🍭 OpenAI fired them. Here’s their side 👀

Better Tools Need Better Checks.

Good morning. The safety news is back. The three dismissed OpenAI employees working on safety responsed. Their take: OpenAI agree to third party safety auditors, preserve monitorability of models, let openAI reaffirm its prior open culture of safety, affirm open and transparent dialogue culture. Claude terms and conditions now addresses bully behaviour toward its models.

Dynamic AI chat versus dynamic AI work:

On Thursday we learned about OpenAI’s GPT-6 with intelligent UI. Working towards a software that adapts to the user instead of the user learning a rigid software. Presently, in chat mode. Claude released dashboards in motion aiming to make traditionally separate applications dynamic. This is aims to make work mode more dynamic.

Let’s dive in 👇

Today’s Menu 🥠

• Who gets to question AI safety? The labs, the researchers and outside scrutiny.

• Claude builds the dashboard. More threats to UI SaaS?

• Fast Snacks: Revenue math, security, AI in the clinic and fake bylines.

• Toolbox: Read documents, steer tasks and check AI watermarks.

• Snack Quiz: Which robot photo is real?

Friday, October 9 · Read Time: 6 minutes

The Big Bite · Safety and Accountability
🛡️ Who Gets to Question the AI Labs?

The Bite: Two developments put the people writing AI safety rules under a closer look.

Three researchers fired by OpenAI, Tomek Korbak, Jasmine Wang and Mikita Balesni, have published a letter warning that their dismissals could make colleagues afraid to raise concerns or work with outside experts. They say they acted within their roles and the working norms at the time. Their account is disputed: OpenAI denies retaliation and says the dismissals followed violations involving sensitive research information.

In their words: “If conduct that was considered normal last month now constitutes grounds for sudden dismissal, everyone at OpenAI is left guessing where the line is.” Korbak, Wang and Balesni, in their letter.

Separately, Anthropic’s updated usage policy will prohibit extreme, repeated, purposeless abuse directed at its models from November 12. Ordinary frustration and model testing remain outside that restriction. Of course, this is a rule for using Claude, not evidence that the model has feelings. I ask myself, who is bullying the models? 👀

Why it Bites: AI safety should be designed to protect humanity. Establishing an open and safe culture in this domain is critical to keeping bad actors in check. However, no rulebook currently exists. It is not a trivial problem, and it raises the question: who gets to decide these rules? At the same time, it is odd that Anthropic has published rules against abusing its models. While most people using AI are not abusing these systems—except perhaps when they get frustrated trying to babysit their agentic workflows—Anthropic seems confused about model consciousness. Since AI models are not trained on bullying behavior, bullying a model will not teach it to become aggressive toward other users. Where do you land on the spectrum?

Our take: A safety promise becomes more useful when someone outside the company can examine it. Watch what access independent evaluators actually receive, what they can disclose, and whether employees know how to escalate concerns. Those details tell us more than a lab’s preferred definition of safety. As for consciousness, I think thats characteristic of living beings only, not computer models, no matter how large the data center is.

Second Bite · Claude
🤖 Your Next Dashboard Starts With a Question

The Bite: Claude, Anthropic’s AI assistant, is moving further into the work usually done in spreadsheets and presentations. On Thursday, Anthropic introduced Dashboards and Motion, two beta tools that turn a request into something you can share.

Connect company data from platforms such as Snowflake or Salesforce, then ask how this week’s sales compare with last month’s. Dashboards builds the view and updates it as the data changes. You can inspect the query behind a number and see when each chart last refreshed. Motion turns text, charts, shapes and images into editable animations, downloadable as MP4 files. Think a product walkthrough or an animated explanation of quarterly results. It animates code, rather than generating video footage.

Dashboards is in beta on paid plans. Motion is in beta on Team and Enterprise. Both are off by default for Enterprise until an admin enables them. Docs, Slides and Design also leave beta on every plan, including Free, with Enterprise access controlled by admins.

Why it Bites: User interfaces may be becoming disposable. The advantage: getting a useful first version could become significantly easier for people who cannot write database queries or animate a slide. This means fewer apps to context switch between and a much better workflow experience—assuming we can trust the output for live work.

Our take: Start with one question you already know how to answer. Check the query, refresh date and sharing permissions before circulating the result. A beautiful chart still needs someone who understands what it counts. Run a previous task performed in SQL and see if they align.

Fast Snacks 🥡

💵 A $50 Billion Figure Needs a Footnote
Bloomberg reports OpenAI’s annualized revenue is near $50 billion, below earlier estimates around $70 billion. Its report points to differences in how cloud partner sales are counted when comparing OpenAI with Anthropic. Annualized revenue projects a current pace over a year; it is not a completed year’s audited sales. The useful takeaway: check what a headline number counts before deciding what changed.

🔐 Finding Bugs Is Only Half the Job
Anthropic's new Cyber Mission brings Claude models and engineers to defenders of power grids, water systems and other critical infrastructure. It also offers participating open source projects free security scans through OSS Scanner. Those reports arrive without human review and can contain mistakes. More findings do not automatically mean safer systems. Maintainers still need time to verify them, prioritize real risks and apply fixes without breaking what everyone depends on.

🩺 An AI Intake Visit, With a Doctor Watching
Google's AMIE research chatbot met real patients before their appointments in a study newly published in The Lancet. Ninety eight patients completed both steps. Doctors supervised every conversation: none triggered a required safety stop, although supervisors noted one hallucination. Conducted at one clinic in 2025, this small study supports further testing, not replacing doctors. The useful question is whether AI can help people arrive better prepared while preserving the human judgment that keeps care safe.

📰 A Familiar Publication Can Carry a Fake Byline
OpenAI's Thursday investigation describes an operation originating in Iran that used seven fabricated journalist identities. OpenAI identified almost 100 articles published or syndicated under those names. That count does not mean 100 original articles, and the company could not identify the actor behind the operation. This is a new report about older activity. The reader lesson: a recognizable publisher helps a story travel, but original evidence still deserves a click.

Toolbox
🧰 Three useful things to try

📄 LightOnOCR3: October 8 release. This downloadable model turns scanned pages into usable text and can extract chart values as tables, with labeled boxes showing where content sits on the page. Useful for developers building document search or data extraction. It carries an Apache 2.0 license and requires deployment.

🧭 Codex steering: October 8 update. OpenAI is rolling out faster course corrections for running Codex tasks in the ChatGPT desktop app. Useful when your coding agent heads in the wrong direction. Check Settings > General to choose whether new messages steer the current task or wait for the next run. Try correcting one small task before delegating more. Availability may vary during rollout.

🔎 SynthID Detector: Earlier this week, October 7. Google’s public tool checks media for supported AI watermarks from Google and participating partners. Available globally in English, it gives editors and curious readers another check before sharing an image, video or audio clip. Try a file whose origin you already know. Note: no detected watermark does not mean the content is real.

Test Your Skill
Snack Quiz: Choose The Real Image 🔍

One is a real photograph. One is an AI recreation. Take a close look at these robots in orbit.

Image A

Image B

Photo source and answer. Original photograph: NASA / Mark Vande Hei, February 2, 2022. Archival image used for this challenge.