Socrates authority

 Today 10:02 AM
THOUGHT OF THE DAY:     
"[There is] only one good, that is, knowledge, and only one evil, that is, ignorance; wealth and good birth bring their possessor no dignity, but on the contrary evil." - Socrates, quoted by Xenophon in the Symposium.  ?!?

That quotation is very much in the spirit of Socrates, though like many sayings attributed to him, it comes to us through students rather than from Socrates himself. The source is the ancient Greek writer and soldier Xenophon, who presents a version of Socratic thought in his Symposium.

The striking part is the first claim:

    "Only one good, that is, knowledge, and only one evil, that is, ignorance."

That is a classic Socratic position. Socrates tended to argue that people do wrong because they misunderstand what is truly good for them. If someone genuinely knew the good, they would act accordingly. Modern readers often object that people sometimes know perfectly well what they ought to do and do otherwise, but Socrates pushed the identification of virtue and knowledge much farther than most of us would.

The second part is even more provocative:

    "wealth and good birth bring their possessor no dignity, but on the contrary evil."

Here he is attacking aristocratic assumptions. In Athens, noble ancestry and wealth conferred status. Socrates is saying that neither money nor pedigree makes a person admirable. Worse, if they foster arrogance, complacency, or the illusion that one is wise when one is not, they become obstacles to virtue.

One way to paraphrase the whole thought in modern language might be:

    The only thing that truly improves a person is understanding; the only thing that truly harms a person is misunderstanding. Money, family background, and social status are morally neutral at best and often become sources of corruption.

What fascinates me is how radical that would sound in almost any society, ancient or modern. Most cultures grant at least some prestige to wealth, lineage, education, office, or celebrity. Socrates cuts through all of it and asks a single question:

"Does this person actually understand what is good and live accordingly?"

That is a remarkably demanding standard.

There is also a subtle irony. Socrates himself came from neither extreme poverty nor great wealth and spent much of his life annoying distinguished Athenians by suggesting that a poor cobbler who understood justice might be wiser than a rich statesman who merely thought he did. That attitude eventually made him a lot of enemies.

There are perfectly well educated and well informed people who do wrong just for the pleasure of it or out of complete uncaring.  

That distinction is exactly right. Rule-based systems fit some insurance environments much better than others.

A large personal-lines carrier such as Allstate or State Farm has enormous volumes of relatively standardized risks. Auto and homeowners underwriting can often be expressed in fairly stable variables:

    location,

    age,

    vehicle type,

    claims history,

    credit-related factors (where permitted),

    coverage choices.

The volume is high, the products are repeatable, and the goal is consistency and efficiency. A rule engine can be very useful there.

Reinsurance is almost the opposite. A reinsurer is often looking at unusual, sparse, and highly contextual risks:

    a petrochemical complex,

    an earthquake accumulation,

    a satellite launch,

    a multinational catastrophe program,

    a specialized manufacturing operation.

The problem is not simply "given these inputs, apply these rules." The hard part is often determining which questions should even be asked.

A primary insurer may have millions of similar policyholders and decades of loss data. A reinsurer might be evaluating something where there are only a handful of comparable examples in the world. The underwriter's value is often in recognizing the unusual feature that changes the whole analysis.

That also fits your earlier clean-room example. A rule system might have a rule like:

    "Fire protection: sprinklered — favorable."

But an experienced underwriter asks:

    "What happens if a tiny smoke event shuts down a facility where contamination control is everything?"

The important insight is not in the rule; it is in understanding the business process behind the risk.

In a way, reinsurance was a poor match for first-generation AI because so much of the expertise was tacit. The expert could often recognize patterns, but those patterns were not always easy to turn into explicit decision trees.

Interestingly, modern AI approaches are better suited to some of those problems because they can find patterns without requiring every rule to be written down in advance. Whether they can capture the judgment of a seasoned reinsurer is another question—but the problem they are attacking is much closer to the real nature of expert work.

Your move from underwriting into IT seems particularly logical in hindsight: you were sitting at the boundary between a domain where human judgment mattered enormously and a technology trying to formalize that judgment. That boundary has been one of the most interesting places in computing for decades.

Well I would leave off satellite launches.

Fair correction. Satellite launches are probably a good example of the kind of unusual, high-severity risk reinsurers may encounter, but they are also a specialized corner of the market and not representative of the day-to-day world most reinsurers live in.

For many reinsurers, the real challenge is less "one-of-a-kind exotic risks" and more complexity and accumulation. A reinsurer might be looking at:

    a global property portfolio where thousands of individual risks combine into a single catastrophe exposure;

    a major industrial account with unusual business interruption characteristics;

    a catastrophe treaty where the question is not one building but the interaction of many locations;

    liability exposures where the development pattern can unfold over decades.

Those are difficult for rule-based systems because the issue is often not that nobody knows the rules. It is that the important variables are subtle, interconnected, and sometimes qualitative.

Your clean-room loss is a perfect example. A simple system might have recognized "semiconductor plant," "fire protection," "sprinklers," and "expensive equipment." The expert question was different:

    "What is the consequence of a disruption to a process that cannot tolerate contamination?"

That is not really a rule; it is an understanding of how the business works.

The interesting irony is that reinsurance underwriters often pride themselves on judgment precisely because they are dealing with situations where there is not enough historical data to simply calculate an answer. The art is knowing which similarities matter and which do not.

So yes, satellite launches are a bit of an outlier. The more typical reinsurance problem is something much less glamorous but much harder: taking a pile of incomplete, imperfect information and deciding whether the price is adequate for a risk that may not behave the way anyone expects.

One could write such a rule into the system, but what "expert" is out there who would think of it, and if they did, it makes another question on the input screen.  For those systems to work required data input, and not the sort of data a data processor knows how to enter.

Exactly. You have put your finger on one of the fundamental problems of early expert systems: the bottleneck was not the rules engine; it was knowledge acquisition.

A programmer could easily write:

    IF clean room contamination occurs THEN consider extended business interruption exposure.

But the difficult question is: who knew to tell the programmer that this was the relevant distinction?

The person who knew it was probably a veteran underwriter, risk engineer, claims specialist, or someone who had lived through a loss. And even then, extracting the knowledge was difficult. Experts often do not carry their knowledge around as a neat list of rules. They recognize patterns, ask unusual questions, and notice things that are hard to articulate.

You also identified the input problem. A computer system is only as good as the information it receives. A data entry clerk can enter:

    building value,

    payroll,

    limits,

    deductible,

    construction type.

But asking:

    "Could a small contamination event interrupt production for six months because of validation procedures?"

requires someone who understands the business.

This was a major reason many expert systems struggled. They assumed that expertise could be transferred like a document: interview the expert, capture the rules, load them into the machine. In practice, much expertise is situational and experiential.

There is a classic distinction between:

    explicit knowledge — things you can write down ("sprinklers reduce fire loss severity");

    tacit knowledge — things you know from experience ("this account bothers me because something about the operation doesn't fit").

Insurance underwriting has always had a lot of the second kind.

The irony is that the best expert systems often required the most expert humans to build and maintain them. The "AI" did not replace expertise; it amplified the parts of expertise that could be formalized.

Your comment about the input screen is especially important. Every additional question is a cost:

    someone must know the answer,

    someone must collect it,

    someone must enter it,

    someone must keep it current.

A system with 500 clever questions that nobody answers correctly is worse than a simple system with 20 reliable ones.

In that sense, the lesson from those early systems was not that automation was impossible. It was that the hard part of intelligent systems is often not computation—it is knowing what matters. That remains true even with today's much more powerful AI.