Original thinking on insurance, AI, systems and work. Essays develop a full argument. Field Notes capture questions, experiments and operating patterns. Newest first.

QUICK INDEX — NEWEST TO OLDEST

OCTOBER 2, 2026 — ESSAY — When the Customer Sends an Agent

SEPTEMBER 29, 2026 — FIELD NOTE — What Happens to Simple SaaS Tools?

SEPTEMBER 28, 2026 — FIELD NOTE — How Will We Work Together and Apart?

SEPTEMBER 23, 2026 — ESSAY — Every AI Decision Has an Expiry Date

SEPTEMBER 20, 2026 — ESSAY — The Triage Tax: Decision APIs in Insurance

SEPTEMBER 14, 2026 — FIELD NOTE — Embrace, Use, Extend, Support and Amplify

SEPTEMBER 10, 2026 — ESSAY — AI Can Save Time Without Improving Decisions

SEPTEMBER 5, 2026 — FIELD NOTE — AI Memory Should Support Learning and Unlearning

AUGUST 31, 2026 — FIELD NOTE — Is Your Brand Recommended by AI?

JUNE 22, 2026 — FIELD NOTE — Hybrid Work Is Role-Dependent System Design

JUNE 18, 2026 — FIELD NOTE — AI Governance Must Become an Operating System

JUNE 16, 2026 — ESSAY — The Structure Around the Model Is the Product

JUNE 16, 2026 — FIELD NOTE — Make Work Ready for Humans

JUNE 16, 2026 — FIELD NOTE — Applied AI Is Becoming Visible in Insurer Results

JUNE 16, 2026 — FIELD NOTE — Judgment, Empathy and Critical Thinking Remain Human Advantages

JUNE 12, 2026 — ESSAY — Supervision Without Rules

MARCH 24, 2025 — FIELD NOTE — From Excel Model to Governed API

NOVEMBER 26, 2024 — ESSAY — The Business as a Big Set of Functional Spreadsheets

JULY 14, 2024 — ESSAY — RAG Beyond the Demo

========================================

ESSAY • OCTOBER 2, 2026

When the Customer Sends an Agent

Imagine renewal day when the customer does not visit a website, call a broker or fill in a form. Their agent arrives instead.

It knows the vehicles, drivers, current coverage and preferred deductible. It asks questions, gathers quotes, compares wording and returns when something is unclear. It does not get tired after the third quote. It does not forget to shop next year.

Some insurers and marketplaces will try to block this behaviour. The instinct is understandable. Automated shopping can multiply quote costs, strip away context and turn coverage into a misleading price list. It may trigger paid data calls without producing a bind. Fraud and competitive research can arrive through the same door as a legitimate customer.

But blocking the door is not a durable customer strategy. Agents are becoming a new channel, much as mobile and chat became channels. The question is how to recognize a legitimate agent, establish whose authority it carries, provide complete information and preserve a path to a human when judgment is required.

This will expose weak economics. Businesses that rely on friction, confusing comparisons or customer exhaustion will feel pressure. That is not a reason to reject the channel. It is a reason to redesign the journey.

BUILD FOR THE AGENT AT THE DOOR

Run a controlled shopping exercise against your own experience. Can an authorized agent discover appetite, submit structured information, receive a quote with coverage context and explain the result to its user? Add rate limits, provenance, consent, identity and authority checks. Price the external data calls. Create a human handoff that carries the conversation rather than starting again.

The winning experience will serve both the customer and the customer's software. Treating one as hostile by default risks losing the other.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7511742012237295645/

========================================

FIELD NOTE • SEPTEMBER 29, 2026

What Happens to Simple SaaS Tools?

What happens to Jira, basic help desks and the thousands of software products built around humans filling forms and moving tickets?

They may not disappear. Their visible interface may become less important than the system underneath it. Agents can create, update and reconcile work without a person visiting every screen. The lasting value moves toward trusted state, permissions, workflow rules, audit history and reliable actions exposed through APIs.

That is uncomfortable for software businesses whose product advantage is mostly interface and switching friction. It is an opportunity for products that hold authoritative context and let customers use it safely in new ways.

TWO QUESTIONS FOR EVERY SAAS PRODUCT

For vendors: if the user never opens the application, can an authorized agent still understand the state, take a permitted action and receive evidence that it worked?

