[Premium] News on OpenAI, FigureAI, Apple, & More

TLDR;

Technology:

😍 Google’s AlphaChip

🧐 Sam Altman Pitches Biden

🥸 The Architect/Editor Framework

Markets:

🍎 Apple is Finally not Investing in OpenAI

🤩 ScaleAI Quadruples Revenues

🥵 Zuckerberg’s Comments on AI Investing

Product:

🫣 AI-Generated Posts, the Future of Social Media?

🦿 In Talks with FigureAI’s CEO

HARDWARE

Google Gives Insights into its AI Chip Designer

Google has released new information about AlphaChip, its superhuman Reinforcement Learning model that helps the company design the layouts of its new chips, including the new version of its TPU hardware, Trillium.

The model treats chip layout design as a ‘game’ and can find new layouts in hours, while humans can take weeks or even months. It can also reduce the wirelength, the distance of the wires connecting the chip components, by up to 6.2% compared to humans.

The model places chip components into a circuit to achieve the best possible result. By learning to place these components millions of times, the model develops intuitions on how to build a functional chip while optimizing for a set of constraints (reduced wirelength, throughput improvements, etc.), which yield higher rewards the more the model meets these goals.

Therefore, the complete process becomes a game, and eventually, by playing it millions of times, it becomes better than humans. A similar method was used to train AlphaZero to become the world's best chess/Go/shogi player, achieving an ELO no human can compete with.

TheWhiteBox’s take:

I consider Google to be the AI company that is most slept on. Crucially, Google is the only lab globally that has consistently deployed superhuman AI into real settings.

While a strong case can be made in favor of OpenAI regarding Large Language Models, Google Deepmind is still the largest AI lab in the world throughput-wise, and they are head-and-shoulders above the rest in everything besides LLMs.

Besides optimizing chip layouts, Google is already using RL-based algorithms to manage their data centers, achieving a PUE (Power Usage Effectiveness) of 1.1, a number other Hypersclares like Microsoft can only dream of for now.

Simply put, while OpenAI seems slightly ahead when training generalist models (foundation models), Google is the undisputed king of depth (achieving superhuman performance on specific tasks).

The bullish case for Google is that RL is clearly the only way humans know how to achieve superhuman AI performance on a task, and it’s also the secret beyond OpenAI’s o1 models. Added to their overwhelming compute and data superiority, it’s hard to make a case against Google.

HARDWARE

Sam Altman Pitches Biden

Continuing on hardware, Sam Altman, OpenAI’s CEO, is reportedly pitching the Biden Administration on building up to seven different 5 GigaWatt data centers, framing this need as a national security threat to maintain AI dominance against China.

TheWhiteBox’s take:

If these conversations are indeed true, the numbers we drew in our previous analysis of Big Tech’s Big Data Center Plan might soon be dwarfed.

35 GigaWatts of power for AI would consume an astronomical 306.6 TWh per year, more than most…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.