What neuro-semantic reference architecture will survive?

[alex.turkhanov]

Why can the future AI landscape be full of upright bikes and QWERTY keyboards? The social aspect of the tipping point (or product-market fit), a 40-year-old concept and market segmentation based on technological beliefs.

Some sociology background for the standards war in AI that is supposedly coming.

Many engineers ignore the social aspects of technology and think that the technical superiority of one product is enough to win the competition. Let’s consider an ancient example from the history of technics, the standards war of recumbent vs. upright bicycles, that proves otherwise and makes us reflect on the marketing positioning of AI solutions we deliver.

Recumbent bike

The recumbent bike reference architecture beats the upright variant in almost any measurable characteristic. You can check the comparison table in Wikipedia. But you barely see this superiority on the streets. The evolution of solution variants is not linear. You cannot untie all these Agile spirals into one thread and build a consistent narrative of “how we came up with this design.” Some critical requirements cannot be put as requirements because they are about politics, social norms, balance of power, etc. You cannot explicate that, and you cannot use them as inputs to your design process, whether they are model-based or not. It went for bikes 100 years ago and applies to the AI solutions we deliver now. Humans didn’t change much in some aspects.

The upright bicycle design alternative survived because it was emancipation time, and women found the upright variant more appealing:

“During the days of the high-wheeled ‘Ordinary,’ women were not supposed to mount a bicycle. For instance, in a magazine advice column (1885), it is proclaimed, in reply to a letter from a young lady: The mere fact of riding a bicycle is not in itself sinful, and if it is the only means of reaching the church on a Sunday, it may be excusable.”

See the upright bicycle case in the 1984 article: http://dtc-wsuv.org/wp/dtc375-spring15/files/2015/02/The-Social-Construction-of-Facts-and-Artefacts.pdf

Rogers provides the tipping point concept in the 1995 book ‘Diffusion of Innovations.’ A tipping point is when an innovation starts spreading or reaches what is now called “product-market fit.” What is often missed is that it’s not only product features (technical characteristics) that play a role in the diffusion. It’s not only a product roadmap. It’s also technology adoption or even technology domestication activities. There is a whole social dimension to it. As always with society, there are several explanations for the successes and failures of the technologies - two of the progressive camp and two of the conservative camp (an extensive article on that topic is coming).

My take

AI is an interesting set of technologies; they are highly polarized. I see how these four cohorts (progressive-conservative and AI is suitable vs. AI is wrong division) emerge in the political movements and companies. They are techno-skeptical and techno-optimistic companies; some adopt and domesticate technologies until they are fit for current processes and do not require organizational changes, and some strongly believe that AI will bring us back to the medieval ages like social networks did.

That means the standards war in AI will also not be global, like the standards war between phone connectors was, with USB type C finally winning the battle. They will be regionalized within and between these social segments. There will be no global AI companies, in other words, and decoupling policies and “COMMISSION RECOMMENDATION of 3.10.2023 on critical technology areas for the EU’s economic security for further risk assessment with Member States” are secondary effects, not causes for that state of affairs. It’s the social dynamics overall.

On the business level, it has implications. When you position your AI company, you should consider these social viewpoints fit to your branding. Silicon Valley’s slogan, “Our technology will change the world,” can ring all the wrong bells in some segments. Some worlds do not want change, and businesses are not universities. We should not educate the market. We should sustainably address the problems the customers bring us. But many companies and investment funds are ignoring these social aspects, and the reference architectures they promote will not survive, despite the technical perfection. A tipping point will be for one segment but a freezing point for others; see the broader range of reactions to ChatGPT’s success, with bans and legislative rebuff.

The future AI landscape can be full of upright bikes and QWERTY keyboards, so we should not rely purely on technical benchmarks when we build our solutions. We can talk all day long about how technologies are neutral, but we will not convince some of our customers because AI is not unbiased for them.

[alex.turkhanov]

Re: TRUSTLLM: TRUSTWORTHINESS IN LARGE LANGUAGE MODELS

You cannot trust their answers - but again, who said that is always a problem? They can be used to produce good quality Socratic questions - them you should not trust, hence, no problems with the trust.*

• However, very few researchers understand the problem of building trust architectures (not from cyber security, but from the epistemology/ontology engineering point of view), but that’s a bit of a different topic for another day. I can recommend reading Devlin, Keith. «Confronting context effects in intelligence analysis: How can mathematics help», 2005.

[alex.turkhanov]

When we deliver a solution, we provide a product and a technology package that enables its usage. For Terraphim + Atomic Server (as an ontology and data model editor), that would include:

1. Logseq is a tool of choice for the personal knowledge base.
2. Collaboration text processor, Typst GitHub - typst/typst: A new markup-based typesetting system that is powerful and easy to learn.
3. Zotero for reference management.

