The AI Rundown by Lightscape Partners – 12/11/23

Good morning and welcome to the first release of the AI Rundown by Lightscape Partners! We are excited to launch the newsletter and look forward to giving you the rundown on last week’s biggest stories in artificial intelligence.

Last week, Google announced their GPT competitor, “Gemini”, and have already received scrutiny for falsifying the release video. But the tech giant claims their model beats GPT-4 in some benchmarks. Apple enters the generative AI space with their release of open-source framework and model libraries. Also, Mistral, a France-based AI startup raises $737 million as they set their sights on OpenAI. Keep reading to learn more!

Top AI Developments of the Week

Google Announces Next-Gen Large Language Model “Gemini”. Link.

Google unveiled its next-generation large language model (LLM), Gemini, designed to power the future of Bard and other AI products. See an announcement video here.

Gemini boasts several key features:

• Understands video, images, text, and voice simultaneously.

• Offered in three versions: Ultra (most powerful), Pro (general tasks), and Nano (mobile devices).

• Google claims it's the first LLM to outperform human experts on the MMLU test.

Key Points:

• Gemini Ultra achieved superior performance compared to GPT-4 in benchmark tests, according to a technical paper they released.

• Gemini will be gradually rolled out. Bard currently utilizes a fine-tuned version of Gemini Pro.

• Gemini Nano will power generative AI features on the Pixel 8 Pro.

• Access to Gemini Ultra is restricted until further safety testing.

• Google faces pressure from rivals like OpenAI, Microsoft, and Anthropic.

This announcement signifies significant progress in Google's AI development, positioning it to compete effectively within the rapidly evolving field of LLMs. While the announcement video gathered lots of excitement to the capabilities of Gemini, unfortunately, it seems the release was heavily edited.

Apple Makes Quiet Entry into Generative AI with MLX Framework and Model Libraries. Link.

Apple, previously considered conservative in its AI approach, has released MLX, a framework and model library designed for generative AI applications on Apple Silicon chips. This marks a significant shift towards more advanced and publicly available AI tools from the company.

MLX Framework:

• Open-source, inspired by PyTorch, Jax, and ArrayFire.

• Shared memory architecture for efficient execution on Apple devices.

• Capable of training powerful AI models like Meta's Llama and Stable Diffusion.

Target Applications:

• Generative AI apps on MacBooks and other Apple devices.

• Potential integration with existing Apple services.

This release represents a significant change in Apple's AI strategy, moving beyond embedded machine learning towards more powerful and public-facing tools. MLX and MLX Data are open-source and accessible on GitHub and PyPI, encouraging developer adoption. Apple's entry into generative AI suggests potential competition for established players like Microsoft and Google.

Venture

Mistral, an OpenAI rival, nears $3b valuation in latest funding round. Link.

French AI startup Mistral AI is nearing a massive $450 million funding round, valuing the company at nearly $2 billion. This move positions Mistral as a significant competitor to OpenAI in the generative AI space.

• Funding Amount: €450 million ($737 million)

• Valuation: US$2 billion ($3 billion)

• Investors: Andreessen Horowitz (leading with €200 million), Nvidia, Salesforce (with €120 million in convertible debt)

• Focus: Open-source software for generative AI tools

• Founded: 2023 by former DeepMind and Meta scientists

• Previous Funding: US$113 million in June 2023

Additional Notes:

• Mistral AI directly competes with OpenAI in the development of generative AI tools.

• The company's open-source software approach resonates with developers.

• The large funding round highlights the growing interest in generative AI technology.

• Mistral's co-founders are selling equity in the company as part of the deal.

This news signifies the rapidly evolving landscape of generative AI, with Mistral AI establishing itself as a major player alongside established giants like OpenAI.

Vast Data Raises $118 Million to Build Next-Gen AI Infrastructure. Link.

Vast Data, the New York-based startup specializing in AI data storage solutions, secured $118 million in Series E funding led by Fidelity Ventures. This funding aims to fuel their global expansion and solidify their position as a leader in next-generation AI infrastructure.

• The funding round values Vast Data at $9.1 billion post-money, bringing their total raised to $381 million.

• Vast's solution eliminates tiered storage, providing unified storage, database, and compute engine services for seamless data management.

• The platform caters specifically to AI and GPU-accelerated workloads, facilitating efficient data access for these demanding tasks.

• Vast boasts impressive growth, with $200 million in annual recurring revenue and a 3.3x year-over-year increase.

• Notable customers include Pixar and Zoom, showcasing the diverse applications of Vast's technology.

• The new funding will be used to expand their global reach, particularly in Asia Pacific, the Middle East, and Europe.

• Vast faces competition from vendors like Databricks but claims first-mover advantage in the AI data storage market.

