OpenAI is Starting to Look Like Apple in 2008

Some non obvious views on the most important announcement of the week.

Created Using DALL-E

Next Week in The Sequence:

• Edge 343: The fine-tuning series continues with an overview of the Llama Adapter technique including its original paper. We also review the Chatbot Arena framework.

• Edge 344: We discuss another of hte great papers of the year in which Google shows that the combination of LLMs and memory is enough to simulate any algorithm.

You can subscribe below:

📝 Editorial:  OpenAI is Starting to Look Like Apple in 2008

The OpenAI Developer Day conference dominated the generative AI news this week. You probably heard the overhyped headlines about the thousands of startups that were “killed” by the magic spells of the OpenAI announcements 😉, so there is no need to discuss that nonsense. What I find more interesting is the strategic move of OpenAI trying to position themselves in higher layers of the generative AI stack, such as agents, distribution marketplaces, and several others. Given its dominant position in the generative AI market, we can trace parallels with another tech giant, Apple, after the iOS and app store launch in 2007 and 2008, respectively.

The comparison with Apple is relevant, given the similarities in approaching a new market and compute platform (mobile phones, LLMs) with a close tech stack across different layers of the stack. That strategy certainly worked out for Apple and seems somewhat logical for OpenAI, but there are a few interesting differences:

1. The generative AI market is evolving multiple times faster than mobile was in 2007. In that sense, the landscape is likely to change, and picking the wrong distribution strategy can be fatal.

2. While apps were the only user interface for mobile, it is unclear whether agents are the right interface for foundation models. I am super invested in the agents' ecosystem, but nobody has achieved critical mass in that market. Other modalities might emerge as the ecosystem evolves.

3. Apple was mostly a consumer play, while OpenAI has aspirations for both consumers and enterprises. Only a handful of companies have been able to execute in both models at the same time successfully.

4. Mobile was a two-horse race between Apple and Google almost since the beginning. It is pretty clear that the generative AI landscape is quite more complex.

I think it's fair to say that OpenAI’s position in the market has changed in the eyes of many concerned startups and investors. But if you look at the history of tech, OpenAI’s position is not more dominant than IBM, Intel, or Cisco at their peaks in their respective markets. Given the frantic pace of the generative AI space, we are likely to experience many more aggressive moves and landscape changes. It’s what makes this market so fascinating. For now, OpenAI does look like Apple in 2008, but that might change tomorrow.

---

📌 ML Engineering Event: Join Meta, PepsiCo, RiotGames, Uber & more at apply(ops)

apply(ops) is in two days! Join HelloFresh, Lidl Digital, Meta, PepsiCo, Pinterest, Prima, Remitly, Riot Games & Uber on Tuesday, November 14, at this virtual event to learn best practices from ML practitioners and connect with your peers around the world.

Databricks’ CEO Ali Ghodsi will also be joining Tecton CEO Mike Del Balso for a fireside chat about LLMs, real-time ML, and other trends in ML. 

Register today—it’s free!

---

🔎 ML Research

LLM Pruning

Microsoft Research published a research paper proposing LoRAShear, a technique for pruning and recovering knowledge in LLMs. The method uses LoRA to understand the knowledge structure of an LLM including which sections to prune —> Read more.

Coding Relevance in the Era of Generative AI

Microsoft Research published a paper discussing the impact of generative AI in end user programming. The paper outline a pragmatic thesis as of why programming will still be relevant despite the evolution of coding language models —> Read more.

Modern Clustering

Google Research published a paper detailing a new clustering algorithm. The technique combines the scalability of embedding models and the simplicity of CA methods —> Read more.

Distill-Whispter

Hugging Face published a paper discussing Distill-Whispter, a smaller and highly efficient version of the Whisper model. Distill-Whisper is almost 6 times faster while maintaining performance within 1% or the original model —> Read more.

ReAugKD

Amazon Science published a paper proposing ReAugKD, a retrieval-augmented framework that uses a teacher-student model to improve the performance of models. In ReAugKD, the embeddings and predictions created by the teacher model are used to guide the student model —> Read more.

🤖 Cool AI Tech Releases

OpenAI Announcements

OpenAI unveiled a series of new capabilities in terms of models, tools and developer toolkits at their DevDay event —> Read more.

PromptIDE

Elon Musk’s xAI announced PromptIDE, a development environment for prompt engineering —> Read more.

🛠 Real World ML

Video Search at Netflix

Netflix discusses some of the components of the architecture powering its video search engine —> Read more.

📡AI Radar

• Kai-Fu Lee’s new AI startup 01.ai unveiled a new open source language model and a fresh $1 billion valuation.

• Also 01.ai accumulated quite a bit of compute power before the chip ban.

• The highly anticipated Humane’s AI pin

• Elon Musk’s xAI unveiled some details of its Grok model.

• AI coding startup Replit raised another $20 million to provide liquidity to early employees.

• Another AI coding startup, Tabnine raised a $25 million series B.

• Stability AI raised a new financing round led by Intel.

• Alepth Alpha, a German AI startup focused on enterprise and goverment workloads raised $500 million in new funding.

• The former co-founder of Flikpart seems to be incubating a new AI startup.

• AI video platform Ozone raised $7.1 million in new funding.

• Microsoft updated its startup program to entice companies with GPU access.

• Cloud cost optimization startup Cast AI raised $35 million series B.

• Integration.app , an startup that leverages AI for workflow automation, emerged out of stealth model with $3.5 million in funding.