Compare Atomic ontology editor with that one of Ontological Modeling Language v2 (open caesar.io), and you see why formal ontological modelling seldom takes off in commercial companies.

[alex.turkhanov]

Re-reading Ciborra, Claudio. The Labyrinths of Information: Challenging the Wisdom of Systems. Oxford: Oxford University Press, USA, 2004.
It appears to me that there are two architectures for semantic/deterministic AI - one for idealised index-based data architecture, the framework, and another one for real ways of working, personalised and tinkered for circumstances.

[alex.turkhanov]

Almost any business strategy is based on a process model. But when it comes to implementing a process model, the question of what a process is exactly arises.

For example, we can conceptualise it as an actor as Applied Akka Patterns [Book] (oreilly.com) proposes:

The Actor Model

• Reality Is Eventually Consistent
• Deconstructing the Actor Model
• All Computation Is Performed Within an Actor
• Actors Can Communicate Only Through Messages
• Actors Can Create Child Actors
• Actors Can Change Their State or Behavior
• Everything Is an Actor
• Uses of the Actor Model
• Defining Clear Boundaries
• When Is the Actor Model Appropriate?

With such an approach, it is clear how to go from describing processes as “inputs-states-outputs” to their technical implementation.

[alex.turkhanov]

Any product sits on top of some technology package and evolves with it. Its growth rate depends heavily on how fast those foundational technologies grow, and its potential available market share is limited by the number of workplaces the foundational technology package occupies.
That’s why it’s critical to select the foundation of the product and stick to it. This is an architectural choice, and it involves a commitment to the success of the foundational technology. For example, Terraphim uses Logsec and Obsidian to build roles’ knowledge graphs that personalise search results, and part of product discovery is to develop an easy method of building and upkeeping those knowledge graphs.
Also, the product manager needs to consider what products the system under development inserts itself. The search should cover multiple haystacks, and plugging into them should be easy. It should be more than a search; it should be, for example, a feed generation similar to social networks and YouTube. The relevancy function is the Terraphim role, where topics are graph nodes.

[alex.turkhanov]

The AI governance context is different (and doesn’t exist at the moment)

The AI governance context differs from humans’, which matters when dealing with errors. Here is an example. We discussed the news “Air Canada must honour refund policy invented by airline’s chatbot.” And the argument went like this: “People do the same. They make up rules, and the company should follow them.”

But here is the kicker - we fix human errors with ultimately another sequence of actions and controls than we fix chatbots. With people, you make them read actual documents, send them to training, move to another job, or fire them. Your flow with LLMs is entirely different; you can check out with chatGPT. It is a software engineering job you cannot complete with organisational means.

You can substitute people, but you need to build a new governance model, not just train and release a chatbot. There are a couple of hidden presuppositions behind the adoption of LLMs:

1. Chatbots do not copy the behaviour of others, which people do very well. People learn very complex things just by watching them.
2. LLMs can only sometimes communicate as people do. The current governance model is grounded on the fact that the CEO or board of directors can call any human, from top to bottom of the organisation, and have a conversation that fixes the undesired behaviour or gives grounds for contract termination. With software, you need at least prompting skills to do that, and retraining is expensive.
3. People have trust and reputations to protect. Genuine experts exist, and you can trust every word when they are in their field of expertise. With LLMs, you never know how well they “know” the topic. Besides that, they cannot lose face; they don’t care.

So, as with any other introduction of big technology, governance change should follow, but most industry leaders offer only the technological side of the projects, leaving the organisational part to be figured out by others.

[alex.turkhanov]

Personal digital assistants: the missing vault chain

Every month, new solutions for PDAs appear. Dave Karpf wrote a piece explaining why the business model will fail On AI agents: how are these digital butlers supposed to get paid? (substack.com). While I agree, there’s more to it - those vehicles do not have roads, gas stations, and oil refineries to work.
Cars became possible not only because the Ford T was released. For a long time, people used horses to build roads. Also, they discovered oil and learned how to produce gasoline. The whole infrastructure was there before Ford T became possible. And in fact, Ford also built a chain of car workshops to service every vehicle everywhere. The entire technological stack of innovations made the diffusion of it possible.
PDAs are now missing gas stations and refineries (assuming roads are there). What data sources will feed them? How will the data be produced? How do we teach people to fuel the PDAs properly? All those questions now do not have answers, making further progress difficult. A vault chain can be needed to make the progress possible.
Sources:

Billington, David P. The Innovators: The Engineering Pioneers Who Made America Modern, 1996.

Christensen, Clayton M. The Innovator’s Dilemma: When New Technologies Cause Great Firms to Fail. Harvard Business Review Press, 2013.