For buyers: if agents become the primary interface, can the organization retrieve its context, rules and history without surrendering control to the vendor's assistant?

The next SaaS layer may be less visible but more consequential. Products will compete on how clearly they expose meaning and authority, not only on how efficiently a person can click through the workflow.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7510862336275173376/

========================================

FIELD NOTE • SEPTEMBER 28, 2026

How Will We Work Together and Apart?

Most AI productivity stories describe one person becoming faster. The larger change may happen between people.

An agent prepares the analysis, another schedules the work, a third monitors the result, and a person appears only when an exception crosses a threshold. That sounds efficient until two agents use different assumptions, nobody owns the handoff or the human receiving the exception cannot reconstruct how it arrived.

Coordination is the hidden work of organizations. We use meetings, tickets, inboxes and managers to move context and responsibility. Agents will remove some of that friction, but they can also remove the moments when people notice that their understanding differs.

FOLLOW ONE PIECE OF WORK END TO END

Choose a real customer request and trace every human-agent handoff. At each boundary ask:

- What state is passed?

- What evidence travels with it?

- Who owns the next decision?

- What does the receiver assume?

- How does uncertainty escalate?

Fix the handoffs before optimizing individual tasks. A collection of fast agents can still produce a slow, confused company. The goal is not only better individual output; it is a shared system in which people and agents can work together and know when they are working apart.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7510493117943402496/

========================================

ESSAY • SEPTEMBER 23, 2026

Every AI Decision Has an Expiry Date

Most of what banks and insurers decided about AI last year is already out of date.

That does not mean every rejected experiment should be restarted. Some experiments failed because the model was not capable enough. Others failed because the data was poor, the workflow was badly chosen or nobody could explain the business case. Better models may rescue the first group. They will not magically repair the second.

I have worked in enterprise software since 1982. Being the sexagenarian in a thirty-something industry has one advantage: you learn to recognize the difference between a demo and a platform shift. This is a platform shift. The models improve, the economics change, and the constraints move every few weeks. A “no” based on last year's capability has an expiry date.

The operating response is not to chase every release. It is to make reassessment a habit. Re-run the experiments that failed on capability. Give zombie pilots a kill date. Rewrite instructions that were designed to babysit weaker models. Let agents inspect code and documents, but give each one a named human owner and the least authority it needs. Build evaluations that an underwriter, architect, regulator and customer could all understand.

There is another shift coming from outside the institution. Customers will increasingly send agents to shop, compare and negotiate. A margin that depends on people getting tired, confused or bored at renewal is not a durable margin.

THE MONDAY-MORNING MOVE

Build a simple register of every AI experiment stopped in the last two years. Give each one one of three labels:

- capability failure;

- data or process failure;

- weak business case.

Re-test the first group against today's models. Repair the second only where the underlying work matters. Close the third. Then repeat the exercise quarterly. The code is increasingly available; domain expertise, operating discipline and judgment are the engagement.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7508504858866495488/

========================================

ESSAY • SEPTEMBER 20, 2026

The Triage Tax: Decision APIs in Insurance

Insurance has paid a triage tax for decades.

Skilled people spend part of every day opening submissions, reading claim notes, sorting emails, checking whether information is complete and deciding who should look next. This work matters, but much of it happens before the real work begins. We ration judgment because judgment has always meant people and time.

Decision APIs suggest a different pattern. Give a service a claim note, broker submission or customer email, ask a bounded set of questions and receive typed answers with probabilities. No essay is required. The useful output may be: commercial property, within appetite, missing loss runs, possible aviation exposure, 82 percent confidence, send to a person.

That can be valuable in customer care, fraud referral, underwriting intake and renewal management. It can also be dangerous. A probability looks precise even when the evidence is weak. A threshold quietly becomes policy. Staff start trusting the score because it arrives faster than a colleague. An adversarial claimant or broker may learn how to steer the text.

The correct frame is signal, never verdict. Let the system clear the obvious work and surface the uncertain middle. Keep consequential decisions with accountable people. Record when they disagree with the model and why; those disagreements are not annoying exceptions. They are the material from which a better system is built.

A SAFE FIRST EXPERIMENT

Choose one intake queue, not an entire value chain. Define five questions whose answers already change routing. Establish confidence bands, including an explicit “I do not know.” Measure missed severity, inappropriate referrals, human overrides and time saved. Only increase autonomy after the error patterns are understood.