The company's success is a testament to the growing demand for robust AI infrastructure solutions. This news highlights the booming AI industry and Vast Data's potential to be a major player in shaping its future.

Respeecher, an ethics-first voice cloning solution, locks in $1 million in pre-Series A funding. Link.

Ukrainian startup Respeecher, known for its ethical approach to voice cloning, raised $1 million in pre-Series A funding. This investment will help expand their business in media, gaming, and accessibility.

• Respeecher uses voice models to modify speech instead of generating synthetic voices, offering a more natural and nuanced result.

• The company has worked with major names like Star Wars, CD Projekt Red, Warner Music, and Calm.

• Respeecher prioritizes consent and compensation, working closely with rights holders, actors, and estates. They built a library of opted-in actors and joined Adobe's Content Authenticity Initiative.

• The pre-Series A round saw contributions from entrepreneur Gary Vaynerchuk and other investors.

• The funding will enable them to expand their reach and continue developing their technology.

Respeecher's ethical approach distinguishes them from competitors who prioritize widespread adoption over individual consent. This news highlights Respeecher's potential to become a leader in the synthetic voice industry while prioritizing ethical considerations and responsible use of their technology.

Hardware/Software

AMD Announces New AI Chip Offerings and Partnerships to Challenge Nvidia. Link.

AMD unveiled a suite of new products aimed at capturing market share in the growing AI chip industry, directly challenging Nvidia's dominance.

• Instinct MI300X accelerator: Offers 1.5x more memory than its predecessor and boasts superior performance compared to Nvidia's H100 chip.

• Instinct MI300A APU: Combines CPU and GPU for faster processing and features unified memory eliminating data transfer between devices.

• Ryzen 8040 processor: Adds native AI functions to mobile devices with 1.6x more performance than previous models.

This announcement highlights the intensifying competition in the AI chip space, with AMD positioning itself as a viable alternative to Nvidia.

Ethics in AI

Meta Fights Back Against Deepfakes with Release of New Text-to-Image Product. Link.

Meta launched Imagine, a standalone text-to-image generation tool similar to DALL-E and Midjourney. While free to use, concerns around potential bias and image misuse raise questions about ethical safeguards.

• Imagine generates high-resolution images from text prompts.

• Meta's history with biased AI tools raises concerns about Imagine's potential for misuse.

• Meta plans to add invisible watermarks for traceability, but details are limited.

• Rising concerns about Deepfakes and AI-generated child abuse images highlight the need for transparency in generative AI.

Imagine's launch marks Meta's entry into the text-to-image arena, but ethical concerns regarding potential bias and image misuse demand immediate attention and robust safeguards. The growing pressure for clear AI-generated content labeling is evident in regulations and government hearings.

EU Reaches Provisional Agreement on Landmark AI Act. Link.

• EU negotiators reached a provisional agreement on the AI Act, potentially setting global standards for AI regulation.

• The Act mandates risk assessments, transparency measures, and complaint mechanisms for high-impact AI systems.

• Certain AI uses, including facial recognition for social scoring and manipulation, are explicitly banned.

• Fines for non-compliance range from €35 million to 7.5 million or 1.5-7% of global revenue.

• Finalization requires additional negotiations and votes, expected before the end of 2023.

• Debates over biometrics and general-purpose AI models led to delays in the agreement.

• Lawmakers push for a complete ban on biometric surveillance, while governments seek exceptions.

• Late proposals for self-regulation of generative AI models also contributed to delays.

The EU's AI Act marks a significant step towards regulating AI development and use globally, with potential implications for companies, governments, and citizens worldwide.

Other

Meta AI Year-End Update: New Features, Expanding Access, and Safety Initiatives. Link.

Meta AI is evolving, with new features across messaging platforms, Facebook, and a standalone "imagine" experience. The update emphasizes safety initiatives like invisible watermarking and red teaming.

• Meta AI chat assistant is getting smarter, with improved responses, search capabilities, and post suggestions.

• Meta AI is accessible through Facebook, Instagram, Messenger, and WhatsApp.

• Reels integration allows users to discover content related to their queries.

• Facebook tests AI-powered birthday greetings, Feed post editing, and Group topic suggestions.

• AI-generated images are coming to Facebook, including landscape-to-portrait conversion.

• Suggested replies in DMs help creators engage with their audience faster.

• Search is coming to more Meta AIs, including sports-related ones.

• Long-term memory is being tested for select AIs, enabling deeper connections.

• Red teaming exercises help ensure the safety of generative AI models.

This news showcases Meta's continuous development of AI features across its platforms, while emphasizing responsible AI practices and user safety.

Thank you for reading the AI Rundown by Lightscape Partners. Stay tuned for the latest updates in the world of Artificial Intelligence!