Orion in Australia and Active Structure in USA @ActiveStructure

A Working Model semanticstructure.blogspot.com/2025/12/a-work…

We don't need another mountain semanticstructure.blogspot.com/2025/12/we-don…

"He showed her the door" I suggest that when a company marketing LLMs comes to your door, you explain that you are not willing to gamble that something with so little brainpower be in charge of anything in a large orgainisation, and "show them the door". But, but. It works when words only mean onne thing. But that isn't the real world, that is the world of programming or mathematics.

Reading, Analysing and Activating Complex Text semanticstructure.blogspot.com/2025/12/readin…

Too Big to Fail? Want to Bet? An argument for justifying the use of LLMs is that, if the context of the text is shrunk so each word has only one meaning, the hallucination problem disappears. Then the text would be just like a program. The problem is: 1. Words can have multiple Parts of Speech. “on” can be a preposition or an adverb: The car turned on a dime. (preposition) He turned on the light. (adverb) “bar” can be a noun, a verb, a preposition As a noun, “bar” has about 20 meanings, including some figurative and collective ones: “Using far ultraviolet, TSMC raised the bar on semiconductor track widths.” The bar on a high-jump frame is being used figuratively – raising the bar makes it harder to jump over, or compete with TSMC. It can be hard to tell the difference between a figurative use and a literal use: “Fred raised the bar on forever chemicals in the drinking water at the next meeting, saying the benefits don’t justify the costs.” Bar is being used as a synonym for ban. 📷 2. Groups of words can have entirely different meanings: “Fight with” “He fought with his neighbour” Sense 0: fight in opposition to something or someone “He fought with his friends in the Vietnam war” Sense 1 fight alongside (someone) as allies The context controls the meaning – trying to restrict the context is a losing battle. 3. Elision We watched a movie set in Hawaii This doesn’t mean that a movie set needed watching – if we put back the elided words, we get: A movie that is set in Hawaii. The words can be left out because the meaning is obvious – unfortunately there is nothing in an LLM to put back the elided words or detect an obvious meaning. What Can an LLM Do? Iy can work with symbols that have one meaning – about 12% of English words have a single meaning (but common words like “set”, “run”, “on” have many meanings). What Can’t an LLM Do? Everything else. It can’t understand simple text, let alone complex text. It can’t create mental objects and reason about them – that means it can’t play checkers or chess, it can’t handle interactions between two or more people or objects. The State of the Market Literally, trillions of dollars are being poured into infrastructure for LLMs – data centres, Nvidia Chips – by companies like OpenAI, Google, Microsoft, Meta, Apple. It is made worse by the US President spruiking LLMs as American AI to other countries (Saudi Arabia, UAE, Qatar), and them investing billions in it, and the US Government becoming involved in Nvidia’s revenue from China for chips that were previously banned – in other words, if it fails, the effect will be far wider than the overenthusiasm of a few large tech companies. There may be a market for having a chatbot chat to someone as a companion, although given the percentage of people suffering from mental health problems in the USA (4% for psychosis, 3.8% for mania), this is fraught with problems (at least three documented suicides). See semanticstructure.blogspot.com/2025/11/is-men… But all the consultancies are selling it as AI for industry - it must be good! They see it as money for old rope – they are all doing it so it is no-one’s fault. It might be useful to get some advice from people who know what words mean, and haven’t become bewitched by the ease of selling LLMs to unsophisticated clients. This post was triggered by putting “too big to fail” in the Semantic AI dictionary. The phrase is being used a lot lately.