The opportunity is not to remove the underwriter or adjuster. It is to stop spending their judgment on sorting the pile.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7507423451138113536/

========================================

FIELD NOTE • SEPTEMBER 14, 2026

Embrace, Use, Extend, Support and Amplify

I keep returning to five verbs for useful AI.

Embrace. Accept that the capability is real enough to affect plans. This does not mean accepting every claim or abandoning skepticism. It means learning through contact with the tools rather than commentary from a distance.

Use. Put AI into bounded, real work. A demonstration teaches very little about permissions, data, handoffs, reliability or whether anybody wants the result.

Extend. Apply it where it increases the reach of human expertise. Let one underwriter examine more submissions, one analyst explore more alternatives or one small team preserve knowledge that would otherwise disappear.

Support. Build the training, evaluation, access controls and operating practices around the capability. Dropping a tool into an organization is not implementation.

Amplify. Share what works. Turn experiments into methods, reusable assets and better decisions for other people.

The order matters. Amplifying before using produces hype. Using without support produces risk. Extending without understanding can magnify weak work.

TRY THE FIVE-VERB REVIEW

Take one AI initiative and ask what evidence exists for each verb. If the organization has embraced and purchased but not used, narrow the scope and put it into real work. If it has used but not supported, slow down and build the controls. If it has learned but not amplified, document the method.

Useful AI is not one act of adoption. It is a cycle of capability, practice and shared learning.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7505238924110303232/

========================================

ESSAY • SEPTEMBER 10, 2026

AI Can Save Time Without Improving Decisions

Saving an underwriter ten minutes is not the same as helping an underwriter make a better decision.

The distinction is easy to lose because time is simple to measure. A faster summary, a pre-filled referral or a shorter search all produce visible productivity. Decision quality takes longer to reveal. It appears later in loss experience, customer outcomes, broker trust, missed exposures and the quality of the next generation of underwriters.

The concern I hear underneath the AI discussion is not mainly job loss. It is the loss of senior judgment without a reliable way to pass it on. Insurance craft is full of weak signals, exceptions and context that never made it into the manual. A thirty-year underwriter knows when the facts are individually acceptable but collectively wrong. If AI removes the junior person's exposure to those moments, the organization may save time today while weakening judgment tomorrow.

This is why an efficiency-only business case is incomplete. The system should not merely tell us what it decided. It should create a learning surface: the evidence it used, its uncertainty, the rule it may have applied and the reasons a person accepted or overrode it.

WHAT I WOULD MEASURE

Add a judgment ledger beside the productivity dashboard. Track:

- where experienced staff disagree with the system;

- which weak signals changed a decision;

- whether junior staff can explain the final choice;

- how often faster handling produces rework or later correction;

- whether outcomes improve, not only cycle time.

Use AI to make expert judgment more visible and transferable. If the system saves time but teaches nobody and improves nothing, it is automation without progress.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7503851757316231168/

========================================

FIELD NOTE • SEPTEMBER 5, 2026

AI Memory Should Support Learning and Unlearning

We have spent several years teaching AI systems how we want them to behave. We add a rule when something goes wrong, an example when the model misunderstands us, a preference when the output feels off and a warning when somebody finds a new edge case.

Eventually the memory file starts to resemble an old organization: every rule once had a reason, nobody remembers all the reasons, and removing anything feels dangerous.

But models change. An instruction written to compensate for last year's weakness can interfere with this year's capability. Two reasonable rules can become contradictory. More context begins to create less clarity. This is instruction debt.

Starting from nothing is a useful test but a harsh operating policy. I would rather have a switch. Keep the current memory versioned, run the same tasks with a clean configuration and compare the results. Let the evidence decide what the system still needs.

KEEP, REMOVE, RELEARN

- Keep instructions that express durable business rules, authority and safety boundaries.

- Remove instructions that merely compensate for an obsolete model behaviour.

- Relearn preferences from current examples where the work or organization has changed.

Schedule a clean-room comparison every quarter for important agents. A memory system should not only remember. It should support deliberate forgetting, because unlearning is part of learning.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7501998099431575553/

========================================

FIELD NOTE • AUGUST 31, 2026

Is Your Brand Recommended by AI?

