🎉 This week's AI Chronicles: GigaGAN, Privacy, and Peek-a-Boo Code!

This week, GigaGAN is painting the town with pixel-perfect images in a flash, while the U.S. Copyright Office ponders over AI's role in artistic copyright. Italy bids "arrivederci" to ChatGPT over privacy concerns, and Google's Bard AI faces allegations of peeking at ChatGPT data. In the meantime, Vicuna, an open-source chatbot, steps into the spotlight with impressive performance on a budget. Finally, Elon Musk pulls back the curtain on Twitter's algorithm, but leaves the ad recommendations and data training safely hidden. It's a week of revelations, rivalries, and a little bit of mystery! 🧙🏻‍♂️

🎨 GigaGAN: Picture Perfect Pixels in a Jiffy!

A team of researchers developed GigaGAN, a new GAN architecture that leaps past the limitations of previous models like StyleGAN. Unlike its computationally expensive cousins, GigaGAN is a speed demon, whipping up a 512px image in just 0.13 seconds and 4096px at 3.7s! It even crafts high-resolution 16-megapixel masterpieces in a mere 3.66 seconds. With a lower Fréchet inception distance (FID) than competitors like Stable Diffusion v1.5 and DALL·E 2, GigaGAN is the new kid on the block, showing off its disentangled, continuous, and controllable latent space for text-to-image synthesis. Watch out, world; there's a new pixel Picasso in town!

https://www.marktechpost.com/2023/03/13/meet-gigagan-a-large-scale-modified-gan-architecture-for-text-to-image-synthesis/

🤖 AI Art: To Copyright or Not to Copyright?

The U.S. Copyright Office recently clarified guidelines on the copyright eligibility of AI-assisted artistic works. In short, protection hinges on whether AI's contributions are "mechanical reproduction" or truly reflective of the author's "mental conception." While generative AI systems like Midjourney, ChatGPT, and DALL-E are all the rage, most of their output likely doesn't warrant copyright protection. However, fear not, creative humans! Unique arrangements and modifications of AI-generated works can still be copyrighted, as long as the human creator maintains control over the work's expression and traditional elements of authorship. Just remember to disclose any AI assistance in your copyright applications!

https://www.reuters.com/world/us/us-copyright-office-says-some-ai-assisted-works-may-be-copyrighted-2023-03-15/

🇮🇹 Ciao ChatGPT: Italy Says Arrivederci over Privacy Concerns!

Italy becomes the first Western country to block ChatGPT, citing privacy issues with the AI model created by OpenAI and backed by Microsoft. The Italian data-protection authority is putting the brakes on ChatGPT and launching an investigation into its compliance with General Data Protection Regulation (GDPR). Among other concerns, the watchdog highlighted a recent data breach involving user conversations and payment information. OpenAI, meanwhile, insists it complies with privacy laws and looks forward to re-introducing ChatGPT in Italy "soon." Will other countries follow suit?

https://www.bbc.com/news/technology-65139406

🔍 Bard-gate: Did Google's AI Peek at ChatGPT?

Google's Bard AI faces accusations of using publicly shared ChatGPT responses for training, but Google denies the claims. The Information reported that a former Google AI engineer, Jacob Devlin, believed Bard's team was heavily relying on data from ShareGPT, a website where users share their ChatGPT conversations. Devlin, who now works at OpenAI, expressed his concerns to Google's senior management before resigning. Some Googlers thought that such usage might violate OpenAI's terms of service. Google stated that Bard isn't trained on data from ShareGPT or ChatGPT, but the debate continues. Will this AI drama have a sequel? Stay tuned!

https://9to5google.com/2023/03/30/google-bard-chatgpt-training-report/

💬 Vicuna: A Low-Cost, Open-Source Chatbot Rivaling ChatGPT

A team from UC Berkeley, CMU, Stanford, and UC San Diego developed Vicuna, an open-source chatbot based on LLaMA-13B that achieves 90% ChatGPT quality. With 13B parameters, Vicuna is fine-tuned on user-shared conversations from ShareGPT via public APIs. The chatbot, which costs around $300 to train, outperforms LLaMA and Stanford Alpaca in over 90% of cases. Although it has limitations in reasoning and math, Vicuna is a promising and cost-effective solution for chatbot development, with potential for future improvements.

https://www.marktechpost.com/2023/04/02/meet-vicuna-an-open-source-chatbot-that-achieves-90-chatgpt-quality-and-is-based-on-llama-13b/

🐦 Twitter Peeks Behind the Curtain

In a bid for transparency (and to avoid further embarrassment), Twitter CEO Elon Musk has released part of the social network's source code, including its timeline-recommendation algorithm. While it's a step towards an open-source Twitter, the company is still keeping its ad recommendations and data used for training under wraps. Some might say it's like getting a backstage pass to a magic show, only to find out the magician keeps his best tricks in a secret room. Elon promises improvements based on community feedback, but for now, we'll have to settle for this partial peek into Twitter's mysterious inner workings.

https://techcrunch.com/2023/03/31/twitter-reveals-some-of-its-source-code-including-its-recommendation-algorithm/

If you found this useful, do help share it around 🙏🏻 and subscribe if you haven’t!
https://shinynewthings.beehiiv.com/