Complex Text – Why Not Generative AI?

 

Generative AI is very good at the user typing in a prompt - five to fifty words – and it coming back with a much larger piece of text that has been written by one or more humans.

The application of Complex Text is not like that.

An LLM struggles if it must use the text you give it – switching to making a Python graph of the words you have provided, and then trying to make sense out of that. Its best feature – the linking of related concepts through statistics – is lost.

A relevant post - Complex Text and LLMs Right Tool For the Job?

Its other feature – of continually scrubbing the internet for new text – is also not applicable – the complex text is frozen, and can’t be overlaid or interpreted with the newest thing.

Two possible application areas for complex text:

Legislation

A piece of legislation is complete in itself – an Australian Act of Parliament cannot be explained in detail by taking pieces of similar legislation from other states or countries as a guide.

The users:

Lawyers
Administrators
Programmers
The general public
Members of Parliament (fielding inquiries from their constituents)

Robodebt showed it was unwise to allow lawyers to guide programmers on what the legislation meant – they didn’t have the breadth of understanding necessary, or they had an agenda at cross purposes with the legislation. Making the legislation much more transparent to the administrators of the legislation would have helped to head off the catastrophe.

Horizon in the UK is another example of where connection between software specification and software is fraught, Turning the specification into an Active Structure (that is, an abstract working model) would have provided a check on the resulting software, and helped to limit the lies and incarceration of the innocent.

Defence Specifications

A specification for a large military project is a description of various technologies at a moment in time – not going back to a prior technology, or a new technology that becomes available after the specification is written and steel has been cut (although, worryingly, it sometimes includes a new technology to be proven as part of the build). There are other textual  tools, such as Validation and Verification (V&V), which can benefit from a much stronger link to the specification (currently, V&V is little more than a box-ticking exercise, as the Boeing 737 MAX MCAS debacle demonstrated).

The users:

Lawyers for the government and for the Contractor/subcontractor
               Contract managers
               Complex systems engineers
               Specialists – a wide variety. Using a frigate as an example:
                              Propulsion systems
                              Subsystems – helicopter
                              Guns
                              Missiles and Torpedoes
                              Control Systems
                              Communications
                              Radar & Sonar
                              Electronic Warfare
                              Procurement

Over the life of the project, people will come to have a good unconscious understanding of how everything links together. That doesn’t help in the first few years, when many important decisions need to be made.

Humans have a severe limit on the amount of information that can be held live in their conscious mind – the Four Pieces Limit. What is worse, they are largely unaware of this limit, and put its effects down to inattention or lack of knowledge, instead of taking steps to work around it (not easy).

What we are suggesting as a means of handling Complex Text is Active Structure.

The Australian National Audit Office (ANAO) was lauded for its report eviscerating the Hunter frigates project –

“Design immaturity has affected Defence’s planning for the construction phase, led to an extension of the design and productionisation phase at additional cost to Defence, and diverted approved government funding for long lead time items to pay for the extension and other remediation activities.”

 

The report also harps on (140 times) the lack of a value for money analysis. ANAO made no effort to provide one – probably for a good reason.

Defence is planning to spend $350 billion on five nuclear-powered submarines – if Anti-Submarine Warfare (ASW) was effective, this is a lot of money that would be wasted. Instead, for our submarines it is assumed that ASW ships of an opponent will be largely ineffective. This cuts both ways. If the fleet of frigates costing $45 billion could be destroyed in the first few days of hostilities by missiles fired from thousands of kilometres away, and costing no more than a billion dollars to do it, this is a severe economic loss, as well as a loss of capability, which, for a different project (the submarines), we are saying will be ineffective anyway. Why bother to destroy them if they are ineffective – the blow to morale.

 

Of course it is acceptable to have two methods of skinning a cat, but not having oversight across the two methods is not a good idea.

 

The ANAO report points out that as of late 2023, there were 730,000 documents related to the frigate project – far too many for any person to have oversight.

 

How would converting some of those documents to an Active Structure help?     

Two main reasons. The Four Pieces Limit and Specialisation.

The Four Pieces Limit

This reveals itself in several different ways. If the person knows little about the area, they will fasten on up to four different factors, say A, B, C and D, and attempt to find a solution. The fact that changing A changes X, which changes C will escape their attention, as X is treated as part of everything that remains constant. The solution they derive may well be a good one at that instant in time, but drifts to become a poor solution. How would Active Structure change this? Everything is linked together through the connections described in the complex text, so if you change A, X changes, and so does C. The person is forced to find a better solution, but has an abstract working model to help them, a working model where each word has the precise meaning reflecting its particular context in the text.

If the person has spent their professional career developing an unconscious understanding of how things work, they will protect that knowledge. There is a saying in science – “someone has to die for a new idea to take root”. It was true for “The Origin of Species” – the leading biologist of the day was a fierce opponent. Only when he died could the idea gain a foothold. That may seem to have been long ago, but it was just as true when it was suggested that bacteria cause stomach ulcers – one of the discoverers had to give themselves ulcers and then cure it with antibiotics before the idea was taken seriously.  It is bad enough now, that some people protect their knowledge fiefdoms by not allowing their workers to try new things.

By knowing their limitations, and knowing that other people have the same limitations, people can be much stronger in overcoming them by using technology.

Specialisation

With technology becoming increasingly complex, specialisation is needed. That comes with a cost, as the specialist has little knowledge, or interest, in specialities that abut theirs – legal requirements and missile systems, or missile systems and control systems for example. Having words backed up by their meanings in a dense undirected network structure allows a much closer linkage of the different technical dialects, and more collaboration from people who would otherwise remain silent.

Active Structure would seem ideally suited for Complex Text - it should, the causes of multi-billion dollar stuffups in legislation and Defence were a guiding light in its creation.