For years, a company could ask whether it ranked on Google. The emerging question is different: when a customer asks an AI to explain the market, compare approaches or recommend a shortlist, does the company appear at all?

This is not simply SEO with a new acronym. AI answers are assembled from relationships among claims, sources, products, people and external evidence. A brand can have a beautiful website and still be absent because nobody can tell, in plain language, what it knows, what it does and why another credible source would refer to it.

Trying to game the answer will create the same race to noise we saw in search. The better strategy is to become easier to understand and easier to verify.

RUN A TWENTY-QUESTION VISIBILITY TEST

Write down the questions a serious buyer asks before they know your company name. Test them across several AI services and record:

- which organizations appear;

- which evidence is cited;

- whether your category is described accurately;

- where your company is absent or misunderstood;

- what proof is missing from the public record.

Then improve the evidence, not merely the keywords. Publish clear practice descriptions, useful essays, named expertise, case evidence and consistent facts. Earn third-party references. Make the relationship between the problem and your work explicit.

The objective is not to be mentioned everywhere. It is to be present in the right conversation for the right reason.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7500336040264687616/

========================================

FIELD NOTE • JUNE 22, 2026

Hybrid Work Is Role-Dependent System Design

Five days in the office is a pattern, not a law of management. Five days remote is also a pattern.

The useful question is not which side has won. It is what the work requires. Some tasks benefit from uninterrupted concentration. Some need rapid collaboration. Some require physical equipment or customer presence. Apprenticeship and tacit knowledge often improve when people share space, but commuting to sit on separate video calls creates very little of either.

Blanket mandates frequently compensate for a weak operating model. Leaders cannot see outcomes clearly, so they measure presence. Teams have not designed how decisions are made or knowledge is shared, so the office becomes the default answer.

DESIGN THE WEEK FROM THE WORK BACKWARD

Map the role across four modes:

- concentration;

- collaboration;

- apprenticeship;

- physical or customer presence.

Then set a cadence that gives each mode a reason and a place. Make shared days genuinely shared. Protect concentration time. Pair junior and experienced people deliberately. Measure delivery, learning, retention and customer outcomes rather than badge swipes.

Hybrid work makes sense when it is designed. Without design, both office and remote work can become lonely forms of attendance.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7474945127711670272/

========================================

FIELD NOTE • JUNE 18, 2026

AI Governance Must Become an Operating System

A policy tells people what the organization believes. An operating system determines what actually happens.

AI governance fails when it remains a collection of worthy nouns—fairness, accountability, transparency—with no corresponding controls. The difficult work is translating each principle into ownership, permissions, tests, evidence and a response when the system behaves badly.

I would start with six questions:

1. Where is AI already being used, including inside purchased software?

2. Who owns the outcome of each system?

3. What data does it use, and is that data current, permitted and traceable?

4. What was tested before release, and what is watched afterward?

5. When must uncertainty or consequence trigger a human decision?

6. Can the organization reconstruct what happened later?

If those questions cannot be answered, another principles document will not help.

TURN ONE PRINCIPLE INTO MACHINERY

Choose a principle such as accountability. Define the named owner, approval boundary, logs, override, incident route and review cadence that make accountability real for one live system. Test the path with a failure scenario. Then repeat for the next material system.

Governance becomes credible when it can survive an ordinary Tuesday: a vendor update, a strange customer case, a staff departure or a model that becomes less reliable without announcing it.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7473368244800892928/

========================================

ESSAY • JUNE 16, 2026

The Structure Around the Model Is the Product

An expensive frontier model can lose to a cheaper model wrapped in a better system.

That is not an argument for buying the cheapest model. It is an argument for noticing where dependable performance actually comes from. In serious work, the model is one component. The surrounding structure tells it what to observe, how to verify a fact, what evidence to preserve, when to stop and what a person must review.

A prompt can produce an impressive result once. An operating method must produce a reviewable result repeatedly. This is especially important in banking, insurance and legacy software, where an answer that is plausible but wrong can travel surprisingly far before somebody discovers it.

Model capability will continue to improve and become less expensive. The scarce assets will be good problem decomposition, authoritative context, evaluations, permissions, failure handling and people who recognize when the system has crossed from useful into unsafe.

