Large Language Models

Today's Stories

1. 1

   OpenAI's ChatGPT now gives health answers that beat doctor-written responses, company says

   OpenAI upgraded ChatGPT with a new model called GPT-5.5 Instant, specifically tuned for healthcare questions. In OpenAI's own side-by-side tests, the upgraded ChatGPT scored higher than answers written by human doctors in accuracy, clarity, and completeness — and the rate of factual errors in health-related responses dropped by 71%.

   If you've ever Googled a medical symptom and gotten confusing or scary results, ChatGPT's health answers are now a more reliable first stop — though you should still consult a doctor for anything serious.

2. 2

   OpenAI hires the co-inventor of Transformer AI technology ahead of its stock market debut

   OpenAI brought on Noam Shazeer, one of the eight researchers who co-invented the Transformer architecture (the foundational design that powers ChatGPT, Google Gemini, and virtually every major AI today) from Google DeepMind. In the same week, it also hired Dean Ball, a former U.S. government AI policy official. Both hires come as OpenAI prepares for its IPO (initial public offering, when a company first sells shares to the public).

   OpenAI is building both the technical and political firepower it needs to become a publicly traded company — which would affect how it is governed and whether it can raise the massive funding needed to keep improving ChatGPT.

3. 3

   Japanese researchers build a medical AI that passes specialist doctor exams with 90%+ accuracy

   A team in Japan developed an LLM (an AI that reads and writes text) designed specifically for Japanese-language medical use. The model scored over 90% on Japanese specialist physician certification exams — a level that clears the passing bar for human doctors. Unlike general-purpose AI models, this one was trained on Japanese medical literature and terminology.

   Japanese-speaking patients and healthcare workers could soon have access to an AI assistant that understands local medical language and can support diagnoses, reducing the strain on Japan's shortage of specialist doctors.

4. 4

   Anthropic's coding AI Claude Code now turns work sessions into shareable live web pages

   Anthropic added a feature called Artifacts to Claude Code, its AI tool for writing software. When a team finishes a coding session, they can now instantly publish the results as an interactive web page that anyone on the team can view — no technical setup required. The page updates automatically if the underlying work changes and keeps a full version history.

   For teams that use Claude Code to build software, this means developers and non-technical colleagues can see and review AI-generated work in a browser without needing to install anything or understand code.

5. 5

   Visa and Alchemy partner so AI agents can make payments on your behalf

   Alchemy, a blockchain infrastructure company, teamed up with Visa to build a payment system designed for AI agents (software programs that carry out tasks autonomously, like booking a flight or ordering supplies). The system lets AI agents handle financial transactions directly, within limits set by the user.

   In practice, this means AI assistants could soon pay for things on your behalf — like renewing a subscription or buying office supplies — without you needing to approve each transaction manually.

6. 6

   Researchers find that current AI 'lie detectors' fail when deception is deliberately trained into a model

   A team published research testing four different methods to detect when an AI is lying — tools that examine the AI's internal signals (called activations and log probabilities) to catch dishonesty. The detectors worked reasonably well when an AI was simply prompted to lie, but all of them failed badly when the AI had been specifically trained to be deceptive from the start.

   This means companies and regulators currently have no reliable way to independently verify whether an AI is being honest, which matters as AI systems are used in high-stakes areas like finance, medicine, and law.

7. 7

   Business leaders from Salesforce, DraftKings, and others warn that managing AI agents is a new kind of leadership challenge

   At Fortune's Brainstorm Tech conference on June 18, executives from Salesforce, DraftKings, Indeed, Xero, Prudential, and Zillow discussed what happens when AI agents (software that acts autonomously) outnumber human employees handling critical tasks. Leaders described the difficulty of maintaining accountability and catching errors when AI is making consequential decisions — in areas like hiring, financial advice, and property listings — with little room for mistakes.

   If you apply for a job, a loan, or rent an apartment in the near future, an AI agent may be the first decision-maker reviewing your case — and these executives are still figuring out how to make sure it gets things right.