AI (Artificial Intelligence) is all the current rage. It is a large subject from its technical aspects to the many applications it is finding. It is already disrupting many big tech companies and their expensive software engineering staffs. But it is likely to disrupt many other industries and companies over the coming years.

Having myself applied this tool in my workplace, it is amazing in the way it can change a technical workflow, and for some tasks a 10x improvement in productivity is no lie. Watching the better AI models build, deploy, test, debug, and do it again in a cycle by themselves until they reach a solution — and accomplish in ten minutes what would have taken hours — is remarkable. It is probably software engineers who see and feel this most intensely right now, as it is their jobs which are directly in the line of sight.

However, as with any new powerful tool there is a learning curve in how to apply it. The risks of over-enthusiastic application and lack of careful governance can already be seen in the recent experience at Amazon and others. But as with any new technology it marches forward, and as with all things in this world: for better and for worse.

I keep some of the noteworthy news on X as @1A_Lutheran. I also apply the same tools directly to my scholarship: the words and arguments of Early Christianity and Christendom, the Lutheran Reformation, and nineteenth-century American Lutheranism. See, for example, the AI translation demo on “About the Name Lutheran.”

What the term covers

Artificial intelligence is not one product. It is the effort to make computer systems do work that used to require human judgment: learning from examples, recognizing patterns, handling language, and choosing among actions. Most of what the public now calls “AI” is still narrow: good at a bounded task, not a mind.

Three layers sit under the current boom.

Machine learning is the older and still essential core: programs that improve their performance from data rather than from a fully written-out rule book.

Deep learning is machine learning with large neural networks. Image recognition, speech systems, and today’s language models all sit here.

Generative models produce new text, code, images, or audio. They do not retrieve a stored answer so much as continue a pattern they have seen at enormous scale. That is why they can write a lucid paragraph and also invent a citation.

Talk of artificial general intelligence (AGI) — a system that matches or exceeds a competent human across many domains — is now ordinary in vendor marketing. It is not a solved engineering fact. Treat claims of “arrival” as claims, and test the tool on work you can check.

Large language models

A large language model (LLM) is a generative system trained on a vast corpus of text (and, increasingly, code, images, and other traces of work). The dominant architecture is the transformer (2017). In use, the model predicts the next token — a word-piece — over and over. From that simple loop you get summarization, translation, dialogue, and a surprising amount of programming.

They do not “understand” in the human sense. They are extremely good at continuing the kinds of language they were trained on. Two practical consequences follow.

First, hallucination: fluent falsehood. For scholarship this is not a minor annoyance. A model will supply a plausible-looking footnote. You still have to open the book.

Second, tool use. Since 2025 the important change has not only been bigger chat windows. It is agents: models that call a compiler, a browser, a search index, or a file tree, look at the result, and try again. That is the loop behind the 10x software workflow described above. It is also the loop I use, with supervision, on translation and editorial work.

Where things stand (September 2026)

The March 2026 snapshot on this page is out of date. Two different “markets” are now easy to confuse.

Consumer chatbots. Among U.S. web and mobile chatbot users, ChatGPT is still first, but no longer in a class by itself. First Page Sage’s September 2026 estimate (monthly active users) puts ChatGPT at about 51.5%, Google Gemini at 27.6%, Claude at 10.2%, Grok at 2.8%, and Perplexity at 2.0%, with Microsoft Copilot a smaller standalone slice. ChatGPT’s share has fallen through 2026 (it was still in the mid-60s in January) even while its user count grew. People are stacking assistants rather than marrying one. Gemini posted the largest quarterly user growth in that survey; Claude was next.

Those percentages are consumer reach, not “who has the best model.” ChatGPT’s default consumer stack in that report was still the GPT-5.5 family; the GPT-5.6 line (Sol / Terra / Luna) is the current OpenAI flagship on the API and in Codex. Anthropic’s generally available flagship is Claude Opus 4.8, with Sonnet and Haiku as the workhorses and Fable as a restricted/high-end variant. Google’s Gemini 3.x family is the other mass consumer and Workspace path. xAI’s Grok 4.6 is a real frontier model in coding and reliability tests; it is not yet a mass chatbot in the ChatGPT/Gemini sense.

Enterprise spend and coding. If you measure what companies pay for, the picture flips. Menlo Ventures’ 2025 enterprise survey already had Anthropic first in LLM API spend (about 40%), OpenAI down to about 27%, Google about 21%. Ramp’s September 2026 business index still has Anthropic slightly ahead of OpenAI among paying U.S. firms, with growth in cheaper standard models (Sonnet-class, GPT-5.6 Terra, and the like) as finance departments push people off the most expensive frontier SKUs. Token prices have been falling. That is the ordinary fate of a boom: the demo is the giant model; the bill is the mid-size one.

Coding is the clearest industrial foothold. Claude Code, OpenAI’s Codex, Cursor, and similar agent harnesses are how a lot of software is now written. Anthropic spent 2025–26 owning much of that niche; OpenAI, Google, and xAI are all in the same race. NVIDIA remains the shovel-seller: most training and a large share of inference still run on its GPUs, whatever logo is on the chatbot.

Open weights. Meta’s Llama is still widely deployed but has not set the frontier this year. Chinese open models (DeepSeek V4 and others) matter more for cost and for developers who will not or cannot send work to a U.S. API. Enterprise buyers remain cautious; the “open source will eat the API companies” story has not happened on the timetable that was advertised in 2025.

None of these rankings should be treated as theology. A coding benchmark, a chatbot-user survey, and a corporate API invoice will name three different winners. The useful question for this site is narrower: which model, under supervision, helps with a German periodical, a Latin fable, or a long argument in Walther — and which one invents the footnote.

How I am using it here

The tools are already in the workshop. They draft, they collate, they propose a translation, they typeset. They do not replace reading the German, checking the edition, or owning the argument. That division of labor is the whole point of @1A_Lutheran as I have practiced it: news and experiment on one side, the old books on the other, and no confusion about which is which.

Further reading on this site: AI Translation Demo — “About the Name Lutheran.”

Sources (September 2026)

Definitions of machine learning, deep learning, and transformers are standard; see any current textbook treatment or the vendor primers from IBM, Google Cloud, and AWS. Older March 2026 “top five vendors” copy on this page (OpenAI ~60% chatbot share, etc.) is withdrawn.