- AI Tangle
- Posts
- ☕️ Is Microsoft's Recall a Ticking Time Bomb?
☕️ Is Microsoft's Recall a Ticking Time Bomb?
Microsoft's Recall can't seem to catch a break from scrutiny, now facing cybersecurity-related backlash before it's even released. Elsewhere, Google reflects a bit on the AI Overviews feature mess with some improvements to follow, Hugging Face engineers deal with a breach on its AI model hosting platform, and the developer-loved Raspberry Pi teams up with Hailo to bring AI to its latest microcomputer. Join us at AI Tangle as we untangle this week's happenings in AI.
THE BIG AI STORY
Microsoft's recent push towards the "AI PC" led them to showcase Recall a few weeks ago, an AI-based Windows feature that screenshots everything on the computer for users to backtrack later. Though the feature was heavily scrutinized for being a "privacy nightmare" already for its ludicrous concept, it's started to recently surface that it's not just a privacy problem but cybersecurity, too. Privacy concerns aside, the one thing Microsoft guaranteed to media outlets was that a hacker wouldn't be able to exfiltrate Recall activity remotely, and even that seems to be bogus now.
What's the issue with Recall?
Security researcher Kevin Beaumont scrutinized Microsoft's claims to media outlets and its attitude toward cybersecurity in a lengthy X/Twitter thread. Beaumont details that breaking into the local SQLite database, where Recall stores everything on your computer, is laughably easy and that it's also not encrypted, meaning its data gets stored in plain text. Beaumont also points out that InfoStealer trojans are widespread and that Recall would be an easy target for attackers. The thread turned so lengthy that Beaumont turned it into an article, where he answers the many security-related questions about Recall and debunks them.
COLLABORATING WITH TWILIO
What can good data do for you?
Twilio Segment helps every team access good data that's real-time, clean, and accurate. The result? Relevant customer experiences that drive real revenue.
Segment helps 25,000+ companies turn customer data into tailored experiences. With customer profiles that update in real-time, and best-in-class privacy features - Segment's Customer Data Platform allows you to make good data available to every team.
6 QUICK HITS
At the Computex expo in Taipei, Taiwan, Nvidia CEO Jensen Huang announced an array of AI products and partnerships, emphasizing the shift from traditional data centers to AI factories. He unveiled Nvidia ACE generative AI, which is capable of creating lifelike human avatars, an "ultra" version for its Blackwell platform in 2025, and teased a next-gen GPU architecture codenamed Rubin. He also highlighted Taiwan's crucial role in AI infrastructure and how companies like Foxconn and Siemens use Nvidia's platforms for AI-powered autonomous robots.
Also at Computex was AMD, where the company unveiled its Ryzen 9000-series desktop CPUs powered by the all-new Zen 5 architecture. Promising a 15% performance uplift over last-gen, the 9000-series boasts doubled instruction bandwidth, improved AI acceleration, higher IPC across apps, and efficiency improvements for less power consumption. AMD aims to launch these in July on existing AM5 motherboards with DDR5/PCIe 5.0 support and thus outpacing its rival, Intel, who's now on the backfoot to release something later this year.
Google admitted in a recent blog post that its new AI-powered search feature, debuted at the Google I/O 2024, produced "odd, inaccurate or unhelpful" results after facing backlash over comically bad outputs. Examples include anywhere from the classic "put glue on your pizza" to "eat literal rocks to stay healthy." This behavior stems from AI Overviews misinterpreting queries, misunderstanding web content, and being overly reliant on user-generated data on forums like Reddit, despite the company downplaying the latter. However, Google has been at work on it, as the same blog post details numerous improvements and fixes made to the feature.
On an unsuspecting late afternoon on Friday, AI startup Hugging Face, known for its extensive platform for building, training, and deploying large language models, detected a breach in its AI model hosting platform, Spaces. The intrusion potentially compromised private keys and tokens used to unlock protected resources, with no disclosed number of affected users to speak of. The incident caused Hugging Face to thoroughly review its security policies and procedures amid scrutiny.
Microsoft's latest plans include investing a whopping $3.2 billion in AI and cloud facilities in Sweden, leveraging the Nordic country's abundant green energy to put it on the global map of AI. The investment includes adding 20,000 GPUs at its data centers and training 250,000 people in "essential AI skills." Microsoft's Brad Smith highlighted Sweden's potential benefits from AI and the company's commitment to renewable energy, including a contract with state-owned power company Vattenfall AB.
Raspberry Pi, the developer-loved microcomputer-building company, recently announced that it, too, will get on board the AI hype train by partnering up with Hailo. The duo is working on an AI Kit, an add-on for the Raspberry Pi 5 microcomputer that will run the Hailo-8L M.2 accelerator. Heila CEO and co-founder Orr Danon advocates for its power efficiency and passive cooling for a reasonable 13 tera operations per second (TOPS). The kits will be available "soon from the worldwide network of Raspberry Pi-approved resellers" for $70.
4 AI TOOLS
MarsCode - MarsCode is a built-in AI programming assistant with capabilities like code completion, explanation, debugging, and AI Q&A for faster development at scale.
Spellar - Spellar is an AI-driven speech assistant that gives personalized feedback to enhance your speaking skills and refine your style.
Jovu - Go from idea to production-ready code in four minutes with Jovu, powered by Amplication's AI to effortlessly build new services and extend existing ones.
timeOS - timeOS is an AI productivity companion that captures and summarizes your day, organizes relevant information within the right tool, and proactively surfaces the knowledge when you need it.
AI READ & WATCH
Artificial Intelligence - What is it? (10-min read)
With artificial intelligence becoming as widespread as it is, how many people understand what it is? TechCrunch's Devin Coldeway aims to take a deeper dive, covering the concepts behind chatbots and image generators, the risks of hallucinations and bias, and the ever-elusive quest to achieve artificial general intelligence.
Has Generative AI Already Peaked? (13-min watch)
A recent paper suggests that bigger and more complex generative AI models have more and more diminishing returns. Mike Pound, a researcher at the University of Nottingham, chimes in on that in a video on Computerphile.