ML / AI jargon revised/learnt today

 ( a light survey this week, after putting it off for a decade or so ) : 

• - dimension : information-wise differentiated aspect of some quantifiable substrate; grossly synonymous with "rank", "aspect", "degree of freedom", "axis", "characteristic"
• - N-dimensionality : N refers to the number of different aspects needed to describe a datum; usually N is a Natural number ( 0, 1, 2, etc. ); grossly synonymous with "N-scalar", "N-vector", "N-matrix", "N-tensor" because specifically scalars are 0-rank-tensors, vectors are 1-rank-tensors, matrices are 2-rank-tensors, etc.
• - feature : a dimension of data
• - embedding : a set of dimensions used to position a datum
• - intrinsic vs ambient dimensions : "intrinsic" refers to the fundamentally minimum set of dimensions required to position a datum; "ambient" refers to to any ( arbitrarily greater than minimum ) set of dimensions in which a datum is positioned
• - MI / mechanistic interpretability : the name of the concern, when people are running hot macro operations but don't understand the micro operations, and so they loop back to do forensics / epistemology on it
• - RAG / retrieval augmented generation : the general pattern of ( AGENT, RESOURCE ) -> RESULT
• - MCP / model context protocol : in the context of RAG : the name of a particular open-sourced standard protocol, which enables AGENT traversal and manipulation of RESOURCE; AGENTs are on the client-side, RESOURCES are on the server-side
• - rank vs dimension : in the context of TRANSFORMATIONS : where a function is [ a mapping between sets DOMAIN and RANGE ] : "dimension" refers to [ the N-spaceness of either set , whereas "rank" refers to [ the dimension of the RANGE, specifically ]; also see "rank-nullity theory", where KERNEL a.k.a NULL SPACE refers to [ the subset of the DOMAIN which maps to "0" in the RANGE ]
• - SAE / sparse auto encoder : wherein some internal nodes of a cognitive system may have redundant data storage, this refers to forcing-factors applied to encourage specialisation of memory per node 

A succinct description of the modelling concern, for human experience in general ( point of view of a formalist ) : So, the main problem with most people's use of natural language, is that they overestimate the number of INTRINSIC DIMENSIONS required to describe consciousness, because they live in the mess of AMBIENT DIMENSIONS intrinsic to the messy evolution of informal languages.

---

I think you should consider where spatio sensory data structures fall into embeddings. 

Current LLMs are mainly running on embeddings of distances between words and concepts described in words. All words are intrinsically meaningless, so mainly LLMs now are chinese rooms. 

When you switch out the underlying data, replacing words with sensory spatial structures, then the computation of distance between two structures can be done via various methods, to varying degrees of logical rigour. 

That is the future we are moving to inevitably. 

educating boards on AI

AI is fundamentally an augmentation of the human factor. I once designed a framework for organisations where the three main classes of stakeholder ( or in ISO terms, "interested party" / IP ) are 

• 1. investors : moves first with capital, gets paid last
• 2a. staff : the first supplier, gets paid first, closes loop
• 3. customers : pays first

Depending on business model of course, there are other IPs, like 

• 0. the government,
• 2. suppliers ( superset of staff ), 
• 4. neighbours,
• 5. competitors, 
• 6. criminals

etc. 

It becomes interesting to run brief simulations over a cup of coffee about what happens when the cost of doing business for each of those parties gets impacted by AI. Basically if the cost of thought drops for everyone, the government is going to want a higher standard of compliance, investors are going to allocate capital differently, staff either get lazier or more productive, suppliers are expected to increase performance, customers have different behaviours depending on their own operating models, competitors are simply trying to leverage off the same thing in different dimensions, and thieves are getting sneakier.

I generally say that most business problems are talent management problems - and that the only long-term difference between machine and human intelligences is legal personhood. In the pathology of behaviour, we again a simple framework is to determine if the problematic party is (a) stupid, (b) lazy, or (c) malicious - there being various formulations of reduction and expansion about this. So much fun to think about, but more fun to think about as an educator of students who don't understand this thoroughly.

( Someone was looking for board-level non-technical AI-subject-matter training the other day, so I dropped a random comment and buggered off. Thought about it a bit more over the week. )

status of Semantic Web suite adoption

OK - I'm about 60% through surveying Semantic Web standards. My conclusion is that the adoption of these "standards" is still very low. Adoption is high in domain-specific niches : this is a W3C standard, so web-technology companies, and companies that exposure their data on the web are the main champions. Outside of that, biomedical sciences uses it a bit. Almost no one else is doing : RDF, RDFS, RIF, OWL ... etc.

So it's like JavaScript : significant brand-equity within a niche, but the man on the street has no reason to get into it. Yet unlike JavaScript, Semantic Web is much less infrastructure dependent than web browsing - so there's a lot more room for wobble and non-compliance between systems. This may change in the future, but I'm not seeing it fix up right now.

RDF is just way too complicated, without adding a lot of value immediately, and there are simpler, more transparent ways to get the same thing done.