🚓 NVIDIA raided in France 🚓

Today's issue is brought to you by NightCafe - the AI art studio with a twist!

In today’s edition (29 Sept):

📰 Quick News - NVIDIA raided in France and more

📚 Nerd section - How to run Llama 2 70B on consumer GPUs like NVIDIA RTX 3090 or 4090 and more

📰 Quick News 📰

🎙️ Spotify is adding auto-generated transcripts to millions of podcasts: introducing time-synced textual accompaniments for episodes, making content more accessible and skimmable. This move follows their recent AI-driven voice cloning tool for translations, promising to bolster user experience and engagement.

💸 China chipmaker Enflame raises $274M: amid a national push for semiconductor self-reliance, Enflame Technology secures significant funding, led by Shanghai International Group and backed by giants like Tencent. As US sanctions expand, China advances in chip manufacturing, with even Huawei showcasing gains despite import bans. The landscape shift sees global companies like Nvidia adjusting their strategies for the Chinese market.

🛂UK pushes for greater access to AI’s inner workings to assess risks: the UK emphasizes the need for transparency in AI to understand its potential risks, proposing open-access AI models and explainable AI systems while promoting collaboration among global stakeholders to strike a balance between innovation and accountability.

🔒 Your website can now opt out of training Google's Bard and future AIs: granting webmasters the choice to prevent their content from feeding Google's AI training, amidst discussions on consent and ethical data collection practices​.

📚 Nerd section 📚

📎 Microsoft and MIT Researchers Hope to Reduce AI Hallucinations with DoLa: Introducing a novel decoding approach, DoLa prioritizes deeper layers in LLMs to enhance factual output and reasoning, showing superior results in GPT-4 tests, without the need for external knowledge retrieval; however, it hasn't been tested across all domains and leans solely on the model's existing knowledge.

🎨 RealFill, Adobe Forefly alike, leverages reference images to generate personalized inpainting models, achieving authentic image completion that remains faithful to the original scene even under varying conditions, outperforming baseline methods in fidelity and visual quality.

We love to hear back from you!