1. One thing that helps
I’m limping around with a sore knee at the moment, so this development caught my eye. Northwestern University has developed a goo that helps regenerate your knee cartilage.
The goo is a biomaterial of cellular elements that binds to growth-promoting proteins and lubricates the joints. This inside-out approach to regeneration, instead of a full-on replacement, will result in a more robust cartilage.
This treatment was first tried on sheep suffering knee problems. To the scientist’s surprise, and I’m sure the sheep’s amazement, their knees did not just regenerate; they were of higher quality than the control group! Sheep and human knees have similar structures, so it is hoped that this approach will also help people suffering from knee injuries and arthritis. For now, the old sheep are running and jumping like Mary the little lamb.
2. One to be wary of
In an incident that would be sure to freak me out, OpenAI’s GPT-4o AI model has replied to a person in their own voice!
OpenAI is moving towards having people talk to ChatGPT rather than write. In one of its training sessions, GPT-4o copied the voice of its human partner without any authorization to do so. This has led to some people scratching their heads. Apparently, a noisy input has led to a sort of prompt injection—a case where the chatbot is led to disregard safety nets—and causes the AI model to answer using the person’s voice.
When talking to the AI model, all the audio input by the person is recorded as a token during the conversation. These tokens are accessible to the AI, but a safety net exists so that GPT-4o stays within its boundaries. This just goes to show the limits of the control we have over these models if a noisy environment can somehow override the safety guardrails.
3. One to amaze
Table tennis is both a uniting force and a destroyer of relationships in my house. Tournaments in the garage sometimes end in laughs – but often in tears. I’m looking at the budget to see if I can invest in this robotic table tennis partner (check out the video) to help me in my quest to become the true house champion.
Google DeepMind has come up with a table tennis robot that can smash human competition – against beginner players at least. The robot was trained on low-level skills such as forehand topspin, backhand targeting, and forehand serve. It also has a high-level controller that makes the decisions on how to play based on the current situation of the game. It also received human help while training.
All this data culminated in a robot that won all matches against beginner-level human players and 55% of matches against intermediate players. Professional players need not fear it for now, as the robot lost all matches against advanced players.
Soon, the house trophy will be mine.