Three unrelated stories this week land on the same point. A dermatologist pitted a ten-year-old, purpose-built image classifier against two of this year's best-known chatbots on skin photos, and the old specialized tool won by a wide margin. Mistral released a model that does nothing but forecast numbers, instead of asking a general chatbot to guess at them. And two of the free programs companies use to run AI models on their own machines quietly added a cheaper, narrower kind of model that answers with a flat choice instead of a paragraph. None of these are our work. All of them argue for it.
Why we think this matters
The industry spent two years telling itself that one big, general model would eventually do everything better than a small one built for a single job. This week kept chipping at that story, from several directions that do not know about each other.
A general chatbot is trained to be plausible across almost anything you might ask it. A narrow model is trained, or simply built, for one job and nothing else. When the job is well defined — read this kind of sensor data, read this kind of photograph, answer this kind of yes-or-no question — the narrow tool keeps winning, and keeps winning by more than a rounding error.
A general chatbot has to be plausible about almost anything. A narrow tool only has to be right about one thing. This week, three separate groups of people found the same gap between those two jobs.
That is the whole argument underneath VEDA reading a machine's own fault history instead of asking a general model to guess at it, and underneath Nirdesh answering from an OEM's own manuals instead of the open internet. It is not a new argument for us. It is reassuring to see it turn up, independently, in a plant-forecasting model, a dermatologist's spreadsheet, and the release notes of two pieces of software millions of people already run.
What changed this week
- Anthropic handed the format for teaching an AI agent a specific job to an outside foundation, so no single company controls it. "Agent Skills" is a simple way to package step-by-step know-how — a company's own way of handling a purchase order, say — so that any compatible AI agent can follow it, the way a recipe card works in any kitchen. Anthropic invented the format but has now given it, along with its name and website, to the Agentic AI Foundation, the nonprofit group under the Linux Foundation that already looks after the standard that lets AI tools talk to other software. More than 70 agent products already use the format, and roughly a million skills have been published. Confirmed by Anthropic's own announcement, 22 October 2026 (anthropic.com). What it means for us: a skill we write once should keep working across whichever agent tool a customer already has, rather than being tied to one company's product — the opposite of lock-in, which is the property we care about most.
- Mistral built a model whose only job is forecasting numbers, meant to run on a plant's own computer with no graphics card required. Timestral reads past readings — a sensor, a meter, a production count — and predicts what comes next, showing a range of how confident it is rather than one flat number. It can run fully offline on an ordinary office server. Mistral's claims about its accuracy are the maker's own and have not been checked by anyone outside the company, though the model itself is public so anyone can test it. Confirmed by Mistral's own announcement, 15 October 2026 (mistral.ai). What it means for us: it is one more sign that "AI for predictive maintenance" does not require a chatbot at all — the same case we make for building a purpose-built tool rather than asking a general model to guess.
- Mistral and NVIDIA are building open tests for whether an AI agent can actually do real engineering work, not just describe it. The two companies plan shared benchmarks and open datasets for AI used in simulation-heavy engineering and chip design, aimed at aerospace, automotive and semiconductor teams — work that overlaps with what VCAD and Nirdesh already do. Confirmed by Mistral's own announcement, 21 October 2026 (mistral.ai); the benchmarks themselves do not exist yet, so there is nothing to measure against today. What it means for us: well-funded outsiders are starting to take OEM engineering AI seriously enough to want a shared yardstick for it, which is a market validating itself faster than it is being contested.
New models this week
OpenAI's chatbot models and its coding tool, Codex, will soon be sold through Google's cloud marketplace as well as Microsoft's, a first for a company that has sold exclusively through Microsoft until now; no pricing has been published and nothing changes for us today. Confirmed by OpenAI's own announcement, 21 October 2026. Mistral released Voxtral 3, a new family of speech models — turning speech into text, text into speech, and a kit for building voice assistants — adding support for 50 languages, including several Indian ones, and able to run on a company's own servers rather than only in the cloud. Confirmed by Mistral's own announcement, 20 October 2026. OpenAI told developers it will remove four older models, including GPT-4o and GPT-4.1, from its programming interface on 14 January 2027; confirmed by OpenAI's own documentation, dated 14 October 2026 — routine retirement, worth knowing if anyone has code calling those models by name. And Ollama, the free program many teams use to run AI models on their own computer, spent the back half of October adding a wave of small "decision" models — AI that answers with a flat choice or a score instead of a paragraph, the same pattern behind the Jev tool we explained here in September — alongside support for new vision models from NVIDIA. Confirmed by Ollama's own release notes, 12–24 October 2026. Nothing here changes what we ship this week; worth knowing this pattern has gone from one startup's product to a built-in feature of software people already run locally.