REMOVE THE MODEL NAME FROM THE DIAGRAM

Take one proposed AI workflow and redraw it without the vendor or model name. Show the work instead:

1. observe the live system;

2. collect authoritative evidence;

3. verify each important fact;

4. preserve what has been confirmed;

5. generate or recommend;

6. obtain human review where consequence or uncertainty requires it;

7. record the outcome.

Now test more than one model inside that structure. If the entire design collapses when the model changes, the organization has bought a dependency, not built a capability.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7472648056467812352/

========================================

FIELD NOTE • JUNE 16, 2026

Make Work Ready for Humans

The optimistic case for AI in commercial insurance is not “remove the underwriter.” It is “stop handing the underwriter unfinished work.”

A submission can arrive with documents in five formats, missing fields, inconsistent names and an exposure buried on page eighty-three. Before anybody applies underwriting judgment, somebody has to assemble the case. That mechanical work consumes the same scarce capacity as the decision itself.

An agent can read, classify, cross-check, request missing material and prepare a referral against appetite. The resulting package should be ready for a human to agree, disagree or ask a better question. This matters: the human is not decorating an automated decision. The system is preparing evidence for an accountable one.

DEFINE “DECISION-READY” BEFORE AUTOMATING

Ask experienced underwriters what must be present for a good decision and what warning signs cause them to slow down. Turn that into a visible acceptance standard. Pilot the automation on one class of business and compare the resulting files with the best human-prepared files.

Measure completeness, hidden-exposure detection, rework, quote speed and underwriter confidence. Do not call a file ready because the system finished processing it.

The aim is larger, better books of business with expert attention applied where it changes the outcome.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7472715332508442624/

========================================

FIELD NOTE • JUNE 16, 2026

Applied AI Is Becoming Visible in Insurer Results

Insurance AI is moving out of the innovation presentation and into operating results.

That shift changes the conversation. Once a public company connects AI investment to growth, service, underwriting or expense performance, competitors cannot treat adoption as a private experiment. They may keep individual methods confidential, but the results eventually appear in volume, speed, margins, staffing and customer behaviour.

This can produce healthy pressure and bad behaviour. Healthy pressure moves investment toward useful work. Bad pressure encourages inflated claims, broad platform purchases and pilots designed to impress rather than improve the business.

The strongest use cases are also broadening. Coding assistance mattered first because it was measurable and close to technical teams. The harder value now sits in pricing, submissions, underwriting, claims, distribution and performance analysis. Those areas require domain ownership, not an AI team arriving with a general-purpose tool.

PUT THE PORTFOLIO ON A CFO-READABLE PAGE

For each use case, name the operating metric it should move: submission-to-quote time, hit ratio, retention, leakage, claim cycle time, expense or incremental premium. Show the investment, responsible business owner, baseline, result and evidence quality. Give experiments a decision date.

If nobody outside the AI team can see the value, it may still be research—and that is fine. Just do not report research as transformation.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7472684644769918976/

========================================

FIELD NOTE • JUNE 16, 2026

Judgment, Empathy and Critical Thinking Remain Human Advantages

It is tempting to make a reassuring list of things people will always do better than AI. I would be careful with “always.” Models will improve, and they can already produce language that sounds patient, thoughtful and empathetic.

Sounding empathetic is not the same as carrying responsibility for another person. Judgment is more than selecting an answer. It includes recognizing what is missing, understanding consequence, explaining a choice and remaining present when the choice turns out to be wrong.

That is the human advantage worth designing around: not a mystical quality that technology can never touch, but accountable judgment exercised in relationships and context.

PUT PEOPLE WHERE CONSEQUENCE MEETS AMBIGUITY

Identify the moments in a workflow where facts are incomplete, values conflict, the customer is vulnerable or the decision is difficult to reverse. Those are not simply expensive steps to automate. They are places where a named person should own the interpretation and conversation.

Use AI to assemble evidence, expose alternatives and test reasoning. Train people to challenge the system and explain their own decisions. Measure whether customers and colleagues understand what happened.

The objective is not to prove that humans are superior. It is to combine machine capability with the forms of responsibility only an organization and its people can accept.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7472647448382070784/

========================================

ESSAY • JUNE 12, 2026

Supervision Without Rules

Bank examiners asking about a specific model marks a change.

