Welcome to the 610 new readers to AI Chatter!

We’re AI Chatter - the newsletter that's like an encyclopedia, but with more jokes and less dust 📖

Here’s what you missed:

• Deepfakes and Politics don’t mix

• First Big Tech Leak

• AI reads minds

AI can’t kill anything worth preserving.

Are you ready for the next political scandal? Hint: it might be fake

On a recent episode of Pod Save America, a cohost shared an exclusive audio message from President Joe Biden that sounded oh-so-real. But surprise, surprise, it was actually AI-generated audio! And as AI technology gets better at mimicking human speech, this could spell trouble for political media.

Fake videos of politicians, actors, and podcasters have already gone viral, from Joe Rogan debating Ratatouille with Ben Shapiro to Emma Watson “reading” Mein Kampf.

@iktaco

presidential deepfake pro clubs edition #fifa #proclubs #deepfake #aivoice #biden #president

But it's not all fun and games. Deepfakes could easily be used to spread misinformation, like a fake video of Biden criticizing transgender women.

While some deepfakes can be easily debunked, the potential for more convincing deepfakes is cause for concern. Major news outlets and tech giants will have to figure out how to deal with this new challenge in political media before the next presidential campaign.

And with AI technology becoming more accessible, newsrooms will have to learn how to vet stories faster than ever before. So, buckle up AI enthusiasts, because it's going to get better and scarier very fast.

Work Smarter, Not Harder

• 1/ b7labs.co - generate summary of webpage (Free)

• 2/ Devi - monitor social media brand mentions and keywords (Trial)

• 3/ Maskr.ai - picture of yourself to replace with any celebrity of your choice (Trial)

• 4/ everylead.ai - customized ice breakers based on Linkedin bio (Freemium) 

• 5/ ingestai.io - turn your knowledge base into a chatbot-like assistant (Freemium)

4chan Leaks Meta Models

You won't believe what happened last week when Meta's latest large language model (LLM), LLaMA, got leaked! It turns out that a cheeky user named "llamanon" on 4chan's technology board decided to release LLaMA's 7B and 65B models via torrent. And just like that, the Internet had access to this shiny new LLM!

But that's not all, folks. This leaker even added a pull request to the LLaMA GitHub page under the title "save bandwidth by using a torrent to distribute more efficiently," throwing shade at Meta's process of providing access to the bot. Oh snap!

Despite this blunder, the leak was received positively by 4chan users, with AI researchers all over the world excitedly testing out LLaMA. However, they soon discovered one of the biggest issues with the model - its sheer size. We're talking 30 GB of GPU memory to run the smallest model! But not to worry, they were able to optimize it and even run the 7B version on a Ryzen 7900X CPU.

But hold up, is LLaMA even that good? Initial tests show that it may not be as polished as ChatGPT, but let's not forget that ChatGPT is a chatbot built on a fine-tuned LLM. And let's be honest, asking an LLM the meaning of life is like asking your drunk friend at 2am. The responses are all over the place!

However, LLaMA's benchmarks show that even the smallest version of this agent was able to beat GPT-3. This means that the future of LLaMA is looking pretty bright, especially now that it's in the hands of the AI community. Who knows what exciting discoveries will be made with this shiny new toy?

Penny for Your Thoughts

• Publisher Takes a Stand on AI in Scientific Writing

  Springer Nature, the bigwigs of academic publishing, have spoken out about the use of AI writing tools in scientific papers. In short, they've made it clear that software like ChatGPT can't be credited as an author in any papers published in their countless journals. But don't fret! Springer is okay with scientists using AI to help write or come up with ideas for research. Phew! Just make sure to credit it properly, folks.

• YouTube's New Head Honcho Teases

  The new head honcho of the platform, Neal Mohan, has laid out his plans and teased some upcoming features in his first address since taking the helm last month. In his letter, Mohan let slip that YouTube is developing generative AI tools for content creators that will allow them to create outfits and fantastical film settings. We're talking mind-blowing AI wizardry, folks! Details are still hush-hush, but we can expect the rollout of these generative AI tools in the coming months.

• LLMs and Math: A Tale of Love and Disappointment

  If you think AI systems like ChatGPT are good at everything, think again! While they can produce fluent text, they often struggle when it comes to solving mathematical problems that require reasoning. For instance, when asked about a secondary-school level math problem, OpenAI's ChatGPT scored a measly 26%. It's no surprise that they fared poorly on mathematical tasks since they're just trained to recognize patterns in the language. However, Google's LLM called Minerva has surprised researchers with a score of 50% on the same MATH data set questions. Looks like we have a new math whiz in town!