✅DeepMind's AI Fact-Checks LLMs

Sponsored by

Reading time: 5 minutes  

Meet Ava, The World’s Best AI BDR

With a database of 250M+ leads, built-in email warmup, advanced lead research, and the ability to send 1000s of hyper-personalized emails, Ava provides end-to-end cold email automation. With plans starting from $225/mo and a 14-day hire trial, what are you waiting for?

Think you need a BDR? Hire Ava instead.

Today we will discuss-

  • 🔍Google develops AI fact checker for LLMs

  • 🚫OpenAI restricts release of voice cloning AI

  • 🚨Experts raise alarm over overvaluation in AI sector

  • ⚙️9 amazing AI tools you might not have heard of

All this and more - Let's dive in!

👩‍🍳What’s cooking in the newsroom?

Google DeepMind develops AI system that can fact-check responses from LLMs like ChatGPT 

☕News - A team of researchers from Google's DeepMind has developed an AI-based system named SAFE, designed to fact-check responses generated by large language models (LLMs) like ChatGPT.

🤔How does it work? When fact-checking LLM results, humans often rely on search engines like Google to find credible sources. DeepMind's team adopted a similar strategy for the development of their system, named Search-Augmented Factuality Evaluator (SAFE).

Their system uses an LLM to break down generated text into individual facts or claims. Next, it uses Google Search to identify potential verification sources. Finally, it compares the information from these sources with the original response to gauge accuracy. 

🧐So is the system any good? To find out, the research team tested it on roughly 16,000 facts extracted from responses generated by various LLMs. When compared to human fact-checkers, SAFE matched human findings 72% of the time. And interestingly, when there were disagreements between SAFE and humans, SAFE turned out to be correct about 76% of the time!

Another great thing about SAFE is its cost-effectiveness. The researchers found that using the AI system was about 20 times cheaper than relying on human fact-checkers. This is important because, looking ahead, as models generate an ever-growing volume of information, having an affordable and scalable method to verify claims will become increasingly crucial.

OpenAI previews voice cloning AI tool, but deems it too risky for public release

🗣️News - OpenAI has introduced Voice Engine, a new tool that can clone anyone's voice using only 15 seconds of recorded audio. However, the company has decided to limit its release to early testers, citing concerns regarding potential misuse. 

"We recognize that generating speech that resembles people’s voices has serious risks, which are especially top of mind in an election year," the San Francisco-based company said in a statement.

👨‍🔬How do they plan to ensure safe testing? The company plans to cautiously test Voice Engine by following a similar approach as it did with its text-to-video generator, Sora. In a blog post, OpenAI explained that Voice Engine generations are watermarked, allowing them to trace the origin of any generated audio. Additionally, it stated that early Voice Engine testers have agreed not to impersonate anyone without their consent and to disclose that the voices are AI-generated, ensuring responsible testing practices.

🛣️And what's the roadmap? A trademark application filed by OpenAI on 19 March suggests an interest in entering the speech recognition and digital voice assistant market. The company also mentioned in its blog post that it intends to initiate discussions on the responsible use of synthetic voices and societal adjustment to these innovations. Ultimately, they aim to use feedback from these discussions along with the outcomes of small-scale tests to inform their decision on whether and how to deploy this technology on a larger scale.

Experts foresee impending crash for overinflated AI companies

🫧News - Despite significant investment in the AI sector, analysts fear that many AI companies are overvalued, raising concerns about the emergence of an "AI bubble." (For context, a bubble refers to a period where prices of assets or investments skyrocket unrealistically, driven mainly by hype and speculation, rather than their true worth.)

In a recent research note, tech analyst Richard Windsor used an interesting analogy to explain the risks of the ongoing AI craze. He compared the current situation to a game of musical chairs, where money floods into the AI sector without proper checking of company basics, adding that this could spell trouble for investors when the music stops.

Furthermore, Windsor drew parallels between the AI hype today and past speculative bubbles, like the internet boom in 1999 and the enthusiasm for autonomous driving in 2017.

👨‍💼What's more? Other industry experts have also expressed similar concerns: Kai Wu, founder of Sparkline Capital, warns of a “huge boom” in AI investment, with some pursuing exposure at any cost while others anticipate a bubble burst. Emad Mostaque, ousted CEO of Stability AI, predicts that the AI bubble could surpass all others, dubbing it the “dot AI” bubble. Additionally, Jeffrey Gundlach, CEO of DoubleLine Capital, likens the current AI craze to the dot-com bubble, while John Hussman of Hussman Investment Trust warns of extreme speculation.

🙆🏻‍♀️What else is happening?

🌞Interesting facts

  • According to a study by IBM, bots are able to address 80% of regular user inquiries, resulting in a reduction of more than 30% in customer support expenses.

  • In 2018, a machine in Microsoft's translation system achieved human-level accuracy when translating a Chinese news article into English. The accuracy was confirmed by external human evaluators who compared the machine's output to two independent human translations.

  • AI is getting smarter and could outsmart humans in the future. With its ability to learn, AI is expected to reach the intelligence level of adults by 2029, a significant advancement from its 2013 status, which was comparable to a 4-year-old's intelligence.

👩🏼‍🚒Discover mind-blowing AI tools

  1. Elai.io - A platform that allows users to generate AI videos from text without the need for a camera or studio ($29/month)

  2. Undetectable AI - A tool designed to transform AI-generated content into human-like text that cannot be detected by AI ($14.99/month)

  3. Zyng - A platform that allows users to instantly create amazing product posts by uploading product images ($0.05/image)

  4. Beatoven.ai - A tool that helps users compose unique mood-based music to suit every part of their video or podcast ($3.59/month)

  5. Final Round - An AI-powered interview preparation tool that helps job seekers ace their interviews ($99/month)

  6. Teach-O-Matic - A tool that helps users create how-to videos from text instructions (Open Source)

  7. Securewoof - An AI-powered malware scanner that checks executable files for maliciousness (Free) 

  8. 60sec.site - A no-code website builder that allows users to create a landing page in just 60 seconds ($10/month)

  9. PDF Candy - Allows users to convert files to and from PDF format ($3.60/month)


Myth - AI understands context perfectly

Fact - AI does not understand context perfectly. Despite its capabilities, AI struggles with nuanced context, including humor and cultural differences. It may misinterpret information due to limited inference, changing context, emotional subtleties, and unstructured data challenges. Its lack of common sense further hampers comprehensive contextual understanding.

Can you help us, please?

Be our mentor and let us know your take on today's newsletter. Share your honest and detailed thoughts on how we can improve OpenTools for you or just drop a Hi. YES! We personally read all the messages.

Login or Subscribe to participate in polls.