Homescreen
Posts
🏠 A Rare OpenAI L

🏠 A Rare OpenAI L

You don't see this everyday

Stephen Flanders
October 18, 2023

Gm. Homescreen house party tonight. I’ll see you there 🤝

— Stephen Flanders

🏠 AI

Down Goes Arrakis

Since the release of ChatGPT last November, OpenAI hasn’t taken many L’s.

ChatGPT became the fastest-growing consumer app of all time, GPT-4 came out in March, OpenAI’s ARR hiked to $1.8B, and the company is now valued at a smooth $30B. That’s a miraculous turnaround for a business that only pulled in $28M in revenue in all of 2022.

But it hasn’t been all smooth sailing for OpenAI. It’s very expensive to train and operate AI models, in large part due to the AI chip demand surge they set off with ChatGPT. So, shortly after ChatGPT was released, OpenAI began training a new AI model designed to run the chatbot more efficiently.

That model was codenamed Arakis. The goal of Arakis was to use a concept known as sparsity to run more efficiently. Basically, instead of needing the whole model to answer prompts, sparsity enables the model to operate with just the parts of the model immediately needed. In theory, this leads to greater efficiency, and thus, lower costs and less need for chips.

It sounds great, except for one problem: it failed. And it failed for reasons that OpenAI still doesn’t understand. So they pivoted to developing a version of GPT-4 designed to generate responses quicker.

But that doesn’t mean sparsity isn’t a thing anymore. Many researchers still see it as the future of AI models.

And with advancements like Anthropic’s mechanistic interpretability breakthrough, it might not be too long until sparsity is a reality.

The Big Picture: With competition ramping up, OpenAI can’t afford too many more L’s.

🤝 THE LATEST IN…

TECH

You can now stop Instagram from relentlessly tracking you.
The latest FSD demo dropped.
YouTube is really cracking down on ad blockers.
The Meta glassholes have arrived.

AI

Wall Street doesn’t really understand Meta’s AI approach.
MemGPT creates perpetual chatbots with self-editing memory.
GPT-4 still has flaws.
Chinese tech giant Baidu claims their new ERNIE 4.0 model is as good as GPT-4.

SCIENCE

Figure’s humanoid robot is already getting scarily good.
China says it’s discovered a bunch of new ore.
This superficial brain implant could have a big impact.
Scientists are working on gene-hacked plants that suck up extra CO2.

CRYPTO

Reddit is killing the blockchain-based community points.
Asia is quickly becoming a crypto haven.
FTX plans to return 90% of customer funds.
Binance.US is halting dollar withdrawals.

🏃‍♀️ QUICKIES

Raise: Nuclear fusion company SHINE raised $70M.

Stat: $30,000: What Bitcoin reached following the Cointelegraph fake ETF fakeout. For those unware, Cointelegraph reported that a spot Bitcoin ETF was approved, which then led Bitcoin to shoot up to $30K just 10 minutes later. Unfortunately, it turns out that they prematurely shot the gun, and there’s no ETF yet. Still a very good sign for Bitcoin’s eventual resurgence.

Rabbit hole: The Techno-Optimist Manifesto (Marc Andreessen)