This was another incredible week of AI news. Meta continues to dominate open source, AI can now use tools, Google puts ai in robots, and it sure looks a lot like AGI, multiple companies are experimenting with AI features, and the dark web produces another worrisome AI model. And for our main story, we have potentially one of the most significant scientific breakthroughs of this generation.
Meta Dominates AI
Meta has been on an absolute tear lately, publishing open-source AI at an insane rate. Just last week, they released LLaMA 2, and now they’ve come this week with a vengeance.
Meta Chatbots
In what is likely their first entry into leveraging LLaMA 2 in their own products, Meta is reported to be releasing AI-powered chatbots on their social media platforms as early as next month. These chatbots will take on different personas, and according to TechCrunch, some examples are one that advises users on travel plans in the style of a surfer and another that speaks like Abraham Lincoln. Meta views this as a way to boost engagement across their suite of products.
Uncensored LLaMA 2
As many had hoped, LLama 2 now has multiple uncensored versions. But uncensored is much more than just being able to ask how to break into a car. Uncensored models perform better in many cases. Let’s take a look at some examples from ollama.ai:
So what you’re seeing here is beyond censorship, this is LLaMA being completely lobotomized.
LLaMA Drama
Llama llamas got some drama. An open-source watchdog says Llama 2 is not open-source. The open-source initiative has published a blog post accusing Meta of incorrectly labeling llama 2 as open-source even though it does not meet the requirements of the open-source definition. They specifically called out the fact that Meta has restricted the use of the llama 2 models for companies they see as competition, by requiring permission to use llama 2 for apps with more than 700m users. Of course, there aren’t many apps out there with that many users, but that restriction clearly negates the premise of true open-source.
AI Music
Next, Meta has open-sourced Audiocraft, which is generative AI for audio. This technology is awe-inspiring. Type a few words into a prompt, and incredible, AI-generated music comes out. According to Meta: this can be used for “A professional musician able to explore new compositions without playing a single note on an instrument. Or an indie game developer populating virtual worlds with realistic sound effects and ambient noise on a shoestring budget. Or a small business owner adding a soundtrack to their latest Instagram post with ease.” I’ve already made a video on using these tools, so check them out.
AI Labeling
It was reported that Meta will release tools to help identify and label ai-generated content on its platforms. They’re also releasing a “summarize” feature for their messenger app for those friends who always send you walls of text.
Catalyst Demo
Meta has released Catalyst Demo, which uses AI to help researchers accelerate work in material sciences by enabling them to simulate the reactivity of catalyst materials ~1000x faster than existing computational methods using AI. Even Elon Musk was excited about this.
YouTube Summaries
Our next set of stories is about companies embracing AI in different ways. First, Youtube is experimenting with an AI tool to summarize videos. These summaries allow users to decide if a video is worth their time to watch. But I see a few issues with this. First, youtube made it clear that this will not replace video descriptions, which creators write. Also, there’s a delicate balance between giving enough information upfront and a viewer not wanting actually to watch a video.
Stackoverflow AI
Stackoverflow has been in the news lately because of AI, but not in a good way. It has been widely reported that engineers use the question-and-answer service much less often and instead, go to ChatGPT. I can tell you firsthand that when I’m coding, the first place I now go with errors or bugs is not Stackoverflow, but instead ChatGPT. It’s just so easy to get a great answer immediately. Don’t get me wrong, I love Stackoverflow. It has provided me with so much value over the years. So I was thrilled to see them embracing AI and launching a new product called OverflowAI. Overflow AI is their vision of the future of Stackoverflow.
Stackoverflow is leveraging its incredibly valuable dataset to answer your programming questions instantly.
FraudGPT
Following WormGPT and PoisonGPT, a new cybercrime generative AI tool is called FraudGPT. "This is an AI bot, exclusively targeted for offensive purposes, such as crafting spear phishing emails, creating cracking tools, carding, etc.," Netenrich security researcher Rakesh Krishnan said in a report published Tuesday. For a subscription cost of $200/mo, you too, can have FraudGPT. I find it hilarious that saas is still the desired business model, even on the dark web.
AI Influencers
Meet the new batch of influencers with flawless bodies, unattainable lifestyles, and mega deals from fashion giants. The catch? They are 100% AI-generated. In what feels like something straight out of blade runner, many AI influencers are amassing huge followings and landing REAL sponsorship deals with huge brands. Influencers who faked private jet photos and used copious amounts of photoshop were already bad for people’s mental health. Now, AI-generated influencers will worsen things because there are no limitations. Do you agree? Am I being old and out of touch? It seems like Meta’s ai labeling tech would be helpful here.
AI Using Tools
In a massive step towards AGI, AI models can now use tools in the form of API calls. I reviewed one of them this week called Gorilla, but multiple others have been released recently. Using tools isn’t new. Chatgpt has had plugins for months now. But what is new is the ability of these models to figure out how to use tools that weren’t pre-defined for them. Not only that, but since tools change all the time, they also can check for changes to APIs at run time. This leads to a world where we have verticalized models that are good at specific things and a generalized model that can call for help any time it needs. This is almost certainly what AGI will look like.
AI Video Of The Week
Each week, I will share my favorite AI-generated video of the week, and this week is a doozy. Have you heard of power-slap? It’s the game you play with your drunk friends to see who is the toughest…and dumbest, except it’s now professional and owned by the UFC. Two people enter, take turns slapping the $hit out of each other, and whoever gives up or passes out first loses. But this week, an AI robot entered the competition in what is bound to be made illegal.
Ok, so that wasn’t real. Someone took a video and used WonderDynamics to turn one of the…slappers…into a robot. It does look real though. I’m opening submissions for the ai video of the week, so hop in my discord, and there’s a channel called “ai video of the week”. If your video is chosen, you’ll get a call out, and in the future, I may give away some prizes!
RT-2
Regarding robots, Google released a new project called RT-2, which uses AI models to translate “abstract concepts learned from vast web data into actionable knowledge that informs robot behavior.” That means it allows robots to look at their surroundings and use large language models to understand how to interact with their environment. Microsoft released a paper months ago on the same topic, you can find my video review below, but RT-2 takes it to another level. One way RT-2 excels is by understanding tasks but also being able to generalize those tasks. It also exhibits the ability to use more complex prompting, such as chain-of-thought reasoning. And Google’s techniques have shown emergent capabilities. Very exciting, and of course, I’m here to serve our robot overlords.
Superconductivity (LK-99)
Our main story today is not related to AI but is too exciting to not talk about. In what is potentially one of our generation's most significant scientific discoveries, researchers from Korea University have discovered a material that can achieve room-temperature super-conductivity named LK-99. Although the findings are still being peer-reviewed, there has been a wave of excitement throughout the scientific community. Some research teams have validated portions of the claims made by the original researchers. So…why is this so important? Imagine your phone’s battery lasting 10 times longer, electric cars will go significantly further, and computer chips will operate with essentially zero heat, and these are just the beginning. Some even think this could be what powers spacecraft in the future.
The gist of how it works is that it gets hot when electricity passes through a material. That heat is a significant limit of performance. With superconductive materials, that heat generation essentially drops to zero. So you can run vast amounts of electricity through it without heat and efficiency loss. I can’t speak much to the science, but I encourage you to check out other YouTubers who know more about how this works and what it means.
Once again, all of these futuristic technologies are merging, and I’m here for it. I’ll keep bringing you the most exciting AI and technology news weekly.
