Wednesday's Rundown was less geopolitics and more "new tools inside software." I rewrote it the way I would tell a colleague over coffee. Short version: ChatGPT co-creator Diogo Almeida's TypeSafe launched Jev — an AI that only picks between preset options inside apps, with confidence scores, claiming zero hallucinations plus extreme speed and price. Salesforce introduced Koa, its own reasoning model for sales and support agents. A how-to covers OpenRouter so coding agents can use more models. And 30+ Chinese researchers published a roadmap of recursive self-improvement they call "the last AI built by humans." Skipping the You.com, Stack AI, and workshop plugs.
Words worth knowing
| Word | In one line |
|---|---|
| Diogo Almeida / TypeSafe | Ex-OpenAI researcher who helped build methods behind ChatGPT's conversational training; his startup TypeSafe just left stealth with Jev. |
| Jev / System One | TypeSafe's new AI that lives inside software and answers preset questions with confidence scores — not a free-form chat model. |
| Tokens | Chunks of text AI models bill and process. TypeSafe prices Jev at $42 per billion input tokens; output is free. |
| Hallucination | When an AI invents a plausible-sounding false answer. TypeSafe says Jev can't, because it only chooses among options set in advance. |
| Jevons paradox | When something gets cheaper, people use more of it — TypeSafe's nod for why ultra-cheap judgment calls may flood into software. |
| LLM | Large language model — the usual chat-style AI (Claude, ChatGPT, etc.) that generates open-ended text. |
| Salesforce Koa | Salesforce's new in-house reasoning model for Agentforce sales/support agents, built on Nvidia's open Nemotron 3 Super. |
| Synthetic training data | Fake-but-realistic scenarios used to train a model (here: irate callers, deal-closing reps) without using real customer data. |
| CRM | Customer relationship management software — Salesforce's core product for tracking customers, deals, and tickets. |
| AIforce / ClaudeForce | Salesforce tools that push CRM data and permissions into outside AI apps; ClaudeForce (Anthropic/Claude) is now in beta for every customer. |
| OpenRouter | A one-stop API gateway that lets you pick and pay for many AI models from one key — useful for Codex or Claude Code. |
| API key | A secret password string that lets software call a paid AI service. Set a spend limit and expiry when you create one. |
| RSI (recursive self-improvement) | AI that improves itself, then improves the improvement process — up to models that build their successors. |
| World model | An AI that learns how environments behave so it can control things like robots, cars, drones, or game worlds (Odyssey 3). |
| Project Lily | OpenAI program where human contractors read and rate real ChatGPT chats — often without the user's knowledge, per 404 Media. |
| Meta One | Meta's paid tier ($2.99–$499/month) for extra Meta AI usage and creator tools. |
TypeSafe's Jev: judgment calls inside software, not another chatbot
Ex-OpenAI researcher Diogo Almeida's TypeSafe emerged from stealth with Jev — a system that answers preset questions inside software and attaches confidence scores. Almeida helped build the methods that taught AI to talk to people (the research behind ChatGPT). His next model cannot generate free-form text at all — by design.
- Price: $42 per billion input tokens; output is free. TypeSafe estimates that at about 238× cheaper than Claude Fable 5.1.
- Speed: answers in 70–500 milliseconds — roughly 40–200× faster than today's LLMs.
- TypeSafe says Jev "can't hallucinate" because it only chooses between options set in advance. They call it a "frontier-intelligence function call."
- Example uses: sorting requests, scoring records, or screening another AI's outputs for jailbreaks (attempts to bypass safety rules).
Almeida frames it as "more like a database than a coworker." The Jevons-paradox pitch: if it really is this fast, cheap, and reliable, it may become a standard building block inside software rather than a rival to chatbots.
Salesforce trains Koa for its own agents
Salesforce introduced Koa — an in-house reasoning model for its sales and support agents. It is built on Nvidia's open Nemotron 3 Super and tuned for business tasks.
- Training data is fully synthetic: personas like irate support callers and reps closing deals across more than a dozen industries — no customer data.
