AI Update mid 2026
Several months ago I wrote a piece about “Artificial Intelligence (AI)”. (I will reiterate that AI is neither “artificial” nor “intelligent”.)
AI has been around since at least the 1970’s. (I took a course in graduate school entitled AI in 1977). It has burst into wide public consciousness since chatbots were introduced in 2022.
Reactions to chatbots have ranged from doomsday scenarios, to derision at examples of bad chatbot answers.
The chatbots are apps that query a Large Language Model (LLM). LLMs had two flaws:
- They frequently gave incorrect answers, and worse would “hallucinate” when it did not know the answer.
- LLMs were very useful, but you needed to know what to use them for, how to phrase your questions, and verify from other sources whether the LLM was hallucinating.
We can think of an LLM as like a person who has a photographic memory, with all the written material on the Internet in its memory. The LLM can take a query and retrieve from its “memory” what the Internet states about your question. There is no original or intelligent thought. It only knows what humans have previously produced about the topic.
In the past three years, there have been significant breakthroughs in AI due to
the introduction of what are called “Harnesses” and “Agents”.
A Harness controls an LLM. It provides the User Interface (UI), and is able to “plan” and retry queries to the LLM, among other functions. Importantly, it decides what data (files, documents, instructions) to include it is interaction with the LLM. A Harness will often perform multiple calls to an LLM.
Examples of Harness:
- airline APIs (Application Programming Interface)
- email tool
- authentication (e.g. login)
An Agent relies on the LLM and Harness to perform tasks.
Example of agent:
- searches airline flights
- compares prices
- composes summary
- sends email
Going up the hierarchy further, Anthropic has introduced a tool called Claude, which uses Agents. It is particularly good at assisting computer programmers.
However, there is a significant amount of learning required to utilize Claude for programming to its fullest. It will not turn a novice into an overnight expert programmer.
Claude has shined at some other high profile tasks. It was used to plan a route for NASA's Mars rover, Perseverance. Norway's sovereign wealth fund is using Claude to screen its portfolio for ESG (environmental, social and governance) risks. Recently, Claude found over 100 bugs in the Mozilla Firefox web browser, of which 14 were considered high severity.
In the last four years AI has clearly advanced from what some have called “parlor tricks“, to being genuinely useful in diverse ways.
It is difficult to correctly predict how new technologies will be utilized, and how they will impact society. (Ironically, AI is terrible at predicting the future.)
We have to simultaneously entertain contradictory views about AI.
Discount the hype coming from those with a financial interest in AI, take seriously knowledgeable people who cite the dangers, and accept that AI is likely to make people in a lot of jobs more productive, and yes eliminate a lot of jobs.
On the whole, is AI going to be good for society?
Over the last two generations, technology has eliminated millions of farm jobs. Most people would agree this has been good for society (though not good for a lot of farm laborers).
My prediction: adoption of AI will be a mixed blessing. Consider automobiles: tremendous improvement in transportation, millions of people killed.
We need guard rails around AI. I don’t trust those with a financial interest, and the current federal government is not up to the task of providing guard rails around AI.