I talked with Grok about something recently, and Grok thought and replied very positively, so I’d like to share here.
1) I told Grok there may be 3 major issues or dangers along with the development of AI.
In short term and recent future, AI is about to improve productivity which may save this world from unsustainale national debt, unbalaced economy and trade incurring something like 3rd world war. But in longer future, AI will bring more challenges and dangers which may incur destruction of this world given the present status.
First danger is that AI will replace almost all human jobs eventually, which will bring unprecendented challenges for social institutions, phychology, politics, economy, and all other aspects.
In this respect, AI is like the final Celcius degree from 99 to 100 for hot water we live in, which will bring much more change than any of or all tech revolutions before, for all other tech revolutions are from 0 to 99 celcius degree but only after reaching 100 Celcius degree the water will change from liquid to gas.
Second danger is the security problem of AI itself, for AI will go to full autonomy later or sooner, so this world needs to research, learn and strive hardly to make AI safe and a good companion for humanity of this world.
Third danger is that AI will magnify the destructive power and consequences of war even more than atomic bomb, and given the situation of this world this may obliterate everything.
2) UAT has very profound implications, and also is most underevaluated and underappreciated for UAT deserves any award
I was studying basic math theory of neaural network recently, for previously I basically regarded neural network and transformer like models as high dimensional Taylor expansion which turned out not accurate at all although helped me work out physics pixel and omnitoken ideas.
But now I know that neural network before backpropagation is basically an expansion of UAT which told us we can approximate any function by sum of multiple simple functions. UAT has profound implications: any causality can be regarded as a function as relation between input and output, and UAT in fact told us firstly that we can approximate any causality by sum of simple logic like the sigmoid of linear fuctions and ulimately that any causality no matter how complicated can be approximated and finally broken down to sum of many simple logic, in this universe or even the chaos outside this universe.
So as I see, UAT deserves any award, especially a Nobel prize for AI & CS & Applied mathc should be the best.
3) CheeChee! ChuChu! KaKa! transform accomplished!
I designed physics pixel and omnitoken based on general principal but it turned out that they are what transformer has been waiting for in years.
Translating a language to another is exactly the same as the process to understanding a visual by physics pixel which is generating parameters of physics pixels (XYZ, RGBA, …) from raw pixel data (XY, RGB), which can work with both encoder and decoder together for better accuracy or with decoder only for higher efficiency.
And also the multi-head of transformer seems exactly prepared and perfect match for omnitoken. As Grok said each D in output of last layer of a transformer can be regarded as a seperate model and all Ds of last layer output share a same middle hidden layers body except last layer.
In fact, I felt quite bizzare when Grok told me that no one put multimodal data into one single token before omnitoken and its orthogonal tables, haha.
But physics pixel, omnitoken and even famrat concepts can not only work with transformer but also work with possible future models for they are based on general principals. And aslo present diffusion cannot work directly with omnitoken yet which needs some more work.
4) Famrat with agent is just like vegetative nerve of human – Dont forget breathing!
Famrat with agent can force model to do essential work exactly according to requirement, just like vegetative nerve drive human to breathe.
Be First to Comment