- Starter AI
- Posts
- Sutskever announces SSI, Microsoft’s Florence-2, OpenVLA’s approach
Sutskever announces SSI, Microsoft’s Florence-2, OpenVLA’s approach
It's all about safe development.
Sthefania
June 20, 2024
Hello, Starters!
AI may be a controversial topic, as many companies and researchers differ in their opinions about its development. Some focus on commercial outputs, while others believe safety is the key.
Here’s what you’ll find today:
Ilya Sutskever’s latest endeavour
Microsoft introduces Florence-2
OpenVLA: An open-source robotic approach
Luma AI announces new features
Geoffrey Hinton backs CuspAI
And more.
👀 Ilya Sutskever’s latest endeavour (1 min)
Just a few weeks after departing from OpenAI, Ilya Sutskever has revealed that he's starting a new AI startup called Safe Superintelligence Inc., which will focus on developing safe AI systems. This is not far from Sutskever's intentions at OpenAI, which ultimately were part of his imminent exit since there were different visions for the company's future.
Now, with a project of his own, which also includes Daniel Gross, former AI lead at Apple, and Daniel Levy, an ex-member of OpenAI's technical staff, Sutskever is focusing on what he believes is the right way to develop AI systems, prioritising safety, security, and progress without worrying about commercial pressure.
👁️ Microsoft introduces Florence-2 (2 min)
Microsoft's newest model can be found on Hugging Face, directly from their Azure AI team. They've developed Florence-2, a model destined for vision and vision-language tasks, available in 232M and 771M parameters. Florence-2 stands out for its capabilities in object detection, segmentation, captioning, and visual grounding.
This foundation model is yet to be tested in real-world domains, but it's expected to reduce costs, streamline fine-tuning tasks, and provide an overall approach to most vision applications.
🤖 OpenVLA: An open-source robotic approach (5 min)
To showcase how foundation models are set to transform robotics, a team of researchers from Stanford University, UC Berkeley, Toyota Research Institute, and Google DeepMind has developed OpenVLA, a vision-language-action open-source model that has been trained in 970k robot demonstrations.
OpenVLA has several benefits, one of which is that it has been optimised to run on consumer-grade GPUs, reducing fine-tuning costs. Also, based on its training, this model outperforms similar contenders, as it adapts easily to new tasks and setups.
🎥Following its recent launch, Luma AI has announced a new feature for its video generation model, "Dream Machine." Users can now create high-quality clips with an extended length of 1:20 minutes. Also, they're planning to add more editing features and prompt controls soon.
🔎Geoffrey Hinton, well-known as one of AI's godfathers and for being vocal about the technology's potential risks, has recently agreed to serve as an advisor for CuspAI, a new company that uses AI to design next-gen building materials that can aid in the fight against climate change.
⚡️Quick links
What did you think of today's newsletter? |