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.

I'd say natural language >> programming language

Though, isn't a programming language meant to abstract away the CPU in some form? Here, the "CPU" is operating on the natural language directly, so it's more like an instruction set. But the specific thing I'm wondering is if the "instruction set" in use biases the cognition of the natural language computer in some way. For example, in English, we use the terms "dot product" and "cross product" for two vector operations, and English speakers tend to think of these things as (non-commutative, for cross) forms of multiplication, because the word "product" is used. In Russian, the terms translate to "scalar composition" and "vector composition" and I've not met a Russian speaker who thinks of these as multiplication. So, I wouldn't be surprised if two LLM-based mathematics engines, one running English internally and another running Russian internally, had very different approaches to theorems involving vector math.

Yes, could happen but this is where prompt processing mechanisms like chain of thoughts, self reflection and other things may be helpful

While many look at the size of the model, prompting techniques are the second vector of development that will significantly increase the value of LLM. A recent paper on a new prompting technique has achieved results that are significantly better than previous ones, and the difference between GPT3.5 and GPT4 models is negligible https://arxiv.org/pdf/2311.04254.pdf https://cdn.discordapp.com/attachments/1173652086160892004/1177633120925778060/image.png?ex=6573374f&is=6560c24f&hm=d4ef6c54bc2c37b94c2f9c7fbe6c406945536b26fc579afcf3cbabb11bffa496&

The size and architecture of models can be compared to a CPU. Prompting techniques are like abstractions in programming languages. If programming language abstractions had not evolved, we would not be able to extract the value from CPUs that we have today

There is a good explanation of the XoT technique here https://medium.com/@raphael.mansuy/xot-a-new-prompting-technique-for-ai-the-secret-sauce-to-level-up-your-llms-reasoning-prowess-3e19703ab582

Too bad the article is paywalled

Do you mean something like https://www.secondstate.io/

We will focus on a higher level of abstraction than runtime

calling external apis and accumulating / processing stuff?

Right. Focus on agents as an entity, agent groups, communication between agents, task scheduling

would be cool. thinking of handrolling some stuff like that myself. with k8s tools