- On an internal CRM benchmark, Koa made 3× fewer errors than top models and matched or beat them on jobs like updating deals or routing tickets.
- Salesforce hosts Koa itself, so customer requests stay inside its systems instead of going to an outside model provider.
- It also launched AIforce (pushes CRM data and permissions into outside AI tools) and opened ClaudeForce to every customer as a beta.
The Rundown's take: Salesforce is customer, distribution partner, and now competitor to the big AI labs at once. High-volume agent reasoning can stay in-house; ClaudeForce still helps frontier labs reach Salesforce customers.
OpenRouter 101: more models for Codex and Claude Code
The letter's training piece is a guide to wiring OpenRouter into coding agents so they can try cheaper or specialised models.
- Sign in to OpenRouter, browse the catalog, filter for your task, sort by price, and check a model's capabilities and cost.
- Create a new API key named for the project or agent; set a spending limit and an expiration date.
- Ask your agent to help set OpenRouter up, store the key privately, then give it a model name and a small task (their example: cheaper hat-ad mockups).
- Compare results and check usage in OpenRouter — their mockups were useful for direction but did not match the real hat or a GPT Image 2.5 version.
Pro tip from the letter: tell your agent to create a reusable skill that uses OpenRouter for cheap bulk brainstorming variations.
Chinese researchers map "the last AI built by humans"
More than 30 Chinese AI researchers (including from ByteDance, Tsinghua, and Shanghai AI Lab) published "The Last AI Built by Humans" — a roadmap of five levels of recursive self-improvement (RSI), ending with AI that builds its successors.
- Level 1: AI carries out upgrades humans designed. Level 2: it diagnoses its own weak spots and decides how to fix them.
- Levels 3–4: it takes over what it learns next and how it adapts after launch. Level 5: it overhauls the improvement process itself.
- They say coding has the clearest path to RSI (fixes can be tested instantly); robotics, science, and medicine face slower, costlier feedback.
- They sorted 491 existing papers onto the ladder: about 75% at Level 1 or 2; under 6% at Level 5.
The Rundown's take: Level 5 is what Western labs list as a risk in safety frameworks (OpenAI, Anthropic, Google). The Chinese paper treats RSI more like a milestone than a threat — consistent with Beijing's pushback on slowdown talk earlier in the week.
Quick hits
- Community: a Rundown reader used ChatGPT as an emergency vet coordinator — researched practices, sent personalised Gmail outreach, kept a live tracker, contacted 20+ clinics, and got a next-morning evaluation/surgery offer for a dog with a dental abscess and eye-pressure damage.
- Gemini 3.8 Live: Google's new voice models that can keep speaking while they think; an Extended Thinking variant topped AA's speech-to-speech quality ranking.
- Odyssey 3: a world model aimed at controlling robot arms, humanoids, self-driving cars, drones, and games.
- Project Lily: 404 Media reported OpenAI uses hundreds of human contractors to read and rate real ChatGPT chats, often without the user's knowledge.
- Meta One: Meta's paid tier from $2.99 to $499/month for extra Meta AI usage and creator tools.
- Also named: StepAudio 3 (five-model audio suite for voice agents, transcription, music); Ask the Community in the Hub.
What this means for a GP
None of this is clinical advice, and none of it should change how you treat a patient tomorrow. It is why a busy doctor might still skim the letter. Jev is the conceptual shift: not another chatbot for patients, but cheap, fast "pick from a list" judgment calls that software vendors may embed for triage-like sorting, scoring, or checking another AI's output — useful vocabulary when a practice tool claims "AI inside." Salesforce Koa is the enterprise pattern (train on synthetic scenarios, host in-house, keep data on-premises) you will see more vendors copy. OpenRouter is optional ops geekery if you ever wire coding agents to cheaper models. The RSI paper and Project Lily are the patient-inbox risk stories: self-improving AI as a milestone vs a threat, and the reminder that chat logs can be read by humans for rating. Meta One and Gemini Live are consumer pricing/voice noise — not clinic protocol.