In our last newsletter, we spoke about the latest and shiniest large language model (LLM) on the block, Meta’s Llama 3.1. Soon after we published the newsletter, France’s Mistral AI came out with its own rival model–Mistral Large 2.

Calling it ‘large enough,’ Mistral claimed in a blog post that their new model surpasses Llama 3.1 in some benchmarks despite being trained on fewer parameters. Excelling in code generation, mathematics, and multilingual support, Mistral Large 2 aims to reduce hallucinations and provide more precise answers.

Previously, the model was kept proprietary behind their API, but now they are opening the gates to the Large 2 model. While Llama 3.1 is open source, Mistral Large 2 is not fully open source and requires a paid license for commercial use.

Speaking of open-sourced models (or lack thereof), there was speculation back when OpenAI announced GPT-4o that we might be getting our hands on a search engine from the AI company, and here we are two months later. 

OpenAI’s search engine SearchGPT is finally on the market. Well, sort of.

Just like it did with Sora, announcing a product to create much hype around it before they publicly release it, OpenAI hasn’t made SearchGPT open to the public yet. However, people can sign up for a trial round.

OpenAI also announced that it plans to roll these features into ChatGPT eventually. Plus, they're partnering with publishers to make sure sources are cited and linked properly, as is done by Perplexity AI and Google AI Overviews (only available in the US).

All eyes are on OpenAI to see if SearchGPT delivers better results than Google’s AI-driven summaries, which have had a rocky start. Will this be the next big thing in search? It’s unlikely since Google enjoys over 90% of the market share and it is less likely to roll over and give its number 1 position to someone else unless the product really knocks our socks off.

And also, there have been reports of SearchGPT making mistakes already.

Speaking of mistakes, haven’t we all (at a low point in our life) turned to Google and sought out Reddit's genuine human responses for some authentic, relatable answers? ‘What REALLY happens if you swallow your gum reddit’ or ‘Would you rather be the best in the world at one skill and below average in everything else reddit.’

C’mon, we have all been there. Well, now Google has snagged all the Reddit goodies, leaving the rest of the search engines out in the cold. The search engine has become the only one with a VIP pass to Reddit's repository of user-generated content, reported 404media.

This move not only highlights Google's near-monopoly in search but also Reddit’s decision to limit data scraping for AI training to Google, which secured a special multi-million dollar deal for continued access.

Other search engines, such as Bing, DuckDuckGo, and Qwant, can't display recent posts from Reddit, as the microblogging website has restricted access to its data.

This comes after Reddit took a firm stance against AI companies scraping the site to train LLMs. Last year, it disrupted the ecosystem of beloved third-party apps by introducing fees for accessing its API, making it financially unfeasible for many of these apps to continue operating, added the 404media report. 

This decision was a significant move to protect its content and maintain control over its data. So, if you're looking to search for those embarrassing questions specifically on Reddit, you'll now have to do it through Google.

Sparking concerns about the misuse of AI as the United States Election Day approaches, a video featuring an AI-cloned voice impersonating Vice President Kamala Harris went viral after Elon Musk shared it in an X post.

In the video, Harris is saying things she never actually said. The fake ad cleverly mimics a real campaign ad but replaces her actual voice with a convincing AI-generated one, highlighting the potential for AI to mislead and the importance of discerning truth from fiction in political content. You can read more about it here.

Anyway, the video went viral within hours, and Musk acknowledged it was satire a day later, but by that time, the damage had already been done.

And in the category of 'AI outperforming humans,' an AI won a silver medal at this year's International Mathematical Olympiad (IMO), marking the first time a model such as that was placed on the podium. 

The IMO is a highly respected competition for young mathematicians. Its challenging questions usually require a level of mathematical skill that AI systems generally lack.

Developed by Google DeepMind, AlphaGeometry could solve some IMO geometry questions. However, it wasn’t tested live and couldn’t handle other mathematical areas needed to win an IMO medal. 

So, Google DeepMind upgraded AlphaGeometry and introduced AlphaProof, which together solved four out of six IMO questions correctly. This earned them a silver medal with a score of 28 out of 42 points, just shy of the gold medal threshold. You can read more about it here.

To sum up the other news of last week:

Adobe rolls out new AI features

Adobe adds new generative AI tools to Illustrator and Photoshop, like Generative Shape Fill and Text to Pattern in Illustrator and Generate Image and Enhance Detail in Photoshop, to speed up and simplify creative work.

Kling AI is now available to all

Remember Kling AI, the Sora killer? We spoke about it at length in a previous newsletter. Well, the latest and exciting update is that the video model, which was restricted to China and Chinese phone numbers only, has now gone global. 

Social media users claim that it’s uncensored and that you can create videos of celebrities. But can you? I obviously had to try out the ‘Will Smith eating spaghetti’ prompt. More about that in the tutorial section below.

Google Gemini users get quicker 1.5 Flash AI model

Google is enhancing Gemini AI with its new 1.5 Flash model, making it faster and more efficient, claims the company. This upgrade improves response speed, reasoning, and image understanding on both web and mobile, even for users without Gemini Advanced. 

