E37: 4/26/24
The training data race continues, Microsoft/Meta/Apple all double down on on-device SLMs, Snowflake enters the LLM arena, several AI-copilot-for-code startups raise >$1B val. mega rounds, and more...
Here’s this week’s scoop! Also, inspired by Parsa @ AI2, I'm adding a new section for upcoming AI events in the Seattle area. If you know of any events I missed or you think I should add, please let me know!
PNW AI News:
Amazon updates AWS Bedrock to enable developers to import their own customer AI models, rather than just the off-the-shelf Anthropic Claude 3 Opus LLM suite
They also announced support for Meta's LlaMa 3 foundation models, and soon will include Cohere's Command R and R+ models for enterprise AI
Microsoft joins the small language model (SLM)race with the Phi-3 Mini, a 3.8B-parameter model designed for on-device performance
MSFT beats Q1 earnings expectations as quarterly revenues rise 17% to $61.9B, profits rise 20% to $22B
According to Satya, Microsoft “is seeing a new era of AI transformation driving better business outcomes across every role and industry”
Microsoft says cloud AI demand is exceeding supply even after 79% YoY surge in capital spending into AI workflows + infra
The UW's Institute for Protein Design is yielding several high profile AI-powered protein design startups, including SF-based Xaira Therapeutics co-founded by head of UW's IPD, which emerged from stealth with $1B in funding from Arch Venture Partners, NEA, Sequoia, Lux Capital, Menlo Ventures, Lightspeed, and others
Key AI Product Updates:
Apple has released OpenELM, a suite of 8 different open-source LLMs on Hugging Face designed for on-device execution
Ranging from 270M to 3B parameters, encompassing both pre-trained and instruction-tuned variants
Snowflake breaks its silence in the AI wave, announcing "Snowflake Arctic", a fully open-sourced 10B+128x4B mixture-of-experts (MoE) LLM, which claims to beat rival Databrick's DBRX on every benchmark
10B+128x4B MoE = a 10B-parameter general context model that routes prompts between 128 different specialist "expert" 4B-parameter models
OpenAI introduces "instruction hierarchy," a new method to enhance the security of LLMs against prompt injection attacks and jailbreaks
This method prioritizes system instructions from developers over user input or third-party tools
Leading Hollywood talent agency CAA is building CAA Vault, allowing A-list clients to create AI clones of themselves to license for 3rd-party creative use
OpenAI co-founder Andrej Karpathy is claiming current LLMs are "undertrained by a factor of maybe 100-1000X"
Coincides with Meta's LLaMa 3 unveil with an unprecedented data set of 10M high-quality scenarios and 15 trillion tokens
Exceeding Deepmind's recommended “training-data:model-parameter” threshold by 75X, while being lightweight enough to run locally on most iPhones
Meta just announced multimodal capabilities for the Ray-Ban Meta smart glasses, integrating AI to process and understand a user's surroundings
Ask Q's like "what species is that bird?"
Profluent just developed OpenCRISPR-1, the world's first open-source AI-developed gene editor capable of editing the human genome
Adobe just released VideoGigaGAN, is an AI model designed to upscale low-resolution videos by up to 8X in definition without hallucination
Generative adversarial networks (GANs) are a popular ML framework that create new data instances that resemble the initial training data
Sakana AI just released EvoSDXL-JP, a text-to-image generator specifically tailored for Japanese-style images 10X faster than Stable Diffusion
Chinese company SenseTime just launched SenseNova 5.0, a 600B parameter LLM that was trained on >10TB of synthetic data, that beats GPT-4 Turbo across nearly all key benchmarks
Synthesia unveiled a new generation of “Expressive Avatars” that can convey a wide range of human emotions, advancing the hyper-realism of video outputs
Key AI Business/Investment Updates:
Moderna expanded its partnership with OpenAI to integrate AI across its entire org, aiming to accelerate the development of life-saving mRNA treatments using 750 different internal GPTs
Viral deepfake videos of A-list Bollywood actors criticizing Indian PM Narendra Modi have been spreading ahead of the election, raising concerns about AI’s use for misinformation
Seattle-based TrueMedia founded by AI2 leader Oren Etzioni hoping to combat this for the upcoming US election
Elon's 10-month old xAI is raising $6B at a $24B post-money valuation from investors including Sequoia and Future Ventures, up from the rumored $3B on $18B post just last week
6-month old AI startup Cognition raised $175M at a valuation over $2B from Founders Fund, despite no revenue and mixed sentiment on its ‘Devin’ AI coding agent
Founded by ex-Microsoft software engineers to challenge Github Copilot, Augment emerged with $252M in funding at $980M post from former Google CEO Eric Schmidt, Index Ventures, Lightspeed, Meritech, and more
These valuations of AI-for-code startups like Cognition/Augment are absurd
Run:ai, a Kubernetes-based software that optimizes the workload of AI apps on GPUs, was acquired by Nvidia for $700M in cash
Apple acquired Datakalab, a Paris-based AI startup that specializes in data compression and image analysis
Legislators in 24 states are working on bills or have passed laws to combat AI-generated sexually explicit images of minors
PNW AI/ML Fundings:
DropzoneAI (Seattle, WA) raised a $16.9M Series A led by Tom Tunguz' Theory Ventures, joined by Decibel Partners, PSL Ventures, In-Q-Tel, and several angels
Software platform intended to deliver pre-trained autonomous AI security agents that work alongside human analysts on security operations teams - TFTD $20.5M
CEO: Edward Wu
Rendered.ai (Bellevue, WA) raised an undisclosed post-Series A round from In-Q-Tel
Data engineering tools designed for generating synthetic datasets for training AI and ML computer vision systems - TFTD $12M+
CEO: Nathan Kundtz
Today (Seattle, WA) raised $5M in Seed funding co-led by Sfermion and Big Brain Holdings, joined by Compute Capital, Collab+Currency, Spirit DAO, Metavest Capital, LIF, and more
Web3 social simulation game, offering interactions with GenAI backed NPCs, and no-code tools for virtual world development - TFTD $5M
Shipped (FKA Invisible Commerce - Seattle, WA) raised $625K in Pre-Seed funding from undisclosed investors
Ecommerce AI customer success agent that can actually take actions and execute resolutions - TFTD $625K
CEO: Jonathan Wu
Lawlink (Kirkland, WA) raised $550K in Pre-Seed funding from undisclosed investors according to an SEC filing
Stealth AI LegalTech startup - $550K
CEO: Pat Wilburn
Upcoming PNW AI/Startup Events:
Seattle AI Frontier Summit - VC Insights, AI Innovation, Startup Pitch and Networking
April 27th: sign up here
Dent:AI Startup Showcase Reception
April 29th: sign up here
YoungTech Seattle Fireside Chat III: Ascend x Yoodli
May 1st: sign up here
Seattle's AI Tinkerers
May 20th: sign up here