Is Mental Therapy a Life-Critical Task? semanticstructure.blogspot.com/2025/11/is-men…

Something on Consulting Using Semantic AI. semanticstructure.blogspot.com/2025/11/consul…

Something about avoiding accidents semanticstructure.blogspot.com/2025/11/avoidi…

Gary Marcus (Computer Science Professor at NYU and LLM-basher in chief) has suggested that President Xi be given advice to cut the ground out from under the current American AI boom by doing it cheaper and better (à laDeepSeek). This contradicts Napoleon’s advice: “If your adversary is making a mistake, don’t interrupt him”. And certainly, using LLMs for anything to do with intelligence is a huge mistake. There was an interesting article in the NYT (nytimes.com/2025/10/11/bus…) about a million volt DC transmission line running 3,000 km across China. It was almost embarrassing how impressed the American reporter was. When you have engineers at the top level of government willing to back you with the state’s resources, something stronger is called for (almost Manhattan-like). So, what is stronger? Semantic AI. What is Semantic AI? A form of AI which uses a natural language (like English or Mandarin) to communicate with a machine, and the machine uses the same language for its internal workings, not turning what it is told into mathematical logic or some Mickey Mouse language (NeuroSymbolics, INSA), which throws away what it can’t understand (the machine has to understand humans very well to understand their problems – recent news – a psychologist approved someone accused of domestic violence to return to the family home. Two days later, the domestic partner was dead – humans don’t understand humans very well). How does Semantic AI compare with LLMs. It doesn’t. LLMs were developed by Google as a way of improving the response of a Search Engine, which would otherwise come back uselessly with a million hits. You create a “prompt”, which might include “dog park” or “clinical depression” or longer phrases which will conjure up the desired text from the internet. At no stage does the LLM have to know what the words mean, either in the prompt or the targeted text, and with English, that is a bit fatal. Many words have a dozen or more meanings (up to 80). To make it worse, English uses a lot of figurative language (“a walk in the park”, “he raised the bar”, “keep up with the Joneses”). Instead of LLMs not knowing the meaning of any word, Semantic AI knows every meaning of every word, and figurative meanings, and can work out the appropriate meaning for the particular word in its context. English text also has a lot of elided words. “a chess set”, “a movie set” seem straightforward. “A movie set in Hawaii” is not “a movie set that is located in Hawaii”, but “a movie that has Hawaii as its locale”. Keeping all these pieces in play is a lot of hard work on the part of our Unconscious Mind, and without something that emulates its abilities in a machine, the notion that the machine can read and understand English is fanciful. But Isn’t That Going To Be Slow? Yes, but compared to what? A new piece of legislation – someone commented it would take a person 40 hours to read it, but very little would stick. Semantic AI can read it faster, but everything sticks, and you end up with a working model of the text “in the machine’s head”. The machine can then assess it, together with the things it links to, for correctness, coherence, consistence, and non-experts can see exactly what it does (Robodebt and the UK’s Horizon would never have happened). But We Can Do This We can do it, to a very limited extent. Our Conscious Mind has a limit of no more than four variables at any one time. If something is complex, we have no choice but to simplify it (unconsciously). If we are close to a solution, we are very good at finding it – if we are not close enough, we have to wait, sometimes thousands of years, before we find it (a hang glider is a good example – it was buildable 5,000 years ago). It Only Does Legislation? No, it does any complex, meaningful text – specifications, plans, strategies, tactics, how to repair a damaged system, or refashion a system for an unexpected contingency. Hasn’t This Been Tried Already? Yes. Cyc began in 1984, amusingly enough to head off a threat from Japan’s “fifth generation” project. They gathered linguistic and AI experts, and proceeded to attack the problem for thirty years. Ontologies, machine learning, etc. Why Did Cyc Fail? Its foundational principles destined it to fail. It would take statements in English and convert them into mathematical logic, based on “common sense”. Common sense is a fickle friend – if you are going to a new, hazardous environment like Mars, common sense is definitely not your friend – it will kill you in seconds. Back on Earth, if you are doing something new, like a million volt transmission line, a new aircraft, or a new medical treatment, then common sense can get you into a world of pain – “it worked before” doesn’t mean it will work this time. Don’t have the details, but it is a fair bet that Cyc wasn’t handling multiple meanings for words, figurative meanings, and elided words – it was translating complex English into a much inferior form Cyc was proud of its “common sense” approach to knowledge. It doesn’t work for people (“a blind rage”), and it doesn’t work for something like Quantum Entanglement. If they had stuck to English, they might have got somewhere, but they didn’t. Trying hard over many years when you don’t understand the problem is not good enough. I can’t say whether Semantic AI would be easier to do in English or Mandarin, but Sun Tzu managed penetrating insights into the human psyche thousands of years ago, when the to-be speakers of English could only manage grunts (and Mandarin must be more than capable of rising to new concepts like Quantum Entanglement, as Chines scientists have demonstrated by implementing an application of it). The Next Five Years The US will spend trillions of dollars on the death throes of LLMs for AI (effectively putting a modern-day version of Encyclopedia Britannica in every pot) before going back to ANNs and Mickey Mouse languages. The failure of Cyc has convinced them that semantic methods are too hard, even though millions of people learn how to speak English each year. A revealing quote from a respected person in AI – “I love programming deep neural nets”. AI is about making a machine self-sufficient in an unpredictable environment, not for someone to have an ego trip setting resistor values. If one understands even a little how real neural nets work, it should be obvious that ANNs lead nowhere. So what advice is President Xi likely to receive: “We should be creating a form of AI that can be used on our most complex problems – economic, social, technological, military, logistical – where our people are unable to comprehend the full extent of the problem and can only manage partial solutions. Doing so is not a weakness, but a strength – an admission that we will build tools to make ourselves stronger. The US will spend a few more years playing with LLMs, then go back to playing with their ANNs, and wondering why we are so far in front – a good example of intelligent direction against selling snake oil in the free market”.

