Welcome to Verses Over Variables, a newsletter exploring the world of artificial/ intelligence (AI) and its influence on our society, culture, and perception of reality.

AI creates abundance in communication. I have been saying this for a while now, in briefings, workshops, and client calls, in slightly different terms depending on who was in the room. This month, the concept really hit home for me when GPT-6 Astra launched, and the world suddenly found Blender, the 3D modeling tool.

The demos made the argument visible:  the people using Astra already had something they wanted to make. What changed was their access to the skills and tools required to express it.

Expertise and ideas come first, but their form has always depended on the skills, tools, and time available. When an idea needed a dashboard and I couldn't build one, I had a few choices. Learn the software. Hire someone who already knew it. Or turn the idea into the slide I knew how to make.

AI changes that. It puts hard skills I never trained in and tools I never learned to operate within reach. I pair those new capabilities with the expertise and judgment I already have. I can reach for code, data visualization, animation, video, or software I never expected to use when one of those capabilities serves the work.

That is what I mean when I say AI creates abundance.

The same shift applies beyond communication. A tool that once required hours of tutorials, a specialist vocabulary, and a practiced hand is no longer automatically out of reach. AI can help operate the interface, translate an intention into the tool's language, and work through the unfamiliar steps with you. However, expertise still determines whether the result is useful.

A box of crayons with 64 colors does not make someone an artist. It can produce a fairly spectacular mess. The cost of making something fell, but the quality standard stayed the same. More capability creates another responsibility: choosing what is worth making.

For communication, start with the person receiving it and what they need to do: understand a situation, inspect the evidence, learn how something works, compare alternatives, or make a decision. The outcome, call to action should guide the form.

A spreadsheet can become a dashboard when someone needs to explore the numbers behind the conclusion. A process stored one step per row can become a diagram when the relationships and loops are the point. An explanation can become a prototype when trying the mechanism will teach more than another paragraph.

Sometimes the right answer is still the paragraph. Abundance matters because you can choose it after considering what the person needs, rather than letting the familiar format drive the decision.

I am demonstrating this live at AI Unscripted on October 14. If you are in LA, join me to see how these new capabilities can change the way an idea is explained, explored, and used. If you want to look at where the form of your work is still being set by the tool you happen to be fluent in, schedule a call.

Most AI models are built to make something: an answer, an image, a plan, a block of code. Jev is built to make a decision, which is a surprisingly modest job description for an AI model in 2026. Jev is TypeSafe’s first “System One” model, named for the fast, intuitive judgments people make without working through a problem step by step. It doesn’t generate free-form text; it returns a probability, a score, or a choice from a predefined list.

Think of Jev as a smart if-statement. It handles classification and quick decisions inside a larger workflow. It does not reason through a long task or generate the finished work. It only accepts text, so it cannot look at an image, listen to audio, or watch video. Its context is also relatively small: 64,000 tokens per request, with a 32,000-token limit for the state plus the longest question. This model knows its lane.

Another way to think about Jev: Jev makes quant (machine learning) accessible to the people who are used to generative AI. If you think of all the quant traders on Wall Street who use fast technical trading, Jev lets non-technical people approximate that task.

TypeSafe also describes Jev as unable to hallucinate. This is a very specific definition of hallucination. Jev cannot invent an extra category or return an essay where the software expects a probability. It can still choose the wrong category with impeccable formatting. The guarantee covers the answer's format, not its truth.

This narrower job changes the economics. TypeSafe prices Jev at $0.042 per million input tokens, with no output-token charge. It is built to make many small decisions quickly rather than conduct one long conversation. I tested it on thousands of recent AI research papers. I wanted to see which papers could help a nontechnical person use AI for research, writing, analysis, information-finding, and decision-making. Jev judged each paper's usefulness, the topic it addressed, who would benefit, and how strong its evidence appeared. I reviewed the top 150, and a different model wrote plain-English takeaways and helped format the library.

The full Jev run made 73,854 decisions in a few minutes and cost $0.95. Claude estimated that the same workload on Haiku would take more than six hours and cost about $26. Setting up Claude to generate and format the finished library still took longer than Jev took to classify the papers.

Jev becomes interesting when unstructured information needs to end in a limited action. Data classification is the obvious case. I have also seen demos that looked like live digital trading because the model’s entire job was to return buy, sell, or hold, and color-palette generators built around the same kind of bounded choice. The interface can look sophisticated. Underneath, the model has three buttons.

Organizations are full of decisions people know how to make, but software cannot touch because nobody has ever written down the rules. Jev lets a team describe the judgment in plain English, run it against real material, and inspect the mistakes. That is a much cheaper way to find out whether a decision belongs in software rather than building a custom model first.

After years of AI models trying to say everything, Jev's charm is that it knows when to stop talking and pick a button.

I used to joke in my design classes that business people make things difficult to understand because that is how they get paid more. (Just think about all the explanations of the bond market you have been hearing lately with the rise in bond yields.) Developers make things hard to understand for a different reason. They live inside a bubble where they talk mostly to other tech people, and as we’ve discussed, they are really bad at naming things.

