Kling AI is now available

Wizard fam,

One of the links in my previous newsletter mistakenly triggered some phishing alerts in Gmail – please mark as safe sender to ensure it doesn’t happen again and that my future newsletters don’t end up in spam.

PS I’m taking some days off 🌞 Next newsletter coming Monday next week.

‍

Dario’s Picks

The most important news stories in AI this week

‍

1. Kling AI, the Chinese AI video generator, is now available worldwide. You might remember seeing clips of Kling from a few months back, as it went viral on social media (like this cat slicing chicken). Well, now it's here and it has text to video and image to video. You get a few daily credits for free, and paid plans start from 5$/month; this also gives you gives you videos without watermark, up to 3 minutes in length.

‍

‎ Why it matters‎ ‎ Along with Runway's Gen-3, we now have another powerful text-to-video model available – while we impatently wait for Sora.

‍

2. Runway Gen-3 used YouTube videos and pirated films in its training data. Runway has been asked before about the source of their training data, but only offered vague answers. Now, a leaked internal spreadsheet obtained by 404 Media, supposedly shows that Runway used popular videos from creators and brands on YouTube, as part of a company-wide effort to collect high-quality videos for training.

‍

‎ Why it matters‎ ‎ It's getting increasingly clear that few, if any, of the leading AI companies have clean hands when it comes to the rights of their training data. Some are working to improve the situation (at least in the public's eye) by partnering with renowned publishers. Unless we want creators to lose their incentive to create over the long term, some combination of regulation for AI data sourcing and better business models that fairly attributes a share of revenue to creators are likely needed.‍

‍

3. OpenAI launches experimental GPT-4o version for long output. It's available for those on the Alpha program through the API, with the gpt-4o-64k-output-alpha model name.

‍

‎ Why it matters‎ ‎ Being able to generate longer outputs would be particularly useful in certain use cases, such as coding. This is still only available to a limited amount of devs, and not in ChatGPT. However, it'll likely become part of ChatGPT at some point.

‍

4. Grok "secretly" training on your data. There’s a setting on X (previously Twitter) that, by default, gives your consent to use your post and actions on the platform for training Elon's next Grok model. You can opt-out, but only if you’re on the desktop version and the setting is kinda hidden.

The way to opt out is to go to Settings --> Privacy and safety --> Grok --> Toggle off the box labelled "Allow your posts as well as your interactions, inputs, and results with Grok to be used for training and fine-tuning".

‍

‎ Why it matters‎ ‎ AI models are only as good as their training data. Consider the “opt-out-of-training”button the new “cancel my subscription” button on the web: you know it exists, but they’ve made it damn hard to find.‍

‍

5. Reddit wants Microsoft to pay for using their data. Reddit has already struck licensing deals with OpenAI and Google, which allows these companies to scrape and use their content in search results. However, other search engines and AI models have been actively scraping Reddit, acting like all internet content is free to use, including Microsoft, Anthropic and Perplexity.

‍

‎ Why it matters‎ ‎ It's not only traditional media publishers that are affected by their content being scraped and used by AI models. New types of publishers, including internet forums like Reddit, are demanding payment for using their content to train AI and blocking web crawlers from the ones they don't have agreement with using robots.txt.

‍

6. A tool to understand how LLMs "guesses the next word". Maxime Labonne built a tool to help you more intuitively understand how an LLM works. You may have heard that LLMs simply "predicts the next word". The tool features sample sentences with a word missing, and you can try adjusting different parameters and see visually how the probabilities and content of that "next word" changes in real-time.

‍

‎ Why it matters‎ ‎ AI can feel like magic sometimes, and in a way it kind of is. However, this tool makes the process in which they come up with answers seem a little bit more logical, and like there's math rather than magic behind. Play around with the sliders and see how things change.