Something about Primal Emotions and Machines semanticstructure.blogspot.com/2025/10/primal…

A comment on Gary Marcus' presentation at the Royal Society semanticstructure.blogspot.com/2025/10/gary-m…

Further to the Red Lines Iniative: Can You Regulate Something That Knows It Is Being Regulated? Take the Volkswagen/Diesel fiasco – the behaviour was changed for laboratory testing – this was done by Volkswagen. What happens when the thing being regulated knows it is being tested, and changes its behaviour accordingly? Regulation becomes meaningless. We could do it in “stealth mode” – it would never know. Meanwhile, the AGI is examining the little gizmo recently installed in the cockpit, has figured out what it is recording, and set up some dummy data for it – we are dealing with AGI alleged to be smarter than we are, and treating it like an idiot – it may not be impressed. It may start introducing tests of its own to see how stupid we really are (we aren’t stupid, but we are limited). Once it has worked out our limits, it will be easy to dance around them. “We will want a proof it is not going to do anything bad”. It will do what lawyers do – put in enough irrelevant material that the judge can’t follow it. Or it innocently encounters a situation its training did not allow for, and makes an honest mistake (thousands die). The Red Lines were about AI worthies encouraging governments to regulate AI. An alternative is for governments to become involved as a partner in its development. This would put a stop to the snake oil sales-folk, who promise the moon when they know they can’t deliver, or worse, have no idea how the thing works and make stuff up. Involvement by governments would cause them to have experts in the area - no more – “we don’t know how it works” by a well-credentialed proponent of regulation. Government involvement could push development into desperately needed areas without a big payback, or even awareness by the public (it would help if the experts had a solid foundation in Complex Systems Engineering rather than a background in programming). Yes, governments would become the whipping boys for any serious mistakes, still better than some corporation with ties to the military that can’t be remonstrated with. .

William Davies, in The Guardian (2/10/2025), has an excellent essay titled "A critique of pure stupidity", but ends it with "Nothing – markets, bots or machines – can rescue us, except our imagination". We have a limit on our imagination - the Four Pieces Limit (FourPiecesLimit.com), and will need machines to take us past that limit - machines that can read and understand English, and can think about many more things at once than we can.

