Ooh a new channel, exciting! I guess I may as well be the first to ask: how does the discipline of developing AI agents differ from more conventional machine learning? (What topics belong here vs. there?)

OK, I'll bite. How are AI agents different from conventional machine learning?

I could see this channel discussing what AI agents are best for certain types of tasks.. / how to achieve the best results.. aka.. prompt engineering..There is a grey area though.. cus once you get into tuning, performance and deploying.. it could also go in machine learning 😔

We have an idea to evolve the Turing Pi platform towards self-hosted AI agents. We are still working on this direction

AI agents > Smith was an AI agent of the system and, like other Agents, had the role to police and maintain the Matrix by eliminating potential threats to the stability of the system, which were mostly seen to be Redpills and defunct programs.

Is this TMs's way of announcing a TPi2 board powered by energy harvested directly from humans? 👀

Soylent-Pi?

Some interesting parallels between Digital computers and Natural Language Computers

Natural Language Computer | Digital Computer
---------------------------------------------------
Base components
--
LLM                       | CPU
LLMCache                  | CPUCache
Context                   | RAM
Prompt                    | Code block
Inference                 | Execute code block in CPU

Interrupts
--
Functions                 | Interrupts
Functions Schema          | Interrupt vector table

Storage
--
VectorDB                  | Disk storage
Data Indexing             | File system indexing
RAG                       | Disk API for Retrieving/Querying

Applications / Services
--
AI Agent                  | Application / Service
AI Agent State            | Application state (e.g. TERM)
Send/Reply methods        | Inter-process communication (IPC)
Templates? (early idea)   | Executable formats like ELF, PE 

Advanced
--
Virtual Context           | Swap file / virtual memory (MemGPT)

Now I'm wondering whether the Sapir-Whorf hypothesis applies to NLCs: Does the natural language that the LLM is trained on affect its cognition?

I am not familiar with this hypothesis, but I will look into it

The TL;DR in your analogy would be:

Natural Language          | Operating System

The hypothesis is that a person's native language shapes how they perceive the world and think.

Transformers was made possible by word-to-vector work. This makes it possible to perform vector operations on words So LLMs work with strings/words, which opens up the possibility of building new types of computers that work with strings/words instead of numbers. A digital computer works with numbers (bytes encoded in bits), while LLMs work with strings (tokens encoded in vectors) A digital computer performs bit operations A natural language computer does vector operations

From Scalar processor to Vector processor It seems to me that we are in a transitional phase where the architectures of hardware solutions for natural language computers are just being formed. We are designing this architecture on the basis of scalar processors. But in the future, when this architecture is more mature, it is quite likely that there will be specialized vector processors

It's pretty universally acknowledged as false for humans, but perhaps for machines.

I suspect there would be a huge amount of confounding that would apply in the case of natural language, though.

I think the consensus position on Sapir-Whorf is that the 'strong' SWH (language constrains thought) is false, but there are weaker hypotheses that might be true - especially where optionality is at issue.

For example, marking by grammatical gender is optional in English but not in some Romance languages. Does grammatical gender make gender more salient to speakers of Romance languages (because they have to keep track of it)?

There are some pretty big claims there

I just wonder whether the compute power is going to be there. Self hosting of any kind of generative model, especially something as big as an LLM sophisticated enough to act as an Agent, is at the very limit of what e.g. an Orin NX can handle.

In this channel we will later announce a platform for AI agents, let's say an operating system for Agents. We will not rely entirely on Orins and there will be the ability to query Open AI and other providers. But the agents themselves, their memory, functions and data will be hosted locally

Aren‘t there plenty of studies about the different perceptions of colour? I remember seeing that lots of times for green vs. blue in forest vs. island scenarios. Also for areas that come up with >10 types of snow in their language whereas most just have 1-2.

The colour thing is real, the '10 words for snow' thing is a myth

again it's about optionality - if you have to pay attention to the shades of colour, you will.

Even in higher age with reduced neuroplasticity? I once read a book called "How we learn" and it even claimed, that the difference between the sounds R and L cannot be learned at a certain age.

I'm only a little familiar with the L1 research from my undergrad more than 10 years ago, not sure how the story changes for L2

Fwiw what you're referring to with respect to age is called the critical (or sensitive) period hypothesis. In fact there are measurable differences in kid's abilities to distinguish sounds where the difference is not phonemic (needed to correctly understand their native language) from 10 months. R and L in Japanese (lateral vs central alveolar approximant), dark/velarized vs clear L in English (lady and bold have different Ls), word-final voiceless/voiced consonants in Dutch/German ("haff you seen my kitt?") etc. As far as we can tell, most kids have no problem learning a new language if they get some input before about 6-7, and then after that it depends a lot on the individual

The book argues, that neuroplasticity reduces for different brain regions at different age brackets and thereby makes it close to impossible to learn that differentiation you mention after a certain age.