E69: 12/19/24
Microsoft's Phi-4, Seattle-based Carbon.ai gets acquired by Perplexity, OpenAI continues to ship, Databricks' mega round, and more...
Hi everyone, This is the last AI newsletter of 2024! Will resume in a few weeks the holidays wind down. Enjoy!
PNW AI News:
Microsoft just released Phi-4, a 14B parameter small language model (SLM) that outperforms massive competitors like GPT-4o and Gemini Pro 1.5 in areas like mathematical reasoning despite a drastic size difference
Microsoft’s "AI for Good" Lab unveiled a new open-source project and AI-powered edge computing device designed to improve biodiversity monitoring, with the goal of helping outside researchers better understand the planet from afar
UW Computer Science researchers craft method of fine-tuning AI chatbots for individual taste, dubbed "variational preference learning"
Key AI Product/Research Updates:
OpenAI's "12 days of Ship-mas" continues:
Day 7: OpenAI released "Projects" in ChatGPT, enabling users to curate and structure files and conversations within customizable folders.
Dav 8: OpenAI lifts the paywall for ChatGPT’s search, unlocking real-time, web-sourced results for all registered users.
Day 9: OpenAI rolls out its full o1 reasoning model API, but with access limited to Tier 5 developers with at least a $1,000 monthly API spend.
Day 10: OpenAI makes ChatGPT available for phone calls, texts, or WhatsApp by dialing 1-800-CHATGPT.
Google announced Veo 2, an AI video generation model that is beasting OpenAI's Sora Turbo in most benchmarks. It can respond to intricate filmmaking prompts, including lens specifications, camera effects, special effects with realistic physics, etc.
Elon's xAI has rolled out an upgraded Grok AI chatbot (Grok-2) that now operates 3x faster than its predecessor, rivalling Anthropic's Claude Sonnet 3.5 in its ability to follow prompts, with marked improvements in accuracy and language support
Google Labs has dropped Whisk, a genAI tool that reimagines image creation by focusing on visual inputs instead of the usual text prompts, using Google’s Gemini 2.0 Flash model
Google DeepMind has introduced FACTS Grounding, a new benchmark that tests AI models' ability to provide accurate, document-based answers to factcheck AI outputs
NVIDIA introduced the Jetson Orin Nano Super Developer Kit, a $249 compact genAI supercomputer that delivers 1.7x performance gains at half the previous model's price
Key AI Business/Investment Updates:
Databricks raised a $10B funding round at a $62B valuation led by Thrive Capital, with the company planning AI product expansion, joining the "AI talent war", and exploring potential acquisitions
One of the biggest VC rounds ever
The US is preparing new export regulations for AI chips that would position companies like Google and Microsoft as global gatekeepers for access to this technology outside the US, ideally reducing reliance on TSMC in Taiwan
Ilya Sutskever, OpenAI co-founder and former chief scientist, announced a shift in AI development at NeurIPS, declaring, "We've achieved peak data and there'll be no more. Pre-training as we know it will unquestionably end."
He believes the AI industry is approaching the limits of new data, likening it to the depletion of fossil fuels, with the internet representing a finite resource