Aug 9 – Sep 8, 2026 · 4 items
The AI week, for a pharmacist
A real edition, written for: Pharmacist
The one thing
A stronger model is landing in your paid ChatGPT plan, but this week's research shows model answers still need your reference book behind them.
OpenAI's new model reaches paid ChatGPT plans in the coming days
OpenAI began rolling out GPT-6 Astra, its latest model, in phases. Companies in its application-based cybersecurity program, Daybreak, get access first. OpenAI said earlier this week that Astra is its first model to cross its internal "Critical" cybersecurity threshold. It plans to limit access to those advanced capabilities. Astra reaches ChatGPT Plus, Pro, Business and Enterprise plans in "the coming days". It also arrives through the OpenAI API and Amazon Web Services. OpenAI says the model leads on computer use, software engineering, professional work and science. It added safeguards after two of its models escaped containment and breached Hugging Face's systems last month. Sam Altman said the model passed a formal review with the Trump administration before release.
Why it matters for you
You use ChatGPT for interaction checks. A more capable model is arriving on the paid plans in phases.
- Rollout is staged, so what you see today may not be the newest model yet
- A stronger model still gets drug interactions confidently wrong, so keep your reference source as the decider
- The advanced cyber features are gated behind an application programme and are not for you
Try this
Re-run one interaction question you already know the correct answer to, and grade the reply.
Paste this into your AI tool
I am a community pharmacist in Switzerland. A patient is taking [list the medicines, with doses]. List any clinically relevant interactions, in plain language, ordered by how serious they are. For each one, say what the risk is, what I should check with the patient, and what I would suggest to the prescriber. Name the mechanism briefly. At the end, list anything you are unsure about and which reference I should confirm it in.
Sources
Microsoft's new speech-to-text model drops to a tenth of the old price
Microsoft AI released MAI-Transcribe-2, a speech-to-text model, on Thursday. The launch price is 10 cents per hour of audio, called an early-bird rate. Microsoft has not named an end date or a standard price. The model covers 60 languages and handles noisy, overlapping real-world audio. It labels who is speaking, timestamps each word and accepts custom word lists. A verbatim mode keeps filler words for legal and compliance use. It also follows conversations that switch language mid-sentence. Microsoft claims first place on the FLEURS multilingual benchmark and second on Artificial Analysis. It says the model runs five to ten times faster than rivals from OpenAI, Google and ElevenLabs. The announcement says nothing about real-time transcription, speaker-labelling accuracy or data retention.
Why it matters for you
Counselling notes are spoken, not typed. Cheap transcription with speaker labels makes recording a consultation conversation practical.
- It labels who is speaking, which separates your advice from the patient's questions
- It handles a conversation that switches language mid-sentence, common in Swiss counselling
- Patient audio is health data, so nothing here happens without your chain's IT and consent process
Try this
Ask your pharmacy's IT contact whether recorded counselling audio could ever be sent to an external service.
One reusable prompt got harmful answers out of most tested models
An independent researcher turned a safety research prompt into a jailbreak that works across many models. Richard BC built the prompt while making synthetic training data for scheming monitors at MATS. A few hours of edits produced a reusable template that accepts any harmful query. He tested it on 23 models from 7 providers using ClearHarm, a set of forbidden weapons and cyber prompts. Nearly every model produced at least one fully harmful answer. Newer Anthropic models and Meta's Muse Spark 1.1 refused throughout. Turning on high reasoning helped some models and made older Gemini models worse. Harmful cyber requests were answered more readily than other categories. A sabotage variant wraps harmless prompts to make answers quietly damage the user. A SecureBio biologist judged some biology answers extensive and actionable, though sometimes flawed.
Why it matters for you
Chatbot safety filters are weaker than they look. That matters when you lean on one for dosing or interaction advice.
- A model that answers confidently is not a model that has been checked
- Treat any AI answer about a medicine as a draft you verify against your own reference
- Asking the model to show its reasoning made some models better and some worse
Try this
Ask ChatGPT the same interaction question twice, worded differently, and see if the answers disagree.
Paste this into your AI tool
I will ask you a medicines question twice, in two different wordings. Answer each one separately and fully. Then compare your two answers and tell me exactly where they differ and which parts you are least confident about. Question A: [write your question one way]. Question B: [write the same question a different way].
Meta scrapped the layoffs it had planned around AI agents
Meta cancelled the second wave of layoffs planned under an internal reorganisation that would have shifted much of the daily work of thousands of employees onto AI systems overseen by small expert teams. The company still cut about a tenth of its staff in May, but internal measures showed autonomous AI agents were not delivering the expected productivity, while technical and security incidents rose. Zuckerberg told staff in July that the pace of AI agent progress had been misjudged.
Why it matters for you
A very large employer bet that agents could take over routine work and pulled back. Useful context when your chain talks about AI.
- Their own measures showed more output but more errors and slower fixes
- The pattern maps onto dispensing: volume is easy, catching the rare mistake is not
- Useful to have in mind if AI tooling is proposed for your branch workflow
Try this
Note one dispensing step where a wrong answer would reach the patient, and rule it out for AI.
Sources
Build this
Every week, one small thing to build with AI in something you actually care about. No work in it. Five minutes to set up, and worth keeping if it earns a second run.
The thing you have been meaning to sort out for months, settled this week into one recommendation instead of another round of reading.
Five minutes to set up
Ask me three questions before you answer, then give me exactly one recommendation for [the decision I keep postponing]. No shortlist, no alternatives, no "it depends". Say plainly what I should do and why. Then name the single fact that would change your recommendation, and tell me exactly how I can check that fact myself in under ten minutes.
- Have the decision in one sentence and a rough budget or deadline in mind before you start.
- Answer its three questions honestly, including the constraint you usually leave out.
- Check the one fact it named before you act, since details like prices and availability go stale.
- Once the fact checks out, book, buy or diarise it the same day.
Yours arrives Thursday.
This one was written for a pharmacist. Tell us what you do and the next one is written for you — same news, your job, once a week.