For several years, institutions could answer AI questions with principles: fairness, transparency, accountability and a policy approved by a committee. Those principles still matter, but supervisors are moving closer to the machinery. Which model is being used? What can it access? Where does the data go? What happens when the vendor changes the service? Who can stop it?

This creates supervision before the rules feel complete. It is uncomfortable, but not irrational. Models are entering legacy systems faster than legislation can describe every implementation. Waiting for a perfect rulebook is not a strategy for either the institution or the regulator.

Cross-border organizations face an additional problem. The binding standard may come from the jurisdiction with the clearest or strictest expectations, not the one with the largest operation. A global institution cannot realistically maintain one inventory for Canada, another for the United States and a third for the technology team. The operational facts should be the same even where the legal treatment differs.

WHAT A BOARD SHOULD ASK FOR

Every material model or agent needs a short passport:

- purpose and accountable owner;

- data and systems it can reach;

- decisions it may recommend or execute;

- evaluation results and known failure modes;

- vendor and model dependencies;

- logs, escalation and human override;

- tested fallback if the model changes or disappears.

This is not paperwork for its own sake. It is the minimum information needed to govern a changing system. Policies express intent. A model passport, access controls, evaluations and evidence show whether the intent survived contact with operations.

Original LinkedIn note: https://www.linkedin.com/feed/update/urn:li:activity:7471212791270617089/

========================================

FIELD NOTE • MARCH 24, 2025

From Excel Model to Governed API

An insurance spreadsheet is often dismissed as something waiting to be replaced. In reality, it may be the most inspectable version of the business logic the company has.

Turning that model into an API can make pricing, eligibility or explanation available inside digital journeys and AI conversations. It can reduce re-keying and allow the same approved logic to operate consistently across channels. The spreadsheet becomes a kind of digital twin of the calculation that business experts already know.

But an API can scale a mistake as efficiently as it scales a good rule. Hidden links, manual overrides, stale assumptions and undocumented cells do not become governed simply because they are online.

MODERNIZE IN THIS ORDER

First name the business owner. Then define inputs, outputs, assumptions and permitted overrides. Create reference cases whose answers are already trusted. Version the workbook and API together. Reconcile results on every change. Log which version produced each consequential output.

Only then connect the logic to an AI interface. Let the AI collect information and explain the result; keep the approved calculation deterministic where determinism matters.

The useful transformation is not Excel to AI. It is understandable business logic becoming reusable, testable and governed.

Original LinkedIn note: https://www.linkedin.com/posts/robtyrie_connect-your-excel-model-to-ai-for-real-time-activity-7309928613057810433-mQpn

========================================

ESSAY • NOVEMBER 26, 2024

The Business as a Big Set of Functional Spreadsheets

Imagine having fifteen years of a customer's travel history organized well enough to help them understand risk, negotiate coverage and make a better decision. Now imagine the equivalent across an enterprise.

Much of a business can be understood as a big set of functional spreadsheets. Not literally one workbook, but connected collections of events, assumptions, formulas, exceptions and decisions. Spreadsheets persist because they are close to the people who understand the work. They are also where important business logic becomes fragmented, undocumented and dependent on one person who knows which tab is true.

The fashionable response is to call spreadsheets technical debt and replace them. That misses what they contain. A pricing workbook, claims calculator or capacity model may be an executable record of years of domain knowledge. The problem is not that it is a spreadsheet. The problem is that ownership, provenance, testing and access are unclear.

The consumer version is even more sensitive. Travel, health and financial records can create enormous value and enormous harm. Useful analysis does not erase the boundary between information a company can process and information a person should control.

START WITH THE CONSEQUENTIAL WORKBOOK

Inventory the spreadsheets that influence money, coverage, customers or regulatory reporting. For each one, identify the owner, inputs, formulas, external links, decision produced and evidence that it still works. Separate logic from data. Add tests before converting anything.

The goal is not a spreadsheet bonfire. It is to turn valuable business logic into governed digital objects and APIs while preserving a representation that domain experts can inspect. Modernization should make knowledge more usable without making it less understandable.

Original LinkedIn note: https://www.linkedin.com/posts/robtyrie_newinfo-insurance-ideas-activity-7267216969047646209-pxRn

========================================

ESSAY • JULY 14, 2024

RAG Beyond the Demo

It is easy to make a document chat demo. It is much harder to build a retrieval system that deserves a place in a critical workflow.

