Welcome to our roundup of this week’s freshest AI news.
This week AI took blurry videos and made them look awesome.
Microsoft released a tiny model that punches above its weight.
And an AI pilot engaged humans in a real-world F-16 dogfight.
Let’s dig in.
Looking sharp
Adobe teased us with demos of VideoGigaGAN, a model that upscales blurry video to look 8x sharper. We’ve seen attempts at video upscaling before but this is the first to generate smooth and sharp output from low-res videos.
This could give us HD versions of historical archive footage and even HD versions of old TV shows. There’s also the danger of your dad putting it to work on old family videos, so…
Generative AI is powering a surge of deep fake porn and a lot of it is ending up on social media platforms.
Meta’s Oversight Board has been assessing its handling of explicit deep fakes as Meta tries to devise a strategy to deal with the blurred lines of content moderation.
Deep fake porn and political propaganda aren’t likely to go away anytime soon. At least some people are having a little fun with generative AI and politicians.
News: “OpenAI video-generator Sora risks fueling propaganda and bias, experts say”. The internet:
byu/pergessleismydaddy insingularity
New models
This week saw more language models released with exciting developments in some very small models.
Meta dropped two versions of Llama 3, an 8B and a larger 70B parameter model. Benchmark testing shows these are now the leading “open-source” models with Llama-3-70B even competing with Google’s Gemini 1.5 Pro.
Meta says a huge 400B and multimodal versions of Llama 3 are currently being trained.
On the really small side of the language model scale, Microsoft launched Phi-3 Mini, a tiny but powerful LM. Phi-3 Mini proves that bigger isn’t always better.
Microsoft took an interesting approach with curated synthetic data to train the 3.8B parameter model which outperforms models twice its size.
This is the kind of model we’re likely to have on our phones soon.
Tech tumble
NVIDIA’s meteoric share price surge eventually took a breather. The stock tumbled 10% as other US tech stocks also took a battering. A drop of 10% represents $200B being shaved from the company’s valuation.
Is this a sign that investor AI enthusiasm is waning? Unlikely. Once next week’s Big Tech earnings reports are processed we’ll probably see normal service resume.
The right AI stuff
DARPA has been developing an AI fighter jet pilot for a while now with good results in simulated flights. This week they revealed that they tested an AI-piloted F-16 jet in a dogfight against human fighter pilots.
Who won? They’re not saying. But I’m guessing the next generation of fighter jets won’t need seats.
AI’s ability to act autonomously could be a boost for cybercrime. Researchers found that AI agents powered by GPT-4 can autonomously exploit cybersecurity vulnerabilities.
The fact that these vulnerabilities are normally published publicly and the AI agent only required 91 lines of code isn’t comforting.
It may not be as high-tech as piloting a jet or hacking software, but AI is taking jobs from teenagers at this Wendy’s.
Talking AI
This week we had the opportunity to interview some very interesting people involved in the AI, robotics, and blockchain space.
There’s a growing intersection between these technologies and the future will likely be shaped by a combination of them.
Cogniteam CEO Dr. Yehuda Elmaliah explained how their software accelerates robotics development, deployment, and remote management.
Radovan Kavicky, an AI & Data Science evangelist at AIslovakIA gave insights into the importance of safely developing AGI and why he thinks OpenAI’s approach isn’t the way to go.
In other news…
Here are some other clickworthy AI stories we enjoyed this week:
- Microsoft’s VASA-1 model turns a single image and audio into lifelike talking face videos.
- Artificial intelligence can predict political beliefs from expressionless faces.
- MIT study reveals an AI model that can predict your next move.
- The ethics of advanced AI assistants. Exploring the promise and risks of a future with more capable AI.
- OpenAI CEO Sam Altman invests in solar power firm Exowatt to fuel AI datacenters
- Boston Dynamic’s new electric robot looks pretty ominous.
“boston dynamics robot with dune music isn’t real, it can’t hurt you”
boston dynamics robot with dune music: pic.twitter.com/fvSjtrMYzo
— frye (@___frye) April 17, 2024
And that’s a wrap.
Let’s hope Adobe lets us play with VideoGigaGAN soon. The demos look amazing. Which old TV show or movie is first on your list to make HD?
Would you fly in a fighter jet piloted by an AI? I’d love to hear an interview with the human pilots it went up against. Someone in Hollywood is probably already working on an AI-inspired version of Top Gun 3. Maybe they’ll use Sora to make it.
It’s been great getting to interview people in the industry and we’ll be posting more conversations soon.
Let us know if you have contact details of someone you think we should interview next, and keep sending us links to AI stories that we may have missed.
