Liberty’s Highlights

Liberty’s Highlights

Share this post

Liberty’s Highlights
Liberty’s Highlights
366: OpenAI's Financials & Compute Costs, Nvidia, Snowflake + AWS, Cloudflare, Heat Pumps, and Batman
Copy link
Facebook
Email
Notes
More

366: OpenAI's Financials & Compute Costs, Nvidia, Snowflake + AWS, Cloudflare, Heat Pumps, and Batman

"hyper-social connectors like Lyndon B. Johnson"

Liberty's avatar
Liberty
Dec 16, 2022
∙ Paid
11

Share this post

Liberty’s Highlights
Liberty’s Highlights
366: OpenAI's Financials & Compute Costs, Nvidia, Snowflake + AWS, Cloudflare, Heat Pumps, and Batman
Copy link
Facebook
Email
Notes
More
7
Share

Let me never fall into the vulgar mistake of dreaming that I am persecuted whenever I am contradicted. —Ralph Waldo Emerson

✍️ TIL that Plato wasn’t a fan of the written word:

Plato mourned the invention of the alphabet, worried that the use of text would threaten traditional memory-based arts of rhetoric. In his “Dialogues,” arguing through the voice of Thamus, the Egyptian king of the gods, Plato claimed the use of this more modern technology would create “forgetfulness in the learners’ souls, because they will not use their memories,” that it would impart “not truth but only the semblance of truth” and that those who adopt it would “appear to be omniscient and will generally know nothing,” with “the show of wisdom without the reality.”

The more things change…

🤖🎸🥁🎶 Seth Forsgren and Hayk Martiros trained the vanilla StableDiffusion 1.5 model on labeled spectrograms of music. The resulting custom-trained AI model can generate new music in the styles that the user asks for with a text prompt!

This is amazing 🤯 but it makes perfect sense when I think about it.

From the model’s point of view, music and images are just patterns. Training on data to extract patterns is similar regardless of whether the data is perceived as audio or pictures by humans.

Of course, the original model wasn’t designed for this, but it gives us a glimpse of what a model specifically trained on a very large corpus of music could do. Especially if the output can be a full song rather than just a small section (due to the size limitation of SD 1.5’s image output).

I’ve speculated about similar things recently:

In the intro of edition #326 I wrote about a “DALL-E for music” and in edition #359 I wrote about “neural remasters’ of existing recordings (to improve the sound quality by having an AI “hallucinate” a better sounding version, not just do a typical remaster by fiddling with EQs and a few Pro Tools plugins).

🤔 If you could have any job in the world *that isn’t your current job*, what would it be and why?

Leave a comment

💡 Here’s a common misconception:

It is rarely necessary to wait 24 hours before filing a missing person report. When there is evidence of violence or of an unusual absence, it is important to start an investigation promptly.

Criminology experts say the first 72 hours in a missing person investigation are the most critical

So if I ever go missing and there’s evidence of violence, don’t wait! 😬


🏦 💰 Liberty Capital 💳 💴

📈💸 OpenAI’s Financial Projections + ChatGPT’s “Eye-watering” Compute Costs 🤖🤖🤖🤖🤖🤖

First, a few financial data points:

Keep reading with a 7-day free trial

Subscribe to Liberty’s Highlights to keep reading this post and get 7 days of free access to the full post archives.

Already a paid subscriber? Sign in
© 2025 Liberty RPF
Privacy ∙ Terms ∙ Collection notice
Start writingGet the app
Substack is the home for great culture

Share

Copy link
Facebook
Email
Notes
More