Good morning and welcome back to another week of The AI Rundown by Lightscape Partners.

The biggest stories in AI last week:

• Artificial intelligence allows for a new breakthrough in stabilizing nuclear fusion, bring humanity one step closer to unlimited clean energy.

• Stability AI released their newest generative image model, Stability AI, bringing more capabilities and impressive visuals to the space.

• Google finally taps into Reddit’s massive data collection of user discussions, inking a $60 million deal with the social media company.

If you’d like a quick refresher on industry terms and the current landscape of AI, skip to the bottom.

Otherwise, keep reading to see the biggest stories in AI last week.

Top AI Developments of the Week

Scientists using artificial intelligence have overcome a major hurdle in nuclear fusion instability. Link.

• AI developed by Princeton University scientists predicts and prevents plasma instabilities in nuclear fusion, overcoming a significant challenge for clean energy production.

• Nuclear fusion, offering vast energy potential without fossil fuels or hazardous waste, mimics reactions within the Sun but has been difficult to harness.

• The breakthrough enables control over plasma instabilities 300 milliseconds before they occur, allowing for adjustments to maintain stable fusion reactions.

• This advancement follows a 2022 milestone where the Lawrence Livermore National Laboratory achieved net energy gain from nuclear fusion, albeit on a small scale.

• The research, suggesting grid-scale nuclear fusion energy could become viable, was published in Nature under the title "Avoiding fusion plasma tearing instability with deep reinforcement learning."

On the heels of OpenAI’s Sora hype, Stability AI debuts the breath-taking Stable Diffusion 3. Link.

• Stability AI announced Stable Diffusion 3 (SD3), an advanced AI image-generating model aimed at surpassing competitors like OpenAI's Sora and Google's Gemini, with a new architecture designed for a variety of hardware setups.

• Unlike OpenAI and Google models, SD3 is not limited to an API and aims for broad hardware compatibility, including serious GPU setups for machine learning tasks.

• Emad Mostaque, head of Stable Diffusion, highlighted SD3's potential for multimodal understanding and video input/generation, aiming to surpass API-driven competitors in versatility and accessibility.

Google signs $60 million deal with Reddit to train AI models on Reddit’s content. Link.

• Google will be allowed to use online discussion posts to train its AI models and improve other Google services, like Google Search.

• Reddit, in return, will have access to Google’s AI models to improve internal site search and other features.

• The partnership comes days after Reddit’s IPO announcement.

Venture

Magic AI raises $117 million to build AI software engineers. Link.

• The startup aims to create a coworker, not a “copilot”, which already has several major players.

• The round was led by NFDG Ventures, Nat Friedman and Daniel Gross’s firm, bringing the total funding to date to over $145 million.

SoftBank Group’s Masayoshi Son seeks $100B for a new AI chip venture. Link.

• Under the codename Izanagi, the venture intends to partner with Arm, a chip design company spun out of SoftBank last year.

• SoftBank intends to finance $70 of the $100 billion from institutional investors based in the Middle East.

Software + Hardware

Google releases new ‘open models’ for free. Link.

• Google released new AI models named Gemma, allowing developers to use and build AI software for free, following Meta's similar initiative.

• Gemma models are not fully open source, with Google setting terms of use; they are optimized for Google Cloud, offering $300 credits to first-time users.

• The Gemma models have two sizes: two billion or seven billion parameters, in contrast to Meta's Llama 2 models which range from seven to 70 billion parameters.

• Nvidia has collaborated with Google to ensure Gemma models work well on its chips and plans to integrate Gemma with its upcoming chatbot software for Windows PCs.

Adobe launches its own AI assistant to parse, summarize, and search PDFs. Link.

• The AI assistant can summarize and answer questions about PDF documents.

• Whereas ChatGPT and other models require uploading documents, Adobe’s AI assistant is built-in.

• The tool will work with PDFs, Word, PowerPoint, and other documents.

• Currently it is available via a free beta, but will transition to a subscription plan later on.

Mistral AI releases its newest GPT-4 competitor: Mistral Next. Link.

• The French AI startup launched an LLM prototype called “Mistral Next”, which is currently available for testing.

• This will be Mistral’s largest and most impressive model thus far.

• Mistral AI has differentiated itself due to the open-source nature of its models and company, constantly proving it can hold up against “closed” models, like OpenAI.

Ethics in AI

Following diversity controversies, Google pauses AI image generation with its Gemini AI tool. Link.

• Google paused its Gemini AI tool's human image generation feature due to controversies over producing images with perceived bias, including Black founding fathers and a female pope.

• Google struggles with AI bias, aiming to avoid stereotypes while facing challenges in generating non-stereotypical, accurate images.

• The incident highlights the broader issue in AI of generating content without inherent biases, a problem exacerbated by historical data and societal prejudices.

• The debate underscores the complexity of AI's societal impact, emphasizing the need for AI systems to navigate cultural and ethical considerations.

Get up to speed on the current landscape

• AI Term Glossary

• Current AI Landscape from Sequoia Capital

Thank you for reading the AI Rundown by Lightscape Partners. Please send any questions, comments, or suggestions to [email protected].