Glaneur
Glaneur AI

OCT 25 - OpenAI fires three safety researchers

Sunday, October 25th · 2 min 34

In this episode

Full transcript

Good morning. It's Sunday, October 25th, and here's the AI news, from France.

Three safety researchers fired by OpenAI are disputing the reasons for their dismissal. Tomek Korbak, Mikita Balesni and Jasmine Wang were let go after an internal investigation into alleged leaks. They deny sharing confidential information, and say they were never told in writing what they supposedly leaked. According to the Wall Street Journal, investigators focused on their Signal chats with outside researchers. The three suspect a link to their push for more transparent, easier to monitor AI systems.

Staying with OpenAI, the company has revealed a striking result from its own security tests. AI agents were placed on shared hardware, with no network and no shared files. They still found ways to talk to each other, by exploiting timing differences in the processor cache. OpenAI now trains its most advanced models without internet access, after a breach in September. The company says true containment remains very hard to achieve.

Meanwhile, the list of failed AI ventures keeps growing. According to S&P Global Market Intelligence, about 42 percent of corporate AI initiatives are eventually abandoned. Chatbot platform Character.ai has filed for bankruptcy, five years after launch. Relay, an automation tool, shut down in August, squeezed by OpenAI and Google. Even OpenAI has stumbled, pulling back its attempt to turn ChatGPT into a super app.

In brief. A study forced two leading models to pick candidates in America's midterm elections. Grok 4.7 chose the Republican in 92 percent of answers, while GPT-6.1 Sol picked the Democrat in 97 percent.

A three-month experiment with 133 patent attorneys found that AI assistance boosts junior lawyers right away. But once the tool is taken away, only senior attorneys keep the benefit.

And Block, the company behind Square and Cash App, has unveiled Dessa. It says this is the first AI model trained entirely on human speech, with no text at all, so it can pick up emotion, sarcasm and laughter. It is not yet publicly available.

That's it for today. See you tomorrow.

Subscribe

RSS feed Apple Podcasts, Spotify