☕️ Did DeepSeek Just Take The Western AI World by Storm?

Hailing from China, DeepSeek takes the Western tech world by storm and alarm with its latest model, causing massive disruption in the stock market, but could it be a blessing in disguise? Other key highlights of the week include:

• Perplexity updates its proposal to merge with Tiktok US to instead become a new holding company called NewCo

• Grok 3 was spotted in the wild as users on X/Twitter got access to it for a short while before being revoked

• Apple turns its AI suite on by default and temporarily axes AI news summaries in its latest batch of software updates

Join us at AI Tangle as we untangle this week's happenings in AI!

DeepSeek shocks the tech stock market as its DeepSeek-R1 model sows doubt in western AI spending

Tech and chip stocks from Nvidia to Microsoft dropped sharply on Monday as Chinese AI startup DeepSeek's latest open-source AI model, DeepSeek-R1, not only challenged US dominance in AI but also put its enormous investments under the spotlight. DeepSeek officials say that the startup trained its models with way fewer Nvidia chips than industry leaders while beating out all of them and matching OpenAI's o1. However, though it could be seen as a disruption, many analysts claim this could be a net positive for chipmakers in the long run if DeepSeek's approach is something to learn from.

​

​And it just recently released a new family of image models on top​

If crashing Wall Street's AI market wasn't enough, DeepSeek also unveiled a family of "novel autoregressive framework" image models, called Janus-Pro, that it claims can outperform OpenAI's DALL-E 3. The models are available on Hugging Face for anyone to download and range from 1B to 7B parameter variants, which roughly corresponds to a model's problem-solving skills. All of this had led to OpenAI losing its position as the #1 most downloaded app on Apple's App Store to DeepSeek, but the widespread attention has made the company fall victim to "large-scale malicious attacks," reportedly. As a result, DeepSeek has temporarily hit the breaks on user registrations, though existing users can continue to use its services as normal.

Perplexity revises its proposal to merge with Tiktok US in a bid to ByteDance

AI-based search company Perplexity on Sunday updated its proposal to merge with TikTok US to form a new holding company, "NewCo," in which the US government would own up to 50% after a future IPO with a $300 billion valuation. The proposal indicates that ByteDance would sell TikTok US to investors while retaining its core recommendation algorithm. Perplexity stated it would also be willing to be acquired by the new entity if its investors received NewCo equity. As part of a temporary 30-day measure, TikTok services were recently restored after legal action shut them down in the US.

​

Some users on X/Twitter inexplicably got access to xAI's Grok 3 AI model for a few hours

Elon Musk’s AI company xAI may soon release its latest model, Grok 3, which has reportedly been accessed, be it by accident or not, by some users on X/Twitter, including popular reverse engineer Alexey Shabanov. Early tests revealed its ability to tackle logical reasoning, riddles, and coding, though it seems to do minor slip-ups in the latter. Not more than 2-3 hours later, access was revoked. Grok 3 was trained with 10x more compute than its predecessor, according to Musk, at its giant Memphis data center and leaks indicate it may debut with a voice mode.

​

Apple's recent iOS 18.3 updates turn Apple Intelligence on by default as it nears a more complete rollout

Whether users like it or not, in Apple's recent iOS 18.3, iPadOS 18.3, and macOS Sequoia 15.3 software updates, the company has turned its Apple Intelligence suite of AI features on by default, as stated in developer release notes. Apple's infamous AI-powered news summary push notifications, meanwhile, have been turned off for constantly twisting its news notifications to display inaccurate facts, brought out into the limelight by the BBC. Apple Intelligence's rollout has been slow as Apple tries to ease its users into it, though users still uninterested in the suite retain the option to turn it off in the Settings app.

​

Security researchers discover a critical security vulnerability in Meta's Llama AI Stack framework

Researchers recently discovered a critical vulnerability (CVE-2024-50050) in Meta's Llama Stack AI framework that could allow attackers to remotely execute malicious code. The vulnerability, which poses significant risks such as data breaches or theft of computing resources, stems from the misuse of the PyZMQ library and was rated 9.3/10 in severity by experts but only 6.3/10 by Meta itself. However, according to the researchers, Meta was quick to patch the issue in version 0.0.41, and PyZMQ maintainers updated their documentation to prevent future occurrences.

​

Meta's Zuckerberg states the company plans to invest upwards of $65 billion in AI infrastructure in 2025

Calling it "a defining year for AI," Meta CEO Mark Zuckerberg posted on Facebook to announce Meta's plans to invest up to $65 billion in 2025 to expand its AI infrastructure. On top of that, Meta also wants to construct a more-than 2-gigawatt data center in Manhattan, ramp up hiring for AI-centric roles, and end the year with over 1.3 million Nvidia GPUs in its possession. Zuckerberg aims to make Meta AI the market's leading assistant and Llama 4 the leading state-of-the-art model, along with developing an "AI engineer" that could start making a difference in its codebases.

​Tellers - Tellers is a powerful AI video editing platform that accelerates production by taking care of the tedious, AI-handleable work while maintaining excellent quality, leaving just corrections for editors.

​Supernormal - Spend less time writing, polishing, and sharing notes and more time on the work that matters with the help of Supernormal's AI-powered meeting transcriptions.

​Dicte - Dicte is an AI-powered meeting assistant that turns voice notes and recorded meetings into reports and minutes automatically with seamless integration.

AI-Developed Drugs Are Coming Full Circle (​3-min read​)

Aside from coding, one area that AI has been increasingly more prominent is in healthcare, and at the World Economic Forum, Google DeepMind CEO Demis Hassabis talked in length about his enthusiasm, expecting AI-designed pharmaceuticals to be in clinical trials as early as the end of 2025.

How Far Away Really is Human-Level Intelligence? (​3-min read​)

Whether AI is overhyped or underhyped by one public figure or another, AI is advancing at rapid speeds. Although he's not calling it AGI, Anthropic CEO Dario Amodei believes the technology could overtake human capabilities not long after 2027.