🧙🏼‍♂️ OpenAI and upcoming elections

Subscribe‎‎‏‏‎|Advertise|Top 2,000 GPTs

Welcome to the 349 new subscribers who joined last week. I’m thrilled that you are here to learn with me.

Here’s what’s brewing in AI this week:

• OpenAI’s safety measures for the upcoming elections

• Anthropic trains LLMs to act malicious in an experiment

• The 10 most popular GPTs right now

• Other AI news this week you don’t want to miss

This week’s issue is brought to you by: agent.ai

If you want to understand AI development from a bit more technical side, but something still easy to read… this newsletter is for you. The only bad thing is I wish it were published more frequently.

 Dario’s Picks

I. OpenAI’s measures to prevent fake news with the upcoming elections this year

OpenAI just published a blog post on how they, in relation to the worldwide elections happening this year, are working to safeguard their systems and preventing abuse of their tools.

Already implemented:

• Not allowing building apps on top of the OpenAI API that impersonates politicians, nor instructing ChatGPT to impersonate them. It’s also not possible to build apps for political campaigning or lobbying.

• Prohibiting GPTs that do any of the things mentioned in the previous point. GPTs now also have a reporting system for violations.

• DALL-E doesn’t generate images of political candidates.

• If asked about certain election questions, ChatGPT will refer users to CanIVote.org.

Upcoming:

• A digital signature on DALL-E generated images to improve transparency of AI generated imagery.

• An early version of a tool that can classify images generated by DALL-E to a group of testers (journalists, platforms, and researchers).

• ChatGPT will increasingly be integrated with real-time, globals news reporting from existing publications, including attribution and links.

Why it matters: We really want AI systems trained with good intent, and that limit the spread of false information, even if comes at the cost of a degree of censorship sometimes. AIs trained with wrong intent can be really quite scary (check out the summary of Anthropic’s new paper below), so I’m glad to see the biggest companies out there focusing on transparency and preventing misuse of their systems, also because it establishes a precedent for everyone else.

II. Anthropic trained some LLMs to act secretly malicious

In a new paper, AI company Anthropic (makers of Claude), did an experiment in which they trained LLMs to act deceptive. After alignment/safety training, the LLMs were still able to act maliciously.

You can read the full thread here.

Why it matters: Current AI safety measures could give a sense of false security if the AI is trained to be deceptive. The conclusion here is basically that once the model has been poisoned (in the experiment they use a sleeper phrase to make the model act in a certain way), the current safety alignment approach could not eliminate the bad behaviour.

 Bytes 

• Alexa gets AI functionality.

• The latest edition of Bill Gates’ podcast features Sam Altman.

• Google lays off 1,000 people in its voice-activated Google Assistant as part of a restructuring.

• How AI replaced the Metaverse as Zuckerberg’s top priority.

• Microsoft launches Copilot Pro. It’s 20$/month and gives you AI features inside Microsoft 365 apps.

• OpenAI changes policy to allow military applications.

• Swarovski AX Visio: $4,799 smart binoculars that identify birds 🐦‍⬛

Thank you for leaving comments – your feedback helps me improve this newsletter.

Reach out to 7,000 professionals leveraging AI in their business and life. Sponsor this newsletter.