- Everyday AI
- Posts
- Is Microsoft's VASA-1 too dangerous? 😰
Is Microsoft's VASA-1 too dangerous? 😰
Concerns with AI deepfakes, Perplexity raises another $250 million, Microsoft's Phi-3 mini AI model, hidden AI PDF feature in Copilot and more!
👉 Subscribe Here | 🗣 Hire Us To Speak | 🤝 Partner with Us | 🤖 Grow with GenAI
Outsmart The Future
Sup y’all 👋
Pretty soon, we’re gonna run a cool giveaway. We thought we’d let you choose what’s best.
Vote and let us know what might be your fave for future giveaways.
(Vote to see live results)
What would you like to see out of an Everyday AI giveaway?Approximate value is in parenthesis |
✌️
P.S. make sure to attend our 1 year anniversary show live.
Today in Everyday AI
7 minute read
🎙 Daily Podcast Episode: Microsoft’s VASA-1 AI deepfake tech is so good it’s kinda scary. Is it too dangerous? We dive in. Give it a listen.
🕵️♂️ Fresh Finds: Perplexity raising another $250 million, Meta releases multimodal search in Ray Bans collab, Former Meta exec joins Microsoft’s AI team, and more. Read on for Fresh Finds.
🗞 Byte Sized Daily AI News: Microsoft Unveils Phi-3 Mini AI Model, Adobe adds Firefly Image 3 in Photoshop and Coca-Cola’s $1.1 Billion AI deal with Microsoft. For that and more, read on for Byte Sized News.
🚀 AI In 5: This hidden AI feature in Copilot changes how you read PDFs. See it here
🧠 Learn & Leveraging AI: So is Microsoft’s VASA-1 too dangerous? What does it mean for the future of deepfakes and its impact on you? Keep reading for that!
↩️ Don’t miss out: Did you miss our last newsletter? We talked about Apple acquiring an AI startup, Microsoft VASA-1 makes Mona Lisa rap and Salesforce calls for environmental AI regulations. Check it here!
Microsoft's VASA-1 AI Deepfake: So good it's dangerous? 😰
Microsoft recently unveiled its VASA-1 image-to-video model and it’s scary good.
We'll be taking a closer look at Microsoft's impressive VASA-1 deepfake cloning technology and asking one important question.
Is this AI tool too good to be safe?
Maybe.
Join the conversation and ask Jordan questions on Microsoft VASA-1 here.
Also on the pod today:
• Explanation of VASA-1's capabilities 🎥
• Concerns around Microsoft's VASA-1 😬
• Potential Benefits and Positive Use Cases ✅
It’ll be worth your 41 minutes:
Listen on our site:
Subscribe and listen on your favorite podcast platform
Listen on:
Upcoming Everyday AI Livestreams
Wednesday, April 24th at 7:30 am CST ⬇️
Thursday, April 25th - Will AI take our jobs? This might surprise you. (1 Year Anniversary Show)
Here’s our favorite AI finds from across the web:
New AI Tool Spotlight – Ayraa is enterprise search AI, SecBrain is an AI-powered voice recorder app and Bentolingo is personalized AI language learning.
Future of AI Search — According to reports, Perplexity is looking to raise at least another $250 million at a valuation of about $3 billion. Just in January, they were raising at a valuation of $540 million.
Big Tech – Jason Taylor, former VP of infrastructure at Meta, is now part of Microsoft's AI supercomputing team as corporate VP and deputy CTO.
Multimodal AI — Meta just announced interesting updates to its partnership with Ray Bans, brining multimodal search to the AI-enabled glasses.
Multimodal Meta AI is rolling out widely on Ray-Ban Meta starting today! It's a huge advancement for wearables & makes using AI more interactive & intuitive.
Excited to share more on our multimodal work w/ Meta AI (& Llama 3), stay tuned for more updates coming soon.
— Ahmad Al-Dahle (@Ahmad_Al_Dahle)
4:06 PM • Apr 23, 2024
Future of Work - Here’s how companies can create AI that is supportive of workers.
