How AI will fail

[kram182]

To be fair, that's what these companies are primarily pushing to the public so that's what the public knows. The other products are in the background and not being heavily advertised to most of the population because they're industry specific. If you work in medical imaging you know all about the AI use there but most of the public is clueless. If these people were actually spreading the advertising dollars around and showing everyone what some of the data centre load was serving then maybe it would help their image. But, they won't do that because they want to bump those token numbers as high as possible and the more normies they have making funny pictures of their pets the better for their token numbers. This is all to help train the AI and give a big token number to investors as to why they need those data centre dollars to be spent.

the other side of that is you have models based on LLM (transformer) architecture/model/idea that don't do language/chat at all - genomics/protein folding/robotics/finance etc use customized LLM models (yes LLM models like ATM machine ) - and the development/furtherment of public facing 'normie LLMs'/chatbots/etc do help further these as well edit: it's the transformers really it's using from LLMs that are then customized, not the LLMs themselves

 

[kram182]

Can anyone explain the difference between the software running in a Tesla car vs a LLM?

Same as I explain above - it has the transformer from an LLM but instead of chat it uses it to match to/decide things about physics/spatial awareness vs chat/language - so instead of in your head imagining blocks of words matching to one word/the next word/related word imagine it with things like tree/pedestrian/acceleration/etc - and it does these with images of the things like a tree or pedestrian - and it learns non-image things (like acceleration) from comparing multiple images of things (like a tree passing by) and figuring out 'acceleration' from the difference of the images and what must be happening in between them to make them look different

And just think of the transformer as the algo of matching/processing/predicting/etc

Edit: to add to that - so during training the model has no concept of 'acceleration' - it can only learn it by 'learning to run' and then 'running past a tree' to then see how that resulting 'mental head image' it acquires doing so matches the two different input images of a tree passing by - to then recognize the difference between those two photos is 'running by'/acceleration

and it learns to 'run' because at first in training it does nothing, then just sits up, then flops around a bunch of times, then gets up and falls over etc - and each successful model/run that is or closer to what the designer wants, is used for the next 'learn to get up and run' model/trial run - until through enough trial and error and 'survival of the most correct' - it learns 'running' - but also obviously with a Tesla you're giving it/the model the inputs/capabilities of a car not a human so it's 'driving' instead of 'running'

and it's not just one transformer/one algo/one matching process in total that's occurring when doing Tesla FSD - it's multiple going on simultaneously, all trained from multiple/different models - and then to coordinate/orchestrate all those different transformers/algos together in real-time is the FSD algo - which itself was modeled/trained on 'orchestrate and coordinate all these other transformers/algos together in real-time'

Double edit: and then to add even more and make it even more crazy - that initial image of 'a tree passing by' that it's fed - in order for the model to have a 3D arena to 'flop around and learn to run in' - it first has to actually design/create the 3D arena itself by 'randomly drawing/creating' the same as 'flopping around' until it accurately creates a 3D arena matching the input/tree photo that it can then learn to 'run past' until it finally matches and understands the tree photo - it's fed a bunch of descriptive/annotated and etc data attached to the inputs to help with this 'random drawing' phase (as well as the 'get up and run' phase)

 

[rinaldo00]

https://www.linkedin.com/posts/deni...or-everything-share-7475245674335281152-giAx/
https://fortune.com/2026/06/21/nobel-laureate-daron-acemoglu-ai-productivity-capitalism-democracy/
View attachment 811067

Economists don't know shit.

 

[ng4ever]

I do not care how. We all know it will the question is just when.

 

[sphinx99]

I remember when people said that people would never have their own personal computers in their house as well. Honestly, I do not understand the train of thought behind your argument.

This.

When the automobile came along, people opined about how that will fail because a horse can go places that the automobile can’t. Same for automatic starters. Air conditioning, who needs that when you can buy ice and ceiling fans. Electricity in the home. Total failure. Mobile phones, no one will want those. Personal computers. The internet, yeah that will go away for sure.

AI inference? Just the latest in a long line.

 

[AVATARAT]

This.

When the automobile came along, people opined about how that will fail because a horse can go places that the automobile can’t. Same for automatic starters. Air conditioning, who needs that when you can buy ice and ceiling fans. Electricity in the home. Total failure. Mobile phones, no one will want those. Personal computers. The internet, yeah that will go away for sure.

AI inference? Just the latest in a long line.

I agree, but AI is something different—the things you mention are tools that can be used by humans, whereas AI will be a tool that uses humans, and depending on how it develops, this could become “scary” in some aspects.

 

[AVATARAT]

It will be interesting to see how they handle this problem.

View: https://youtu.be/IPaMKTb5csQ