What people are actually trying this week
A dermatologist at a London hospital ran his own test: he gave two of this year's most capable chatbots and a ten-year-old, purpose-built image classifier 200 real skin-lesion photos, all confirmed by biopsy, and asked each to pick a diagnosis. The old specialized tool got 62% right; each chatbot got about 42%, and the specialized tool also caught more of the actual melanomas. It is one person's own test, not reviewed by other scientists, and run on a small set of cases — but he published the photos, the exact questions he asked, and every answer, so anyone can check his work rather than take his word for it. Published 25 October 2026 on the author's own site (lynchlab.net). Worth reading, not yet something to repeat to a customer as settled fact — but it is independent evidence for an argument this newsletter keeps making. Separately, a well-funded new AI lab called Recursive released a free coding model you can download and run on your own laptop, and says its own automated research tool did most of the engineering work of building it, with people supervising rather than doing the work by hand. Recursive's claims about its speed and quality are the company's own; nobody outside Recursive has tested them yet. Confirmed by Recursive's own announcement, 21 October 2026. Watch this one for what it says about AI labs starting to use AI to build AI — not yet something to bring to a customer. And a widely shared analysis asked why OpenAI's and Anthropic's prices have fallen so sharply this year, and concluded it is ordinary competition: cheaper open models and smaller providers keep forcing the larger labs to cut their own prices rather than raise them toward what a question actually costs to answer. Published 25 October 2026 by independent writer Max Woolf (minimaxir.com). Worth knowing as background — cheaper AI from the cloud giants changes nothing about why a customer would still want their own model running inside their own walls, but it is good to know the direction prices are moving.
Explained simply: Agent Skills
In one line. An Agent Skill is a small folder of instructions — written for a human to read, really — that teaches an AI agent how to do one specific job the way a particular company wants it done.
An everyday picture. Think of a recipe card, not a cookbook: a short, plain set of steps for one dish, that any cook in any kitchen with the right ingredients can follow. An Agent Skill is that recipe card for an AI agent — "here is how we file an expense report," "here is the format our invoices use" — rather than a general cookbook trying to cover every dish anyone might ever want.
Why people care. Before this kind of standard existed, teaching an AI assistant your company's own way of doing something meant writing it into a long, one-off set of instructions that only worked with one product, or re-explaining it every time. A shared recipe-card format means the same skill can be written once and followed by any AI agent that understands the format, the way a USB cable works in any laptop that has a USB port. Anthropic published the idea in December 2025, and more than 70 agent products adopted it within a year.
What is real and what is hype. What is real: the format exists, it is simple — mostly plain text with an optional script or two attached — and it is genuinely used, at a scale of roughly a million published skills. What changed this week is real too: Anthropic no longer owns it. The format now belongs to the Agentic AI Foundation, a neutral nonprofit group, with engineers from Anthropic, Google and OpenAI sharing the job of maintaining it. What is still a claim rather than a guarantee: that "neutral" stays neutral. A standard that several large companies all have a hand in maintaining is less likely to quietly favour one of them than a standard one company owns outright — but it is not a law of nature, only a better set of incentives than before.
What it means for us. Where we already teach an agent one of our customers' specific procedures, writing it as a Skill rather than a one-off prompt means that work keeps its value even if the particular agent tool underneath it changes — the opposite of the lock-in we warn customers about elsewhere. Worth knowing the name the next time a customer's own IT team asks what format we use to teach an agent a new task.
Nothing here changes what we ship this week. The thread worth carrying into customer conversations is the one running through all three main stories: a model built for one job keeps beating a bigger one asked to do that job as a side effect, whether the job is forecasting a sensor reading, reading a skin photo, or answering yes or no. That is the case we already make for VEDA and Nirdesh over a general chatbot — this week, several people who have never heard of Varahi made the same case from their own, unrelated corners of the field. We will keep reading the same sources next week, and if nothing moved, we will say so instead of padding it.
Read about AI Transformation →