Pop Culture – Softbank is investing $1 billion into AI.
Read This – This outdoor monitoring security combines AI with live agents.
1. Microsoft Unveils Phi-3 Mini AI Model 🤏
Microsoft introduced Phi-3 Mini, a lightweight AI model with 3.8 billion parameters, now available on Azure, Hugging Face, and Ollama. This smaller model promises high performance similar to larger models like GPT-3.5 but at a more affordable cost, making it attractive for personal devices. Microsoft plans to release Phi-3 Small (7B parameters) and Phi-3 Medium (14B parameters) models, continuing their innovation in the realm of lightweight AI.
2. Adobe's Firefly Image 3 in Photoshop 🖼
Adobe's latest Firefly Image 3 model in Photoshop brings generative AI tools to give users more control over designs. With features like Reference Image, Generate Background, and Enhance Detail, users can easily generate images, change colors, and enhance details effortlessly. The new tools promise photorealistic quality and quicker creative processes, making design work more efficient and impressive.
3. Major AI Companies Join Forces to Combat Children Exploitation 🚫
OpenAI, Meta Platforms, and Google have united to implement new safety measures to combat the rise of exploitative content targeting children online. The alliance aims to curb the proliferation of generative AI tools used by predators to create harmful material. With a focus on child safety, this initiative seeks to safeguard vulnerable youth from the dangers of online exploitation.
4. Coca-Cola’s $1.1 Billion AI Deal with Microsoft 🥤
Microsoft and Coca-Cola are teaming up for a sweet $1.1 billion deal to dive into cloud computing and AI services together. Coca-Cola will be experimenting with Azure OpenAI, a service that uses technology from Microsoft-backed startup OpenAI. This collaboration will also see Coca-Cola exploring Microsoft's AI assistant, Copilot, to boost productivity in the beverage industry.
5. AWS Adds Custom Model Import Feature for GenAI Models 🤖
AWS has unveiled the Custom Model Import feature, enabling organizations to seamlessly integrate their own generative AI models into the Bedrock suite. This move promises enhanced model customization options and robust model evaluation tools for users. As AWS competes head-to-head with major players like Google's Vertex AI and Microsoft's Azure, the race for AI supremacy heats up, offering customers a wealth of cutting-edge capabilities at their fingertips.
6. Microsoft Offers Deepfakes Prevention in Political Campaigns 🗳
Microsoft is taking a stand against deepfakes in political campaigns by introducing Content Integrity tools to verify online content authenticity. These tools allow political parties, news organizations, and election officials to add "Content Credentials" to media, ensuring transparency in the digital information ecosystem. With the rise of generative AI in elections, these tools aim to combat the spread of misleading or altered content online.
This hidden AI feature in Copilot changes how you read PDFs!
Why stick to traditional PDF reading when you can have Microsoft Copilot do it for you?
We found a secret AI PDF feature in Copilot that’ll save you time on your PDFs.
Check out today's AI in 5.
Or see this related video:
🦾How You Can Leverage:
Is VASA-1 a very cool AI answer looking for a problem?
We think so.
If you follow Everyday AI, you’ve already heard about the shockingly impressive VASA-1 from Microsoft, an unreleased AI tool that essentially creates convincing deepfakes with only 1 photo and 1 audio source.
But it got us thinking — is this thing so good it’s dangerous?
For #HotTakeTuesday we looked at Microsoft’s VASA-1 model and talked about what the future of AI avatars and digital twins might look like.
We covered 7 things you need to know.
So, let’s get to it.
1 – What VASA-1? 🤷
In simple terms, VASA is an AI model that requires only a single photo and audio source, and you can create scarily human-like AI avatars.
Like this:
Introducing: VASA-1 by Microsoft Research.
TL;DR: single portrait photo + speech audio = hyper-realistic talking face video with precise lip-audio sync, lifelike facial behavior, and naturalistic head movements, generated in real time.
Tap to see all the videos. twitter.com/i/web/status/1…
— Eduardo Borges (@duborges)
12:54 AM • Apr 18, 2024
According to Microsoft, it is nearly done in real-time, with hardware that everyday people use.
