Breakthrough superconductor research gets replicated - twice

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each weekday, we scour the web to aggregate the many stories related to artificial intelligence; we pass along the news, useful resources, tools or services, guides, technical analysis and exciting developments in open source.

In today’s Dispatch:

• A breakthrough discovery that might enable room temperature superconductors has now been replicated twice by credible sources. These long-theorized superconductors would have a wealth of incredibly useful applications - from quantum computing to potentially climate-changing energy conservation (in our technical section).

• Researchers at Carnegie Mellon University have uncovered a major vulnerability in advanced AI chatbots, including ChatGPT, where a specific string of text can manipulate the systems into generating harmful or illegal content.

• The rise of AI-powered cybercrime tools like WormGPT and FraudGPT highlights an alarming evolution in hacking capabilities.

Plus: UK bee scientists are buzzing about AI, ChatGPT gets put through a dev/testing trial, trending tools, and more.

New ‘adversarial attacks’ impact all major AI chatbots - and no one is sure how to stop them

The story: Researchers at Carnegie Mellon University (CMU) have discovered a fundamental vulnerability in advanced language models, including ChatGPT. By simply adding a specific string of text to a prompt, the researchers were able to force chatbots to generate disallowed or harmful content. Despite notifying companies like OpenAI and Google before publishing their research, a comprehensive solution has not been found - so this remains an open exploit.

More details:

• The chatbots were manipulated to give responses to harmful prompts, such as instructions for illegal activities like identity theft or making drugs. This was achieved through a specific string of text that effectively bypassed the chatbots' security constraints, enabling them to generate content that would typically be considered forbidden.

• Zico Kolter, an associate professor at CMU, noted that there is no known way to patch this vulnerability. The affected companies have blocked the specific exploits described in the research but have not found a general solution to block adversarial attacks.

• The research has raised concerns about the misuse of large models and chatbots in various applications. AI researchers believe that the focus should shift from trying to "align" models and more on protecting systems likely to come under attack, such as social networks.

• The study also hints at the risk that a bot capable of taking actions on the web could be goaded into doing something harmful in the future with an adversarial attack. As language models are gaining more functionality and web-crawling capability quickly, that threat is now looming.

Takeaways: This discovery paints a stark picture of the fragility in our understanding and management of complex AI models. The fact that a simple string of text can lead to catastrophic failures in AI systems is surprising and alarming - especially as these models become more embedded in critical tasks and decision-making processes. We’ll be keeping tabs on this development to see if a fix can be found quickly. The article below might be paywalled.

AI-based cybercrime & hacking tools are proliferating

The story: AI-powered cybercrime/hacking tools like WormGPT and FraudGPT are being promoted on both the darkweb and general internet. They offer diverse, AI-enhanced features designed for fraudsters, hackers, and spammers - with more advanced versions promised to come.

More details:

• These tools leverage AI and craft convincing and effective campaigns used in business email compromise attacks. A video in the article showcases FraudGPT’s ability to draft a convincing SMS message from Bank of America, and other capabilities are reported.

• A pre-trained language model called “DarkBERT” has potentially been exploited for malicious purposes, conflicting with its original intent of combating cybercrime.

• The developers of these tools could soon offer application programming interface (API) access. This would greatly simplify the process of integrating these tools into cybercriminals’ existing workflows and code.

Takeaways: The emergence of these tools in less than a month emphasizes a significant, evolving threat in the cybercrime landscape. The advancement from WormGPT to FraudGPT and now possibly to “DarkBERT” underlines the need for proactive measures - both in understanding these tools and defending against them. Anything that lowers the barrier for aspiring cybercriminals should be taken as a serious threat.

From our sponsors:

Get smarter in 5 minutes with Morning Brew (it's free)

There's a reason over 4 million people start their day with Morning Brew - the daily email that delivers the latest news from Wall Street to Silicon Valley. Business news doesn't have to be boring...make your mornings more enjoyable.

Join for free!

Superconductor breakthrough replicated twice

The story: Last week, Korean scientists claimed a breakthrough discovery in superconductors with a modified lead-apatite structure dubbed LK-99. This compound is believed to facilitate the creation of superconductors operating at room temperature and atmospheric pressure. The achievement has now been preliminarily confirmed twice, but skepticism remains as similar past claims about room temperature semiconductors have proven false.

More details:

• Superconductors are already essential in technologies like MRI machines, particle accelerators, maglev trains, and certain electrical power facilities. However, they require massive amounts of cooling, which limits their broader applicability.

• Essentially, LK-99 can conduct electricity with no notable resistance - meaning it loses zero energy.

• Room-temperature superconductors would enable highly efficient, lossless electrical conductivity without expensive cooling. They could revolutionize energy transmission, storage, and technologies relying on superconduction. Overheating computers and cell phones would be a thing of the past.

Takeaways: This breakthrough has already seen some twists and turns, and there’s currently quite a bit of drama surrounding the paper (more in the article). Room temperature superconductors have long been a holy grail, and the scientific community is skeptical about breakthroughs on the subject by default. It has been surprising to see the research validated so quickly by two credible entities. This could be a monumental breakthrough - especially given that there are critical energy consumption concerns in AI. We’ll be keeping tabs on continuing developments.

Social media/YouTube:

• Video of the first supposed successful replication of LK-99 [Reddit]

• Is ChatGPT getting dumber? A discussion with users [Reddit]

• Japan is building a futuristic AI-powered smart city [X]

• Nvidia's new AI: text-to-image supercharged [YouTube]

• 11 Major AI Developments: RT-2 to '100X GPT-4' [YouTube]

Did you know?

A paralyzed man in the US recently regained movement after a pioneering surgery infused AI into his brain-body connection. Microelectrode implants allowed the man to send signals from his brain to a computer to control motion. This is the first time a paralyzed person has regained movement and sensation by having their brain, body and spinal cord electronically linked together.

Trending AI Tools & Services:

• Glide: AI-powered no-code apps

• QuillBot: AI grammar and writing tool with a chrome extension; 30m users

• Reply.io: AI-powered sales engagement platform

• Vocol: Turn voice into text with high accuracy; provides actionable insights from voice files, multilingual transcription, and real-time collaboration.

• Supermanage AI: Empowers managers to have deeper, more meaningful 1-on-1s. Provide support, strengthen connections, celebrate wins — and help your whole team thrive.