• Starter AI
  • Posts
  • AI safety concerns, Introducing: The Abacus method, AI twins for Zoom meetings

AI safety concerns, Introducing: The Abacus method, AI twins for Zoom meetings

With many benefits, there are also risks.

Hello, Starters!

Developing a new technology that has the potential to change the world as we know it comes with great responsibility. As AI keeps making its way into all parts of our lives, it is normal that there are concerns about its risks.

Here’s what you’ll find today:

  • Ex-OpenAI employees call for stronger AI safety measures

  • Researchers present the Abacus method

  • Your AI twin may join your next Zoom meeting

  • Amazon’s new “Project PI”

  • Tesla’s AI chips go to X

  • And more.

OpenAI continues to make headlines, and this time it is regarding a very important matter: AI safety. Thirteen former company employees have signed an open letter, endorsed by Geoffrey Hinton, calling for more protection for whistleblowers. It has been reported that AI companies are retaliating against those who share concerns and criticisms about AI development.

In the letter, they ask for open pathways to report issues without fear of recrimination, claiming that governments have been slow in offering protection. Although they believe in AI's potential benefits, they are also aware of the many risks surrounding its fast-paced growth.

Researchers from the University of Maryland, the Lawrence Livermore National Laboratory, the ELLIS Institute Tübingen, the Max Planck Institute for Intelligent Systems, the Tübingen AI Center, and Carnegie Mellon University have discovered a method to improve AI models' performance on arithmetic tasks. They call it “Abacus,” and it consists of a series of embeddings that help models keep track of digit positions in long numbers.

Even though models such as GPT-4 are capable of doing basic maths, they often struggle with long additions involving up to 100 digits. Because of this, they cannot perform calculations accurately as they keep losing track of digits. This is where Abacus comes in handy, improving AI's maths skills.

In a recent interview, Eric Yuan, Zoom's CEO, shared his vision of turning the platform into an "AI-first company," which includes a future where a "digital twin" can join meetings on our behalf. However, he believes that this is still a far-off possibility as LLMs continue to advance.

He also discussed the potential for each of us to train our own LLM, allowing them to make decisions for us if we don't want to join a call. Many have disagreed with Yuan's position, arguing that AI is not reliable enough yet, leading to concerns about deepfakes and hallucinations. Yet, the concept of an AI-powered "digital twin" may not be as unrealistic as it seems at the moment; it's just a matter of time to see how it could unfold.

📦Dedicated to integrating AI into many of its processes, Amazon has announced Project PI. This system combines generative AI and computer vision to detect damages, wrong sizes, or colours in items before shipping them to customers. Currently active in several US warehouses, they plan to expand to more sites by the end of the year.

🤖According to emails from Nvidia employees, Tesla's AI ambitions are being put on hold to fuel those of X. CEO Elon Musk has ordered the diversion of the company's AI chips, originally destined for Tesla, to be moved to X instead, most likely for xAI, as Musk continues to push against contenders.

What did you think of today's newsletter?

Login or Subscribe to participate in polls.

Thank you for reading!