E49: 7/26/24

Meta's massive LLaMa 3.1 405B parameter model, paid-tier AI native Alexa, Elon's GPU supercluster for training Grok 3.0, and Bain reveals that GenAI in production is down YoY since 2023

PNW AI News:

• Amazon will be launching a paid tier for Alexa with AI-enabled intelligence as soon as this month, after having invested >$25B into it since 2017

  • Made challenging by Alexa's infra, which relies on 3rd-party apps for specific categories of requests

• Seattle’s Hiya, a startup battling fraudulent calls and providing technology to protect voice identity, acquired deepfake startup Loccus.ai on undisclosed terms

• Microsoft unveils serverless fine-tuning for its Phi-3 SLM, streamlining the adaptation of AI models without managing servers, directly on Azure AI's serverless endpoints

Key AI Product/Research Updates:

• Meta just released Llama 3.1 405B, achieving state-of-the-art performance across key benchmarks, competing with leading closed finetuned models and GPT-4o

  • 128K token context window, 8 languages supported. Doubling down on "open source for good"

• OpenAI announces SearchGPT, its AI-powered search engine that offers real-time access to internet information. It employs  GPT-4 to coalesce disparate info into coherent, attribution-rich summaries

  • In private beta now

• AMD claims its new laptop chips can outperform Apple's M3, boasting improved performance in multitasking, image processing, and gaming

• Kling AI globally release its video generation model, previously limited to China, to challenge OpenAI's Sora in both English and Mandarin prompts

• Mistral released "Large 2", a new AI model claiming to exceed the performance of recent offerings from OpenAI and Meta, despite having far less parameters

•  Elon Musk and xAI just announced the Memphis Supercluster, “the most powerful AI training cluster in the world“, also revealing that Grok 3.0 is planned to release in December

• Nvidia is engineering a version of its new flagship Blackwell AI chip, called the B20, specifically for the Chinese market while adhering to U.S. export controls

• Google's free Gemini chatbot gets 1.5 Flash update, making responses faster and smarter with a 32,000-token capacity—four times more than its predecessor

• Apple just unveiled DCLM-7B, a new 7B parameter open-source AI model that outperforms Mistral 7B

  • Also has a 1.4B parameter SLM variant

• Researchers at EPFL have identified a critical security vulnerability in leading AI language models, like GPT-4o and LLaMa3-8B. By rephrasing malicious queries into the past tense, users can often bypass the models' safeguards

• Google researchers have developed a new AI-powered weather and climate model called ‘NeuralGCM’, by combining methods of ML with traditional physics-based modeling

Key AI Business/Investment Updates:

• OpenAI is reportedly in talks with chip designers like Broadcom to develop its own AI chip, aiming to reduce dependence on scarce and expensive GPUs from Nvidia

• Movie performers are striking over AI concerns, after contract talks with major game studios failed, with the union arguing that current proposals inadequately cover performers’ rights

  • SAG-AFTRA contests the studios’ definitions of who qualifies as a “performer”, the potential for AI to replicate voices and likenesses without proper compensation

• According to a Bain survey, enterprise usage of GenAI in production is actually down YoY since 2023, despite AI pilots and R&D increasing in 2024 across every category

• Coatue ($70B AUM investment firm) released a comprehensive report on the current state and future prospects of AI humanoids and robotics at their EMW conference

• Cohere raised $500M in additional funding from investors including Cisco, AMD, and Fujitsu, bringing its new valuation to a whopping $5.5B

• Five US Senators issued a letter to Sam Altman, demanding details about OpenAI's efforts to ensure AI safety following reports of rushed safety testing for GPT-4 Omni

PNW AI/ML/CV Fundings:

• QA Wolf (Seattle, WA) raised $36M of Series B funding led by Scale Venture Partners, joined by Notation Capital, Inspired Capital, and Threshold Ventures

  • Data-driven QA solution platform intended to discover bugs and offer AI test automation - TFTD $57.5M

  • CEO: Jon Perl

• Archera (Seattle, WA) raised $17M in Series B funding led by HighSage Ventures which included participation from Ridge Ventures, Amplify Partners, and PSL Ventures

  • Developer of a cloud-based financial management platform designed to focus on AWS and Azure cloud cost reduction – TFTD $27.5M

  • CEO: Aran Khanna

• RevealDx (Seattle, WA) raised $4.25M in Seed funding from undisclosed investors

  • Computer vision-enabled clinical decision support technology designed to improve earlier lung cancer diagnosis - TFTD $8.55M

  • CEO: Chris Wood

• GumLoop (Vancouver, BC) raised $3.1M in Seed funding led by First Round Capital with participation from Y Combinator and other angels

  • Developer of a no-code platform designed for automating workflows with AI – TFTD $3.6M

  • CEO: Max Brodeur-Urbas

• Signify (Seattle, WA) raised $2.1M in Pre-Seed funding co led by FUSE, Founders Co-op, and the AI2 Incubator

  • Developer of compliance management software designed to streamline regulatory document analysis for manufacturers – TFTD $2.1M

  • CEO: Martín Ramírez