With all the AI safety issues cropping up in the news, more AI terms are getting batted around. I was listening to a podcast where the speakers kept misdefining those terms while asking for a glossary, so I decided to make one. The full glossary is online now and includes 240 terms. These six are especially useful for following the current conversation.

The Field + Edge AI Glossary has 240 terms, from AGI to zero trust, with a definition, why it matters, and the related ideas worth knowing next.

What it is: Anthropic launched Claude Projects in Claude Code, and if this sounds like we’ve already discussed this, it is because this is the fourth thing that Anthropic has created called Projects. Projects already exist in chat, Cowork, and as a command in Claude Code. This new version of Projects is a single long-running conversation where Claude works as a coordinator. You hand it tasks, and it opens a separate cloud session, or thread, for each piece of work. Every thread starts with the project's repositories, instructions, and memory, works on its own branch, opens a pull request, and reports back. An Overview pane tracks which threads are working, which are waiting on you, and which pull requests are ready.

How I use it: I don't have access yet. The rollout starts with accounts that have used cloud sessions and don't already have Projects in Claude.ai chat or Cowork, which rules me out for now. In Claude Code I already organize by folder, and each folder becomes its own small project holding the context and chats for one client or one topic. This new UI puts a coordinator above that layer, one that decides what becomes a thread, hands the work out, and shows the orchestration as it happens. The catch is that threads are cloud sessions. My folders live on my machine, so anything I want a thread to work on has to go somewhere it can reach.

What it is: Anthropic redesigned the Claude interface, folding Cowork back into Claude after eight months as a separate place to work. Chat, Cowork, and Design now share one interface, and Claude decides whether a request is a quick answer or a longer job with tools and connectors. Anthropic says people used both and found the frustrating part was deciding where a task belonged. Three artifact types arrive with the merge, all in beta. Docs writes living documents inside the conversation, asks its clarifying questions before it starts, leaves comments explaining its choices, and exports to Word or Google Docs. Slides builds a deck from notes, a report, or whatever is already in the conversation, then lets you edit any slide, present without leaving Claude, and download PowerPoint or PDF. Design runs on Canva's Design Engine, so it reads a brand kit and returns something still editable, exporting to PPTX, PDF, HTML, or straight into Canva.

How I use it: I’ve been living in Claude Code and have been making design and MSFT artifacts for quite some time, so the update didn’t change my workflow much. My decks come out of a custom skill that holds my palette and templates, and I built that skill because asking Claude to build a deck produced sloppy ones. Sitting one step closer to the conversation helps. What would move the work is a generator that takes direction from a design system. Docs is a drafting surface for now, since version history is still missing and charts and diagrams stay wherever you put them.

The call came from inside the house: Three years into the New York Times lawsuit against Microsoft and OpenAI, the most damaging material in the case turns out to be Microsoft's own. A January 2023 internal memo from Brent Hecht, the company's director of applied science, called training on unlicensed content "an astonishing theft of unprecedented proportions" and possibly "the largest theft of labor in human history." A year later he measured it. In a January 2024 internal presentation, Hecht showed that Microsoft's Copilot answer engine cut click-through rates to the Times' domain by as much as 93 percent compared with traditional Bing search, and called the pattern a "doom loop." The same brief quotes Nick Turley, who runs ChatGPT, describing publishers as facing an "existential threat" from products that are "largely substitutive" and grow more so as they improve. Satya Nadella testified that anything paywalled should be licensed by whoever wants to use it for training or grounding. Microsoft says Hecht spoke for himself, was not a decision-maker, and was hired to present divergent perspectives. The exhibits behind the brief remain sealed. Judge Sidney Stein is weighing summary judgment, with a decision on whether the case reaches trial expected in 2027.

Show your work: Twenty-five Fields Medal winners, the top prize in mathematics, signed a public declaration on September 11 objecting to how AI companies are using their field as a proving ground. The signatures span award classes from 1978 through this year. Their objection is about the record, not the answers. When a famous problem falls, the proof gets written up, the new methods get separated out and named, earlier work gets cited, and the rest of the field spends years absorbing it. AI solutions arrive "announced in a rush," the declaration says, skipping all of that. "As in all creative professions, this raises severe attribution and plagiarism questions." Three days earlier, OpenAI said it had cracked one of the seven Millennium Prize problems, each carrying a $1 million award, using roughly 10,000 AI agents working at once at a cost the company put in the millions of dollars. NYU mathematician Tristan Buckmaster had spent months on that same obscure line of research with an Anthropic colleague, working through unpublished drafts inside a private OpenAI tool. He asked whether OpenAI's model had benefited from that material and called the match a "bright red flag." OpenAI says nobody searched user data, and separately says it cannot rule out that his input improved the model. Mathematics has a citation standard, a prize committee, and a million dollars at stake, and it still cannot say who found what. Mathematics is objecting to a pace the creative industries have worked at for years. It also keeps better records.

— Lauren Eve Cantor

thanks for reading!

I also provide AI Audits and Workshops. Please feel free to reach out if you’d like to arrange one for you or your team.

if someone sent this to you or you haven’t done so yet, please sign up so you never miss an issue.