Generative AI Unicorn Capitulation

Adept and Humane are looking for buyers.

Created Using DALL-E

Next Week in The Sequence:

  • Edge 399: Our series about autonomous agents continues with an overview of external aid planning. We dive into IBM’s Simplan method for planning in LLMs and review the Langroid framework for autonomous agents.

  • Edge 340: A must read about AlphaFold 3 which expanded capabilities to predict many of the life’s molecules.

You can subscribe to The Sequence below:

TheSequence is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

📝 Editorial: Gen AI Unicorn Capitulation

The week in generative AI was dominated by the news from the Microsoft Build conference, where “copilots, copilots, copilots” seems to be the new “developers, developers, developers.” We also had another OpenAI drama with allegations from Scarlett Johansson about her voice being copied. And yet, today’s editorial is about none of those topics 😉

Instead, I would like to focus on an emerging phenomenon we are seeing in the market related to the “capitulation” of many high-flying startups that raised billions in venture funding. Just this week, we saw news of Adept and Humane AI both looking for prospective buyers.

Adept was founded by some of the authors of the transformer paper with the promise of reimagining productivity agents using foundation models. Specifically, Adept leveraged a computer-vision-based model to learn from user interactions so they can be automated. The company delivered incredible IP in this area with the Fuyu models but seems to have issues scaling the business model.

Humane is another high-profile AI company that tackled the big problem of creating an AI-first consumer device. The idea that AI will power a new hardware form factor makes perfect sense, and Humane’s pin is a very clever design. However, Humane didn’t have a particularly successful launch and is facing the traditional scaling challenges of hardware companies.

Adept and Humane joined recent stories such as Inflection, which proves that building businesses relying on generalist AI models is capital intensive and hard to scale. These early capitulations are the result of a heavily funded market evolving at a lightning-fast speed, making it really hard to identify successful business models this early. Regardless of this outcome, Adept and Humane have brought a tremendous level of innovation to generative AI, and likely a lot of their products will continue to evolve in their next phase.

🔎 ML Research

Inside the Mind of an LLM

Anthropic published a research paper outlining major advancements in foundation model interpretability. The research highlights how to identify interpretable features in mega large models like Claude Sonnet —> Read more.

Key-Value in Cross Layer Attention

Researchers from MIT published a paper outlining an optimization technique for key-value(KV) caching. The method, called l Cross-Layer Attention (CLA), extends paradigms such as Multi-Query Attention (MQA) to share values between adjacent layers —> Read more.

Modular LLM

Microsoft Research published a paper detailing module LLM architecture based on a library of LoRA adapters. The paper also introduce Arrow, a zero-shot routing mechanism for selecting the right adapter for a given task —> Read more.

LANISTR

Google Research published a paper introducing LANguage, Image, and STRuctured data(LANISTR), a multimodal model that is able to learn from unstructured and structured data. LANISTR works with both text, image as well as time-series and tabular datasets —> Read more.

KV-Runahead

Apple Research published a paper proposing a technique for parallelizing key-value cache generation. The key principle of this technique is to minimize the time to first token which optimizes the computation efficiency of the attention layer —> Read more.

🤖 Cool AI Tech Releases

Phi-3

Microsoft’s Phi-3 family of models is now available on Microsoft Azure —> Read more.

More Microsoft Announcements

Microsoft has many AI announcements at its Build conference —> Read more.

Aya 23

Cohere released Aya 23, an 8 and 35 billion parameter LLM supporting 23 languages —> Read more.

PaliGemma

Google’s small language-vision model PaliGemma is now open source —> Read more.

Llama3-ChatQA

NVIDIA released a RAG optimized version of Llama3 —> Read more.

SynthID

Google DeepMind expanded SynthID’s capabilities to insert watermark in AI generated text, images and videos —> Read more.

🛠 Real World AI

Data Management at Meta

Meta engineering discusses the composable data infrastructure powering part of its AI infrastructure —> Read more.

📡AI Radar

By Julio MarchiPosted on