Apple Vision Pro
This week's major news was Apple entered the mixed reality industry with a massive bang with Apple Vision Pro. Although this isn’t AI news directly, it’s too interesting to pass up writing about it!
Apple’s new headset is their first attempt at what they are calling Spacial Computing. In their vision, this is the next generation of how people interact with computers. Their demo videos primarily focused on the computing and media consumption use cases, with very little focus on gaming, which was surprising.
The most jaw-dropping aspect of the launch was not the technology itself but the price tag. Coming in at $3,499 will be out of reach for many consumers. This strategy is similar to the first iPhone, initially priced at $600. Looking back at that price now, it seems inexpensive, but at the time, phones were generally priced in the $100-200 range with carrier subsidies. They quickly brought the price down, but since then, have released many premium iterations of iPhone, and people are used to paying $1,200+ for the top-of-the-line iPhones. Apple Vision Pro will follow the same tactic: get the product out to market to the early adopters, learn from them, and release less expensive versions. Then, as they gain market share, increase the price again.
Apple Vision Pro might be the most sophisticated consumer technology ever released. They have so much tech packed into a small form factor, it’s no wonder the price point is what it is. They even invented a new chip (R1) to handle all the computing necessary from all the sensors and cameras. Multiple cameras, LiDar, multiple screens, 4k resolution in both eyes, they really packed everything into this tiny device.
I’m personally very excited about their take on AR/VR. I’ve used the Oculus and other VR headsets. I’ve been blown away by the technology but have not felt the need to use a device daily. I hope Vision Pro changes that.
AI News
This week was admittedly a bit slow for AI news. There were a ton of technical advancements and launches, so here’s a roundup of all of that news:
Sam Altman’s Trip To India
First, Sam Altman visited India to discuss AI and made some interesting statements:
What does Sam lose the most sleep over? "What I lose the most sleep over is the hypothetical idea that we already have done something really bad by launching ChatGPT." He reiterated that there is an existential risk to humanity from AI and urged for regulation.
Is it worth trying to compete with OpenAI? “Look, the way this works is we’re going to tell you it’s totally hopeless to compete with us on training foundation models you shouldn’t try, and it’s your job to, like, try anyway.”
StabiltyAI Founder’s Bad Week
Next, one of the most influential people in the world of AI, Emad Mostaque, the founder and CEO of StabilityAI, had a negative piece written about him in which Forbes outlines a history of exaggeration. In the article, Forbes writes that Emad doesn’t have a master’s degree from Oxford, as he claims, and Stable Diffusion wasn’t initially created by StabilityAI. Emad already responded to the claims here.
Gen2 Released Publicly
Runway’s Gen2 text-to-video product was publicly released (finally). I’ve been playing around with it and will likely make a video about it. It’s still in the early stages, but the technology promises to change video creation forever.
Open Source Keeps Up Momentum
The open-source AI community continues to keep up strong momentum this week, with the release of multiple new models and research papers: starchat beta, open source text-to-video, and text-to-audio by Meta.
Also, an incredibly fascinating research paper from Microsoft Research was released detailing a new method of fine-tuning that achieves much better results than any other open-source model. Called Orca 13b, it is trained using explanations of logic and reasoning rather than just prompt/answer pairs. This technique was able to achieve much better performance. I made a video about it:
LLM Leaderboard Drama
This week, there was also some drama in how the LLM Leaderboard benchmarks are calculated. The short of it is there is no consistency, and it’s still the wild west. The main issue is many AI leaders were surprised to see LLaMA 65b perform significantly worse than the newly released Falcon model. It turns out it performs much better when all things are equal. Follow the news here.