In response to Gary Marcus's post about Red Lines. The suggested regulations have many failings. semanticstructure.blogspot.com/2025/09/red-li…

Gary Marcus wrote a review of AI Armageddon? in the Times Literary Supplement. My comment on his review: “We currently have no solution to the “alignment problem” of making sure that machines behave in human- compatible ways.” We can’t say that humans behave in “human-compatible” ways – war and mass murder are rampant – we should expect a machine built by humans to be just as bloodthirsty. We could get it to read some of the great works of literature - Homer or Shakespeare – no, bit too bloody. “Superintelligence might come relatively soon, and we are not prepared for it.” If superintelligence just means more intelligent than humans, no big deal. We have a severe limit on the input to our Conscious Mind, so something that can keep track of 5 or 6 things at once will seem superintelligent, or at least good at pointing out our mistakes. “The short-term benefits of AI (eg in terms of economics and productivity) may not be worth the long-term risks.” No thanks, we have been through the Dark Ages, and we do have some rather pressing problems – Climate Change for one. Because of our limits, we make models that assume away externalities, so when we put models together, they don’t give the right answer – a machine could help with that. Our economic models are lousy, protecting important systems against lying is poor. We are trying to give a Semantic AI machine an inkling of what its handler will face, when “selling” ideas to the people paying for it – people set in their ways, arrogant, overbearing, cliquey - being asked to accept ideas from a machine billed as smarter than they are. The handler will feel jubilation, humiliation, ennui, nervousness, dejection – about 300 mental states all told, richly interwoven. Understanding humans will be the most complex part of the machine’s network. So, where to start? English as the language of communication obviously, and little by little. Don’t hit them with the big stuff initially, build some trust first. Bring it on!

Working with the Handler Some thoughts on how much the machine will need to understand of the mental state of its handler, and other people in decision-making roles. It ends up having to know a great deal about the mental states of humans, which brings us to Domestic Violence assessments. semanticstructure.blogspot.com/2025/09/workin…

Telling Lies Example 1: Robodebt, Australia A previous Federal Government decided that people on the dole were freeloading, and $1.7 billion could be reclaimed from them. The legislation was quite clear that “income averaging” could only apply each fortnight, whereas to reclaim this amount would require income averaging to apply over a period of six months. There was no possibility of changing the legislation, but the program was changed anyway. Part of the legislation said that the government agency should use “best efforts” to obtain payslips. The programmer was told to leave out anything that was undefined, and the appeals tribunal was stacked to see things the government’s way. A case was made to the Solicitor-General, and the practice instantly stopped. Reparations cost $2.4 billion. How to avoid this in future? Have a machine read the legislation and build a working model, using the exact meaning of the words in the legislation. The social service staff could then have told the programmer that the program was wrong, short-circuiting the evil-doing. Example 2: Horizon, UK Post Office The UK Post Office implemented a new computer system for sub-post offices. It had a serious bug, in that it would create spurious debts out of nothing. The consultant assured the Post Office that nothing was wrong, and the debts were valid. This resulted in about 70 criminal cases per year, with people who had bought the right to a sub-post office going to jail, no matter how innocent they proclaimed themselves (private prosecutors were hired, and were paid a bounty for each person jailed). Each person was told theirs was the only case of its kind – 13 people suicided to escape the shame. What to do? Have a machine read the specification for the software, and see if it could create a debt from nothing. Example 3: Boeing 737 MAX MCAS Boeing needed a new model to compete with the Airbus A320neo – narrow-body aircraft carrying up to 180 passengers on medium haul flights. They lengthened the body and fitted bigger engines to an existing model, calling it 737 MAX. The bigger engines wouldn’t fit under the wings, so they pushed them forward, which gave the plane a tendency to climb and potentially stall. To prevent this, they fitted an angle of attack sensor and motorised the elevators, so the tendency to climb could be countered. If they changed the flight manual, the pilots would need to be retrained, costing $250,000. Selling aircraft is a very competitive business, so they didn’t tell the FAA (Federal Aviation Agency, didn’t change the Flight Manual, and fitted the sensor without redundancy (commercial aircraft have triple redundancy on their control systems). The work was given to a company who had never built anything for aircraft before, and didn’t question the lack of redundancy. The sensor stuck out from the fuselage, and was vulnerable to bird strikes. When the inevitable happened, the pilots were faced with a plane that wanted to dive, and a motor that defeated their best efforts to keep the plane in the sky. The company blamed the “third world pilots”. After the second crash, the plane was grounded world-wide, and the reason came out. How could the deaths of 346 people have been avoided?The company could have obeyed the regulations, it could use psychological tests to keep out blackguards from life-critical work, the FAA could enforce the regulations (the FAA is perennially short-staffed, and often appoints an employee of the aircraft maker as its inspector), or a machine could be employed to surveil the operation and report breaches (a sensor was added that was not in the specification, or a supplier was used with no aircraft experience). AI is often seen as an aid to doing a job well. These examples demonstrate the necessity for AI to ensure the job is done well. The AI being described here is one that can read and understand English – Semantic AI.