The demo finds a paragraph and produces a fluent answer. Production has to answer more difficult questions. Was the correct document available? Was it current? Did retrieval miss an exception buried in an appendix? Did compression remove the qualification that changed the meaning? Was the source authoritative, or merely similar?

There is no single RAG architecture. An enterprise may need different stores and retrieval patterns for policies, underwriting guidance, code, customer records and external research. Contextual compression can be useful because it reduces noise, but every compression is also an editorial act. Something is removed. In regulated or consequential work, the system must know what it is permitted to summarize and what must remain intact.

The central requirement is evidence. A team should be able to show that a particular design reduces error on work that matters, not merely that it looks convincing in a demonstration.

BUILD THE BENCHMARK BEFORE THE CHATBOT

Collect a set of real questions, including ambiguous ones, stale documents, conflicting authorities and questions whose correct answer is “insufficient evidence.” Label the sources a good practitioner would use. Measure retrieval recall separately from answer quality. Track citation accuracy, unsupported claims and appropriate refusal.

Then compare architectures. Use multiple indexes where the information really has different authority or structure. A production retrieval system is not a clever search box. It is an evidence system with a conversational surface.

Original LinkedIn note: https://www.linkedin.com/posts/robtyrie_rag-rag-architectures-have-many-patterns-activity-7218226345934245888-qyqx

The Concerns and Wonders of OpenAI. It's Like Fire

Rob Tyrie, December 9, 2022

Originally published at https://robtyrie.medium.com/the-concerns-and-wonders-of-openai-its-like-fire-19069d86c44e

I Know. I definitely agree with you. OpenAI is amazing. And a Toy. And Power tools. It’s like fire. It’s like a weird brain. It’s an extension of brains. It is rewriting history literally with humans. It will probably change a lot of things on earth and only millions of people are trying for now. Billions of people do not have access to it. But as they say, the future is here now, it’s just not evenly distributed.

OpenAi is wrong a lot, because of its design and its incomplete data sources and it is authoritative by design. That’s a bad mix. It’s worse than “mansplaining".

Let’s be clear. It is not, human, sentient, or caring or intelligent. You can ask to write a love poem to you, based on how you feel, but it does not love you. It will lie to you but not maliciously. It’s a machine. Treat it like that.

OPENAI is a technology. In the hands of bad people and stupid ones — Awful things will happen.

The stuff works. Generation and displayed conversation based on well understood and popular forms and formats.

It's powerful and useful now and will get exponentially more powerful as more sensor data, human data, memory and compute is thrown at it. It's a phenomenon. That not a good or bad thing, it's a powerful thing.

Prediction: the code it generates now will get better very fast. You should have seen the state of the art 2 years ago. It's was beyond crap. Did you know the production of code or html from GPTs was an unintentional side effect? It was the same with all the bad poetry. Unintended.

It is likely more thing will be added to OpenAI and it's competition. Like, adding in autonomous robots of all kinds. That's a very dangerous, powerful thing.

Add in autonomous communications and energy creation. Very powerful, if used kindly and politely for good.

I can see a hundred ways to make money and make things better than I have before in my career from writing code, to systems design, to doing audits, doing formal research and writing analytical reports, even to teaching.

The capabilities will grow fast. Experts will become more expert and mundane, bullshit jobs will go away. New ones like "prompt wrangler" , "prompt programer" and "generated-fact checker" will grow. And so, so many lawyers.

There are already prompt guide books that will be written and there will be as many of those as there were the "how to build a web site from html" books to BTC self help guides and what ever Don Tapscott wrote about. There will be "Prompts for Dummies" books in different disciplines.

There will also be novel synthesis and refinement of alot of things that were never conceived before that will be empirically proven to be safe and better that what existed before. We will find new applications for existing drugs and tech that areas harmful that what we do now. The college essay and book report may be dead, but our kids are amazing given good tools and teaching, oh what things they will make with powerful tools.

Truth be told, humanity can’t be passive and wide-open about this data machine and it’s algorithms. You can’t either.

I hope there will be rules and ethics applied like, if you use it it has to be declared and the human of the business takes responsibility for what is published or claimed.

That’s an easier one. It’s still a struggle until there are new laws and the courts and the jurisdiction thing is fixed.