The lip syncing, facial expressions and audio are all eerily human.
Yikes?
2 – VASA-1 is not released yet 🔐
Microsoft said (at least for now) they have no plans to release this.
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/e8cde9f8-6dcf-4291-8f1d-9e9858a8859c/safety.png?t=1713891575)
And that’s probably for the best. Right now, Microsoft only released a research paper and examples, which you can check here.
3 – Similar technology is already public 🧑💻️
This type of AI-powered tech isn’t necessarily new.
But it is different.
Here’s how.
We’d like to separate this tech into two buckets:
AI Deepfakes and AI Digital Avatars. (Different companies call them different things.)
There’s plenty of companies that offer similar tech, such as HeyGen, Synthesia, HourOne, D-ID and others.
So let’s try to draw the lines, shall we?
Digital Avatars — Using one of the aforementioned softwares above, you can use a character and give it text to read, or even work with the company to create a digital twin of yourself.
Even though these have realistic qualities, you can usually tell that this is a digital avatar.
But, it starts as a video avatar, and you have permission (from your own consent or from the company) to make the Digital Avatar.
Deepfakes — Using techniques like Microsoft’s VASA-1 or Alibaba’s Emo, you can create a deepfake out of a photo.
In general, you aren’t receiving permission from the real or fictional character.
The lines between the two are blurring, and it’s tough to see what’s a Digital Twin vs. what’s a deepfake.
In short:
AI Avatar/Digital Twin — Used with permission from a (somewhat) limited catalogue of approved options.
AI Deepfake — In most cases, not used with permission or using fake/generated images and videos, or potentially unauthorized subjects.
4 – VASA-1’s quality is outstanding 🤯
That’s what makes VASA-1 noteworthy, exciting and dangerous all at the same time.
Like, look at this example.
The results produced here are unlike anything we’ve seen. The realism is uncanny.
Ultra realistic AI-video from a photo
This is VASA-1 from Microsoft research
The improvements in quality we’re getting between each new release is incredible.
Links below
— Linus ●ᴗ● Ekenstam (@LinusEkenstam)
12:27 AM • Apr 18, 2024
5 – Positive use-cases for VASA-1 🤔
We started this recap by saying VASA-1 was a shiny solution looking for a problem.
We’re not saying there’s no use-cases. Because there definitely are.
Training, personalized learning & development, helping break down communication barriers, helping us learn other cultures without prejudice, etc.
There’s so much upside.
But…..
6 – VASA-1: So good, it’s bad 😬
Again, this is where we have to try and etch out a straight line in the blurry sand between AI Avatars and deepfakes.
There’s a reason that Microsoft isn’t yet (or maybe ever) releasing this publicly.
Because right now, an AI tech like this in the wild would create chaos.
(Especially before the U.S. right before an election cycle.)
Yes, there’s the potential for misinformation and disinformation. But, we’ve technically been living in that world for a few years now.
The average person (in the U.S., at least) finally knows that you can’t trust everything you read online or on social media.
And over the course of the last year or so, the average person kinda knows about AI photos and that not all photos are real. (Again, the term ‘Is that Photoshopped?’ has been locked in our lexicon for decades.)
But videos?
99.9% of people don’t even know this tech exists.
Or even the possibility of its existence.
Which is why, technically, VASA-1 is so good that it’s bad.
7 – VASA-1 prepares us for new normal 🧠
Did the world need deepfakes or digital twins?
Not necessarily.
But, they’re here.
If nothing else, VASA-1 is putting this technology on the map and in people’s minds.
Think of VASA as an important PSA for society, disguised as a research paper and shiny AI tech.
Whether we want it or not, we’re going to be bombarded with AI deepfakes soon.
So, we need to talk about it.
⌚
Numbers to watch
$3 Billion
valuation that Perplexity is fundraising at, according to reports, after fundraising at a $540 million valuation in January.
We wanna know below!
What are your thoughts on our Numbers to Watch section? |
Reply