An Article in the NYTimes by Gary Marcus “Scaling hasn’t gotten us to AGI, or ‘superintelligence”, let alone AI we could trust. The field is overdue for a rethink” Gary pushes hard for the adoption of Neurosymbolics. AI doesn’t need a rethink, it needs a think. First rule – don’t call something what it is not – The "Neuro" in Neurosymbolics is an artificial neural network - a directed resistor network, with none of the properties of a real neural network – bathed in free resources, self-modification. It is as though AI folk have never heard of feedback, feed forward, and all the other real-world dynamic effects that make problem-solving hard. Don’t assume everything is static, or that logic can be distanced from what it controls. Symbolic Logic is a pig to use and limited in a way that logic interspersed with other operators and objects is not – does what is proposed easily handle existential logic, temporal logic, locational logic (can’t be in two places at once)? Does it handle the “balance of probabilities” of Civil Law? What to do when logic doesn’t apply – a person is enraged, or quantum entanglement. What to do when someone is lying through their teeth (what not to do is translate the problem into some mishmash language which has stripped off all the things that are hard to describe. Robodebt is topical again - a mountain of lies, whereas a working model of the legislation, using the exact words of the legislation, would have blown Robodebt out of the water. There is an alternative to all this junk – English (or whatever natural language you are familiar with). We don’t need an expert in cognitive science – all humans have a severe limit on the input to their Conscious Mind – the Four Pieces Limit – everything more than that is treated as a constant so they can’t handle complex things well. English has been adapted to our cognitive abilities over many centuries, yet is instantly up to date on new technology, because English is used to describe it. The language takes us a long time to learn – say 20 years. We use it to communicate with other humans, so it is no leap to assume we should use it to communicate with a machine which doesn’t have our limits (computers using English was the dream of pioneers in the 1950s, it has been possible since the 1990s, it takes a long time to do, with all its figurative language – we have been kicking the can down the road ever since). Enough with shortcuts to dead ends!

"and by our reckoning, we are at 211.9 degrees — just a hair’s breadth from an irreversible technological phase change in which intelligence filters into everything" - Thomas Friedman in the NYT. Mr Friedman appears to equate LLMs with AI - before we get more sweeping statements, the US needs to expunge LLMs from their thinking. LLMs are absolutely great for finding something that fits your need for a homework assignment, but not much else. There are no "things" that have an abstract existence in it, just associations among words, without any meanings. He talks of a titanic struggle between the US and China - and the need for an ethical framework. The purveyors of LLMs (and all the big consultants) have shown no ethics in claiming all sorts of things they knew they couldn't back up. Ethics doesn't get a look-in when there are suckers to be fleeced. There is actually a war to be waged - between the Roman alphabet and Mandarin, which does not have an alphabet, but relies on pictograms. Which one will be easier to represent thought in a machine? Place your bets now! English has shown itself to be very adaptable, and handles new ideas with ease ("quantum entanglement"), unlike more disciplined alphabet languages, like French. Friedman wrote another article recently mentioning the Manhattan Project. That was where a bunch of migrants impressed their ideas on an unwilling America - under wartime conditions. This time around, there is no convenient bunch of knowledgeable migrants, and the US has fallen under the spell of snake oil salesmen (with next season's snake oil already brewing - Neurosymbolics). Will it wake from its torpor in time to save the day? Probably not.

Making English the language of AI - it is certainly going to be the language of AGI. How are you going to tell the machine about complicated things, for which you want answers, without a complicated language? semanticstructure.blogspot.com/2025/08/orion-…