that clay animation remains important to Aardman’s identity, describing it as the company’s USP. “The clay is our USP, and we pride ourselves in that. Authenticity is the most important thing. It’s where the charm is.”...

 

[6Foot2]

Link: Moonpig’s use of AI to design and personalise cards drives up sales | Retail industry | The Guardian

 

Quote

The online card service Moonpig has reported a bump in sales thanks in part to its increased use of AI to help design cards, personalise customers’ messages and answer queries.

 

The company said sales rose 6.7% to £169m in the six months to 31 October and had remained strong in the weeks since then, largely as a result of increased orders and spend per order at its main Moonpig brand.

 

“AI is now designing a lot of cards for us,” said its chief executive, Nickyl Raithatha. He said technology had helped create everything from baby and birthday cards to corporate greetings linked to a particular business.

 

“It is still being managed by our in-house team. We make sure a person will look at it and it is relevant and exciting for customers. We don’t want to fill our site with generic design. We are treading carefully.”...

 

[6Foot2]

Link: Google unveils plans to try again with smart glasses in 2026 - BBC News

 

Quote

Google plans to launch smart glasses powered by artificial intelligence (AI) in 2026, after its previous high-profile attempt to enter the market ended in failure.

 

The tech giant set expectations high in 2013 when it unveiled Google Glass, billed by some as the future of technology despite its odd appearance with a bulky screen positioned above the right eye.

 

Google pulled the product in 2015 less than seven months after its UK release, but is now planning on re-entering the market with smart glasses with a cleaner look.

 

But it comes after Meta has already made waves with its smart specs, which have sold two million pairs as of February.

 

Google's new tech will let users interact with its own AI products, such as its chatbot Gemini.

 

It plans to launch two different varieties - one designed to be provide assistance without any kind of screen, and another which has a display on the glasses themselves.

 

While saying the first of the two types of AI glasses it is working on would come in 2026, Google has not given more information on what form it will take...

 

[dmj]

This documentary is worth a watch

 

Edited December 18, 2025 by dmj

[dmj]

AI has killed stackoverflow

 

https://data.stackexchange.com/stackoverflow/query/1926661#graph

 

I guess they will have to start training on regurgitated AI content

[mikes]

On 07/11/2025 at 15:34, mavhc said:

 

Given a modern LLM with 1 trillion parameters, that's only 100x less than the number of human brain synapses. We're up to the level of a cat or bird. 10 trillion is a baboon

 

It's not necessarily the number though, more like how they are arranged?  The Elephant / Whale has more neurons and more synapses than us but they aren't more intelligent than us or even arguably sapient like we are. (Sentient yes but not sapient). We have no idea how to arrange our "connections" to get intelligence, just scanning text and assigning weights, I really can't see how that will suddenly get us "sapience" or even intelligent "thinking" machines.  Yes it will find very good patterns/connections between text (same thing with pixels in a movie etc) but I've seen no evidence these things can "think". 

[dmj]

6 minutes ago, mikes said:

 

It's not necessarily the number though, more like how they are arranged?  The Elephant / Whale has more neurons and more synapses than us but they aren't more intelligent than us or even arguably sapient like we are. (Sentient yes but not sapient). We have no idea how to arrange our "connections" to get intelligence, just scanning text and assigning weights, I really can't see how that will suddenly get us "sapience" or even intelligent "thinking" machines.  Yes it will find very good patterns/connections between text (same thing with pixels in a movie etc) but I've seen no evidence these things can "think". 

... and they have to spend most of the connections on mundane stuff like staying alive. Opus just needs to be able to spit out code. 

[Ditto]

Staying alive - mundane! Hmm, quite important for us humans. I guess the elixir of life for AI is electricity, one flick of te switch and it's as if it never existed.

[Theblacksheep]

Humans seem to substantially under value the intellgence of non-primates and significantly over value the suposed neuron equivelence in a processor.

 

 

[Rob_D]

36 minutes ago, Theblacksheep said:

Humans seem to substantially under value the intellgence of non-primates and significantly over value the suposed neuron equivelence in a processor.

Yes, but can an elephant run Crysis? 😁

[Andrew_C]

3 hours ago, mikes said:

 I've seen no evidence these things can "think". 

I've seen an awful lot of evidence recently that a lot of humans don't/can't think particularly well either.

[6Foot2]

Link: Meta signs 3 deals for nuclear energy to power AI data centers - CBS News

 

Quote

Meta has cut a trio of deals to power its artificial intelligence data centers, securing enough energy to light up the equivalent of about 5 million homes.

 

The parent company of Facebook on Friday announced agreements with TerraPower, Oklo and Vistra for nuclear power for its Prometheus AI data center that is being built in New Albany, Ohio. Meta announced

 

Prometheus, which will be a 1-gigawatt cluster spanning across multiple data center buildings, in July. It's anticipated to come online this year.

 

Financial terms of the deals with TerraPower, Oklo and Vistra were not disclosed.

 

The Mark Zuckerberg-led Meta said in a statement on Friday that the three deals will support up to 6.6 gigawatts of new and existing clean energy by 2035. A single gigawatt, according to a general industry standard for utilities, can power about 750,000 homes...

 

[6Foot2]

Link: Scientists launch AI DinoTracker app that identifies dinosaur footprints | Dinosaurs | The Guardian

 

Quote

Experts have created an app that uses artificial intelligence to identify dinosaurs from the footprints left behind after they stomped across the land tens of millions of years ago.

 

“When we find a dinosaur footprint, we try to do the Cinderella thing and find the foot that matches the slipper,” said Prof Steve Brusatte, a co-author of the work, from the University of Edinburgh. “But it’s not so simple, because the shape of a dinosaur footprint depends not only on the shape of the dinosaur’s foot but also the type of sand or mud it was walking through, and the motion of its foot.”

 

Writing in the Proceedings of the National Academy of Sciences, Brusatte and colleagues report how previous AI systems based their learning on footprints that had already been labelled as having been made by particular types of dinosaur. But, the team note, if those identifications are incorrect then the AI system will also be flawed.

 

“You never find a footprint and alongside [it] the dinosaur that had made this footprint,” said Dr Gregor Hartmann, the first author of the new research from Helmholtz-Zentrum in Germany. “So, no offence to palaeontologists and such, but most likely some of these labels are wrong.”...

 

[6Foot2]

 

Link: AI Confidential with Hannah Fry - BBC iPlayer

 

Series:

1. The Boy Who Tried to Kill the Queen: Examines how a chatbot encouraged a teenager’s attempt on Queen Elizabeth II’s life.
2. Death by Driverless Car: Investigates liability and consequences when pedestrians are killed by autonomous vehicles.
3. The Assassin and the Algorithm: Explores the link between murder suspect Luigi Mangione and controversial AI algorithms used in healthcare. 

[garethEds]

3 hours ago, 6Foot2 said:

 

Link: AI Confidential with Hannah Fry - BBC iPlayer

 

Series:

1. The Boy Who Tried to Kill the Queen: Examines how a chatbot encouraged a teenager’s attempt on Queen Elizabeth II’s life.
2. Death by Driverless Car: Investigates liability and consequences when pedestrians are killed by autonomous vehicles.
3. The Assassin and the Algorithm: Explores the link between murder suspect Luigi Mangione and controversial AI algorithms used in healthcare. 

 

Wondering if there is any bad language in this - so I can can use bits for work in school with my Year 9

[Rob_D]

11 hours ago, garethedmondson said:

Wondering if there is any bad language in this - so I can can use bits for work in school with my Year 9

It is set to air after 9pm, but I can't immage there would be much given it's positioned as a science/technology documentry.