Christensen, Clayton M., Karen Dillon, Taddy Hall, and David S. Duncan. Competing Against Luck: The Story of Innovation and Customer Choice. New York, NY, 2016.

Rogers, Everett M., Arvind Singhal, and Margaret M. Quinlan. “Diffusion of Innovations.” In An Integrated Approach to Communication Theory and Research, 432–48. Routledge, 2014.

[alex.turkhanov]

Two types of value chains

First, one exciting thing is now clear to me about three strategic analysis tools: the OODA loop, the strategy and tactics tree (Goldratt and the pyramid principle of B. Minto), and the Wardley mapping. The Wardley mapping supports the observation and orientation steps in the OODA loop, and the strategy and tactics tree enables decide and act.

Both techniques are based on the concept of value chains. Opportunities in value chains can be one-time gigs or sustainable businesses. That depends on the top node of the value chain. If you do a one-time optimisation, for example, you analyse with ML the operating mode of the plant or fulfilment centre and execute an efficiency improvement project, you are in the one-time, unsustainable value chain. Once you optimise the top node (plant operation), you have no more business; it’s a gig.

But if the top-level node in your value chain consumes something regularly (like online shopping) or makes regular decisions that your product or service supports (like SCADA or MRP), you’re golden; you have a sustainable value chain.

An important conclusion is that many ML products are suitable for only a one-time optimization and cannot be built into any existing sustainable value chains. They exist in fragile Arctic-like micro-ecosystems, which will disappear when the Sun changes its position on the analemma.

[alex.turkhanov]

An interesting report.

• From Complexity to Clarity: Configuring Success in Industrial Compatibility Management with Semantic Reasoning and Knowledge Graphs

  • links:: SEMANTICS-2023
  • Across various industries and business models, companies face the need to assess compatibility, whether it’s industrial configuration management, terms in contracts, points of a supply chain, buyers and suppliers, and so on.
  • Whether this process is in relation to assemblies, production, or data analysis, it often requires managing and determining millions of combinations to assess whether components fit together and meet certain requirements. Traditional approaches often struggle with the complexity and scale of real-world scenarios, resulting in lengthy calculation times and restricted insights. Knowledge graphs, on the other hand, perfectly fit these needs and overcome the challenges facing current solutions. In addition to performance, such a semantic solution provides data in the context of domain knowledge, using simple yet expressive axioms and rules instead of disparate hard-coded procedures.
  • In this presentation, we will showcase a graph solution developed by Derivo for one of the world’s largest automation companies, which is currently leveraging this technology in their production processes as well as in an interactive customer-facing configuration solution. The demonstration will highlight the cutting-edge capabilities of RDFox, a knowledge graph and reasoning engine, and how it can be used to streamline compatibility assessment in industrial configuration management. With its in-memory architecture and incremental reasoning, RDFox ensures real-time adaptability and scalability, while also providing increased functionality and analytical power over the industry standards.
  • As the complexity of commercial configuration management expands, so too does the need for innovative solutions. Derivo’s partnership with RDFox does just that, propelling us into the future where historical limitations dissolve, giving way to knowledge-based analytics and enhanced decision-making
  • Initial Situation:
  • Our organization, the Fraunhofer Society, faced a challenge in handling the influx of research requests coming in from various external partners and customers. The task of sorting these requests to match with the relevant researchers within our organization (~22.000) was not streamlined and would either cause an information overload for our researchers or be a time consuming and manual task for their managers.
  • Approach:
  • To rectify this, we devised a strategy that leverages our existing internal skill catalogue, which is a hierarchical presentation of about 1000 skills and sub-skills that were self-reported in the Fraunhofer Society. The methodology entailed mapping each incoming research request to the corresponding skills within our organization. To achieve this, we used a pre-trained language model which had been fine-tuned for paraphrasing paragraphs. After extracting features from both the incoming texts and the skills in the catalogue, we determined their similarity. We engineered a pipeline that accommodates parameters such as text segments, similarity thresholds, target levels of catalogue hierarchy, and crafted prompts. We tested this pipeline by labeling an already labeled dataset of research requests.
  • Business Value and Benefits of the Semantic Solution:
  • Our method has proven successful in identifying an optimal set of parameters for labeling incoming research requests. The significant benefit of our approach is its ability to alert relevant employees about incoming research requests that match their skills. This strategy effectively minimizes information overload and ensures employees receive alerts only about relevant details and new research opportunities.
  • Prospects and Recommendation:
  • We are currently integrating this methodology into an internal application, developed by another organizational unit. The application serves as a platform for our external partners and customers to file their research requests digitally. Harnessing the power of language models and organizational resources like a skills catalogue can significantly improve request management and employee productivity.

