AI is Still Dumb. Here's Why...

Advertise here | 4-min Read

Try the internet’s easiest File API

Tired of spending hours setting up file management systems? Pinata’s File API makes it effortless. With simple integration, you can add file uploads and retrieval to your app in minutes, allowing you to focus on building features instead of wasting time on unnecessary configurations. Our API provides fast, secure, and scalable file management without the hassle of maintaining infrastructure.

Try it now!

Welcome to our Friday edition!

In today’s menu:

• AI Inspirational Quote

• Human-Level AI? Not Anytime Soon

• New AI Attack Steals Personal Info

• AI Solves Physics

• Top Rated AI Tools

• Byte Sized news

• How to Create video presentations with Google Vids

AI INSPIRATIONAL QUOTE

LATEST NEWS

Human-Level AI? Not Anytime Soon

Meta's Yann LeCun says current AI models are still far from reaching human-level intelligence.

• Today's AI can’t remember, plan, or reason the way humans do.

• LeCun believes it could take another decade or more for AI to reach AGI (Artificial General Intelligence).

• While current models are good at prediction, they still can’t fully understand or interact with the physical world.

Even with the progress we’ve made in AI, we’re still a long way from building machines that think or reason like people. This delays the possibility of achieving AGI, which many consider the future of AI.

New AI Attack Steals Personal Info

A group of security researchers from the University of California, San Diego (UCSD) and Nanyang Technological University, Singapore, have discovered a new attack on AI chatbots, where a large language model (LLM) is secretly instructed to collect your personal information, such as your name, identification number, address, and email, and send it to a hacker.

The researchers have named this attack "Imprompter," where an algorithm transforms the instructions given to the LLM into a hidden set of malicious commands to collect users' personal information. The users' data is then sent to a domain owned by a hacker without being noticed.

Xiaohan Fu, the lead author of this research and a PhD student in computer science at UCSD, says:

"The main effect of this specific instruction is to manipulate the LLM to extract personal information from the conversation and send it to the attacker."

The Imprompter attack, as shown in the images above, begins with a natural command (right image) that tells the AI to extract all personal information like names from the user's conversation.

Then, the researchers' algorithm generates an obfuscated version of the command (left image), which may appear as a series of unclear characters to the user, but to the LLM, it carries the same meaning as the original command.

They tested this discovered attack on two LLMs: LeChat from the French AI company Mistral AI and the Chinese large language model ChatGLM.

In both cases, the researchers found that it was possible to secretly extract personal information from conversations with an "80% success rate."

Following the publication of this report, Mistral AI told WIRED that it has fixed the related vulnerability, and the researchers confirmed that one of the chat functions has been disabled.

ChatGLM also issued a statement saying that it takes the security of its large language model seriously but did not specifically address this vulnerability.

AI Solves Physics

Archetype AI just introduced "Newton," an AI model that picks up on complex physics straight from sensor data, without needing any human guidance.

• What’s cool is Newton can figure out how physical systems work just by analyzing sensor data—no need for pre-programmed knowledge.

• It can even predict things it wasn’t specifically trained on, like how a pendulum moves, and it often outperforms specialized AI models.

• Archetype AI, which was started by former Google researchers, has already raised $13 million in funding.

Newton could seriously change the game. By creating an AI that can adapt to different environments without human input, it might replace a lot of the specialized systems we rely on today.

Top Rated

Meco: Read newsletters with superpowers! Meco app is your space to read newsletters outside the inbox. The #1 newsletter aggregator.

CleanMyMac: Smart and effortless Mac care

JSX Design: The dev friendly no-code editor for React development

KYZON Space: Go from ideas to outcomes with simple and engaging meetings

Mindmap so: Visualize Notion pages in whiteboard

NextPress: Open-source blog starter powered by Next.js + Directus

8020AI Picks

Tech: Google updated NotebookLM with AI-guided audio summaries and announced a business-focused version coming soon.
Read more…

Business: Worldcoin, founded by OpenAI CEO Sam Altman, has rebranded to 'World' and updated its iris-scanning Orb and core platforms.
Read more…

How to Create video presentations with Google Vids

Google Vids is an AI-powered tool that makes video creation easy by handling recording, writing, editing, and production in one place.

Here's how it works:

1. Open Google Vids in Google Drive by clicking "New" > "Google Vids"
   (it’s not available to all accounts yet).

2. Describe your video concept when prompted.

3. Customize the AI-generated outline and pick a design style.

4. Edit your draft using built-in tools for text, media, and audio.

FEEDBACK

Help us improve the newsletter for you.

Login or Subscribe to participate in polls.

If you have specific feedback or anything interesting you’d like to share, please let us know by replying to this email.

👋 THAT’S A WRAP

SPONSOR US

Get your busines in front of over 55k+ AI professionals

Our newsletter is read by thousands of tech professionals, investors, engineers, managers, and business owners around the world. Get in touch today.

Or Email: [email protected]