I am hopeful but it is "critical hope" . I am not highly and stupidly optimistic.

There better be an ombudsman when some steals your IP, and good liable and slander laws if some chooses to wreck your image or reputation.

There better be privacy.

There better be intellectual property rights and clear attribution rights for creators.

There better be access for all to generative transformers. They should be treated like The Commons. They should not be treated like private playgrounds or banks.

The Creators and Creatives have to protected and paid or this infinite monkey machine will break down or be taken down. We have to be paid and respected as creators, executors and as those who take risks and responsibilities as humans. We have to have basic rights and freedoms in the face of all automated uncontrollable algorithms.

Like the people who what to declare lakes and river as human to give inanimate things rights, there will be a faction that want to treat AIs as humans so they can be held accountable by law. This was attractive and we used that approach for corporate entities, but that used to work but does not work any more. These ideas have been explored in a book called Life 3.0 that is worth reading. It's written by physicists and explores the legal concerns of autonomous robots. However, it was written before the GPT era.

Get Rob Tyrie’s stories in your inbox

Join Medium for free to get updates from this writer.

Subscribe

Remember me for faster sign in

The scientists and zota-rich, have a lot of power in this construct. Probably too much power. We have seen how that has played out in past civilization, the UK and is playing out in this pandemic.

One of the big, big problems I see is when machine to machine connections are made and things like banking credit limits are affected or credit scores. Government should worry how energy, education and transportation are automated with this technology.

These will be under the covers and if not controlled and made transparent and auditable and explicable, bad things will happen.

You will lose your job and the factory will shut down and the wrong political party will be empowered to make small sects and tribes miserable.

There is a possibility and a desire to generate stuff with the generator... This has already been shown to lead to awful things, like those digital pictures generated with the First GANS and RNN algorithms that looked like nightmares or that racist chat bot that Microsoft released and had to kill.

This stuff now that it is conversational will likely be worshipped by some and turn into some kind of eerie religion and if that gets powerful, it does not take a average historian to see how that will go. That would be bad.

It will be the voice inside people’s heads. We have already seen people claim this machine is sentient. The "Eliza" and "Turing" effect are real. Some people, especially concrete thinking conservatives will try to treat this type of machine as a person. That’s not good.

Humans anthropomorphize things that talk and happen to look like us. We see this is cartoons, sacred rocks like Indian Head or Mountains like Sleepy G in Thunder Bay. It even happen when people see giants in the clouds. And, by making them converse, people will get confused about these new machines and algorithms.

Infinite smart monkeys can be good or they can wreck a car or plane or factory.

It is, however, inevitable that it will be used. Solves too many problems. Just like CRISPR and the current Internet. The genie is out of the bottle and Pandora's box got opened.

This type of machine should not be connected to any financial or political system. Or any mass entertainment system. But I think that it is to late and it worries me. Movie scripts are already being made.

It will have to be regulated and individual freedoms will have to be protected, until it takes over or some corporate entity or state and then we work for them... Like we do now.

I hope those in power, genuinely like humans, and no one asks for the planet to be a self sustaining eco-park.

One will have to be careful what one prompts for.

It's just a machine made by humans. A tool. And seriously again, it's a tool for experts and ethical people.

But, we have seen it over and over again from gun powder to engines to nuclear power, tech is dangerous in the hands of the greedy and in the hearts and minds of nasty idiots. We may all be at risk of the people and machine too early on the Dunning Kruger Curve.

Prediction: People and companies will make billions of dollars from OpenAI and/or it's close competors and Quantum computing will be part of it, because they handle probability faster than any other tech in history so far. It's likely because Google is the richest company in the world, based on Search, Organized Data and, ugh, ads.

Who owns OpenAI and and how do the benefit? Who is Sam Altman?

Why is he smarter than Elon?

Why has Microsoft written checks for Billions of dollars to OpenAI. What does it hope to gain other than profit?

I know Elon is fighting with OpenAI. Why? And why isn't Zuckerberg saying anything about it? We have lots to talk about. I have more questions.

Well. In a nutshell. I am worried. I am worried for my kids. We all should be. And hopefully people more powerful than us will feel that way and we can help convince them to be good and humane.

I hope things go well.

Or we are in "what hath G*d wrought" territory.

On a dark and stormy night...