[alex.turkhanov]

• A New Approach to Taxonomy Creation: Combining Human Expertise with NLP to Extract

  • links:: SEMANTICS-2023
  • To mitigate the risk of subjectivity and time-loss in the process of extracting new taxonomy concepts from regulations and law articles, we present a solution used by the Dutch National Police. The application uses entity recognition to mark existing concepts and generates a subset of new possible concepts. To generate this, Large Language Models are enriched with police-specific knowledge from taxonomies defined and curated by our modellers. These suggested concepts are then verified by these same modellers. This human touch, in combination with the Natural Language Processing tooling, increases the accuracy, efficiency and objectivity of the extraction process.
  • Description:
    • The Dutch National Police faces the challenge of defining a vast array of concepts for their taxonomies and ontologies, aiming to create a unified semantic model that can be effectively utilized across the entire police organization. These concepts need to be manually extracted from regulations and law articles. Given the extensive nature of these documents and the subjective nature of the process, the manual selection of the concepts can introduce subjectivity and takes a substantial amount of time.
    • To counteract these problems, we present a solution using entity extraction and supervised Machine Learning to further populate the taxonomies by learning from the existing concepts. The solution enhances accuracy, efficiency, and objectivity, minimizing the potential for human error while keeping the human touch.
    • With this solution, the Dutch National Police, in partnership with Ordina, uses Natural Language Processing to thoroughly analyse the textual documents. It marks the concepts already present in our police taxonomies and suggests new possible concepts. These concepts are selected by building on existing Large Language Models (LLM) and enriching them with police-specific taxonomies. The suggestions are generated by considering a variety of factors, including the semantic context surrounding police-specific concepts. This is achieved through the implementation of pre-processing techniques such as tokenization and lemmatization, as well as leveraging the power of transfer learning. As a result, the model can be trained effectively to produce meaningful results, even with a smaller dataset. The reason for this being that the LLM provides the base, while the taxonomies add another, more subject-specific layer.
  • The outcome of using this tooling is two-fold:
    • An annotated document that marks the existing concepts within the taxonomy and an overview of those concepts.
    • An overview showing all the concepts detected within the document, including newly identified concepts to expand the taxonomy. These are generated in triples, using the SKOS vocabulary.
    • Both will enable the Dutch National Police to decrease the amount of time required to find and extract new concepts.
    • At this date of submission, the tooling has already been used by our modellers during the analysation of new regulation documents, further accelerating the selection of new concepts. In further development, the solution will be enriched so that not only regulations and law articles can be analysed, but also other types of documents. The long-term aspiration is to enhance this solution by not only extracting new entities but also automating the construction of new taxonomies.

[alex.turkhanov]

What is the market size of a personal assistant?

Let’s consider the markets for four types of products: electric scooters, notebooks, bikes, and cars. We’ll see how vendors sell us products in each case, but users tinker with them to create solutions afterwards.

I use electric scooters to bring my kids to school. We take two; one drives the younger daughter, and on the way back, I pack it into a transportation bag and bring it home. On the bigger scooter, there is an additional bag for a jacket and a chain. I need all those elements - transportation bag, jacket bag, chain, to make the solution work for me. Also, there are a lot of tricks on how to ride it on the way to school and back home, especially on how to cross a busy roundabout. It was not obvious how to use a scooter from the first day; it required thinking and training to establish commutes as operations. I built the solution from five products - two scooters, a transportation bag, a jacket bag, and a bike chain. The electric scooter market for me includes all those things, too; they are part of the solution and how I use it.

The same goes for notebooks. They require a table, a dock station, a second screen, a graphic tablet, support and fan, camera lights, and a few cables. I need them all because I need a workplace where a notebook is a central product, but it isn’t functional. I need over 20 other products (leaving software outside) to make it worthwhile. A notebook market may be defined as notebook sales by Apple, but it consists of many different user products for me.

As you can imagine, I can apply the same logic to cars or bikes—neither is helpful by itself; it requires many other products to build a proper solution that meets your requirements and is helpful for your circumstances. Also, each tinkered solution requires particular skills to operate. For example, how can you go down an elevator with your bike or fix it on a bike rack?

The market is built around users’ practices rather than its central product. The same goes for personal assistants. The market for personal assistants will not be limited to big platforms; there will be many other segments and businesses around it. In other words, we sell products, but they buy systems.

[alex.turkhanov]

That is an interesting explanation of DITA:
The Key to Your Chatbot A-Game: DITA | LinkedIn

“To truly excel, chatbots need more than just scripted responses; they require depth, understanding, and the ability to pull in real-time, relevant information. This is where Retrieval-Augmented Generation (RAG) and Knowledge Graphs come into play.”

[alex.turkhanov]

This one is good.