Additionally, Google is introducing a feature to reduce misinformation by displaying links to related content in responses, including potential links to your emails if Gemini is connected to your Gmail account. This feature is currently available in select locations for English prompts only.

Apple Intelligence (AI) comes to iOS 18.1 dev beta

iOS 18's developer beta just dropped a taste of Apple Intelligence. If you have an iPhone 15 Pro, iPad, or Mac with an M1 chip, and you speak US English, you can join the waitlist for AI goodies. 

Features include a revamped Siri, text tools for proofreading with different tones, email summaries, and a smarter Focus Mode. Note: You might have to wait a bit for all the bells and whistles, as some features are still in the fix-it phase. 🤖📩

Scarlett Johannson-like voice to be rolled out to ChatGPT Plus users

After some serious drama, OpenAI is going to roll out ChatGPT's advanced voice mode which features ‘Sky,’ a voice resembling that of actress Scarlett Johannson’s, which was the cause of the delay from May to July.

The use of Johansson's voice is noteworthy because her character in the movie Her is an AI assistant. The actor previously raised objections to her voice being used in this context. Only a lucky few ChatGPT Plus users get to try it out first.

Did a friend forward this e-mail to you?

IE₊ SUPPORT INTERESTING ENGINEERING
_{Invest In Science And Engineering}

Enjoy exclusive access to the forefront of AI content, highlighting trends and news that shape the future. Join a community passionate about AI, delve into the latest AI breakthroughs, and be informed with our AI-focused weekly premium newsletters. With IE+, AI reporting goes beyond the ordinary - and it is Ad-Free.

NEWS

• AI to process 6,000 images in seconds for Hong Kong customs
  Initial research on its AI search engine showed it was 80% accurate, saving officers time.

• AI-backed drones to spot juvenile sharks and issue alerts in US
  AI-backed drones – SharkEye allows for quick detection of sharks even below the water surface and issue rapid text-message alerts to beachgoers in California.

• Breakthrough AI model predicts heat movement in materials 1,000,000 times faster than non-AI methods
  The new technique could find new materials that could greatly boost energy efficiency in global power grids and beyond.

• Congresswoman’s AI-cloned speech creates history in US politics
  By analyzing hours of past speeches and public appearances, ElevenLabs successfully cloned Wexton’s voice within days.

MUST-READ

AI fears rise: Elon Musk’s fake Kamala Harris video post hits 123 million views 

A manipulated video mimicking Vice President Kamala Harris’s voice has raised significant concerns about the misuse of artificial intelligence as the U.S. presidential election approaches.

The video, which portrays Harris saying things she did not actually say, has sparked a debate over AI’s impact on political misinformation and election integrity.

The controversial video first gained widespread attention when tech billionaire Elon Musk shared it on his social media platform, X, on Friday evening. Musk’s post, which has been viewed over 123 million times, called the video “amazing” with a laughing emoji but didn’t mention it was a parody. This has increased worries about AI’s ability to mislead people.

AI PICTURE OF THE WEEK

We are back with an AI challenge for this week. Let’s go! 

AI TOOL OF THE WEEK

I didn’t get to try Kling AI when it came out, since I don’t have a Chinese phone or a number. But here are the results for my prompt: ‘Will Smith eating spaghetti.’

And, of course, it’s not right at all. That man is not Will Smith and is very very Asian. It appears Kling AI may have implemented safeguards to prevent misuse and the spread of misinformation after all.

Here’s how you can use the now globally available Kling:

To access Kling, you have two options:

1. The most feature-rich version of Kling is available exclusively in China. This version requires a Chinese phone number to access. Although there are services that can provide you with a Chinese phone number, these services are often unreliable and not recommended.
   

2. Alternative Option: If you don’t have access to a Chinese phone number, you may need to explore other versions or solutions available outside China, which may have different features. Like the website I used: KlingAI.com.

You’ll have to sign in with an email address and get started making videos.

Easy-to-use interface

You can turn your textual prompts into videos or images with Kling, and your prompts can be up to 2,000 characters long. Best to get as detailed as you want. It seems the more elaborate and thoughtful your prompt, the better the results.

Setting tabs

The settings menu has multiple toggle options. We can use a slider to distribute control between creativity and relevance. The higher the slider on one side, the more it follows that particular prompt.

The ‘Mode’ button lets you choose whether you want higher performance or quality. The video length can be 5 or 10 seconds. In the paid versions, the video length can be extended.

You can also choose the frame ratio between 16:9, 9:16, or 1:1. You can also choose the kind of camera movements you want. For example, by selecting zoom, there will be a gradual zooming into the video as it pans out.

I gave it a simpler prompt: ‘A panda picking strawberries,’ and chose the zoom option. Here’s the output:

While the panda isn’t picking strawberries in the photo, and neither does Kling AI zoom in, the quality of the output is quite realistic. I also did a quick Reddit search on Google to check if pandas eat strawberries, and they apparently don’t. It’s bamboo all the way!