Devin the world’s first AI Software Engineer might be a fraud 🃏, Audio advancements in AI probably just had its chatGPT / Sora moment 🚀 and Adobe Firefly may have trained its model on Midjourney images 🤦. Just another week in the world of Artificial Intelligence.
Let’s jump into it:
- Udio launched their AI music platform and its nothing short of extraordinary. The last few weeks have been a great month as far as audio and AI are concerned. Suno.ai, Stability Audio and of course now Udio. What is remarkable is that if you prompt it correctly, you can even get it to do stand-up comedy. ChatGPT could write stand-up for you (albeit, poorly), now you have also have AI perform it for you. Fun times ahead. Udio also gives you a lot of generations, so head over and check it out.
- Adobe’s Firefly might have some less than reputable source of images as part of its training data according to a Bloomberg report. Of course, Adobe is clear that all their non-human pictures are still safe and everything goes through a rigorous moderation process. The company will double down on the moderation for their Video Generation tool where they are apparently paying creators for content / videos that will be used as part of the training data. We believe them. They’re one of the few incumbents that have kept pace and actually done right by the community for the most part. (Since the main article is behind a paywall – here’s the Tom’s Hardware version of the story for reference.)
- Devin – the world’s first AI software Engineer isn’t as great as it was hyped up to be. First off: We’re sorry for falling prey to the hype. While Devin is still very impressive and does better than other LLMs on certain benchmarks, Cognition-labs definitely over stated its capabilities. Especially with regards to what it can do in the real-world. We’re certain that AI will get better at real world projects, but that is going to require chaining a lot of non-code related tasks as has been pointed out by experts time and time again. Maybe this is where AI wearable companies like Rabbit might be on to something. Training AI on how to perform real world tasks and chaining them together. (BTW, Rabbit in its current form is likely to be very disappointing for anyone looking to make it a part of their daily use cases. We’d wait for R2 or R3 before buying in unless you’re an absolute enthusiast.)
The big lesson here is not to fall prey to the hype. But we’re not entirely sure that we’ve learned the lesson. The Jetsons future couldn’t get here any sooner as far as we’re concerned. The full video for those interested, below:
- Mistral.ai dropped a massive update on X. Mixtral 8x22B. It’s a base model that has 176Bn parameters. Meaning you can’t ask it direct questions the way you might other LLMs. The best way to think about it is as an autocomplete model. All the benchmarks’ we’ve seen on HuggingFace / Reddit say that it’s very good.
magnet:?xt=urn:btih:9238b09245d0d8cd915be09927769d5f7584c1c9&dn=mixtral-8x22b&tr=udp%3A%2F%https://t.co/2UepcMGLGd%3A1337%2Fannounce&tr=http%3A%2F%https://t.co/OdtBUsbeV5%3A1337%2Fannounce
— Mistral AI (@MistralAI) April 10, 2024
- Elevenlabs and Pika AI had their first AI Film making competition over the weekend. A 72 hour AI generated film making marathon that will hopefully push the use case of both platforms. We’re quite excited to see what comes of this. Sora going to Hollywood was just the beginning.
Let's get (post-)real.
— Pika (@pika_labs) April 12, 2024
You have until Sunday, April 14, 11:59 pm PST.
More details in the comments. Good luck! pic.twitter.com/oWi18i92yY
- Humane’s AI Pin was met with a lot of fanfare and confusion back when they first demoed a few months ago. It’s now been in the hands of reviewers for a week or so and unfortunately the overall report is that it’s mostly underwhelming. We are not any closer to a smartphone-less future than we were before its release. 🤷♂️ Nick Shevchenko literally built out an opensource AI wearable: Friend, over a couple of days at a hackathon. He was also sent a ‘cease & desist’ by Humane. That’s one way to get people to hate you even before you’ve launched something.
Bonus – Deepfake Ryan Gosling and Chance the Rappers explaining AI. This was our favourite discovery from this week. Videos below for your perusal. All thanks to How to Fly AI. Enjoy!
