Feels like a faster and cheaper fable, which yes is a big jump.

A lot depends on how good open source models do. I know I can buy a 20k rig or whatever to get model X. If someone wants to charge me 3k a month to write code I just buy the hardware.

I’m interested: how would you spend $20k to get close to performance parity (speed and ability) with a frontier model?

Maybe mid priced is a better term for it lol.

Sure I can accept that lol

5T I believe is the guess, though not sure its public anywhere.

I believe most speculation is that Opus is 5T model, Fable a 10T

Elon claimed Opus was 5T in April, and I think it's fairly likely this is accurate: https://x.com/elonmusk/status/2042123561666855235

The difference in complexity between my go and my Python docker image is actually hilarious. 100s of files and package managers and npm and whatever else.. vs 1 file.

Where does NPM play in to the build?

Beats me but somehow we do!

Every day I drop 1k lines of Python code and replace with go. Will be fully go soon!

Npm always finds its way in, usually for building web artifacts.

Ditto. Nothing I want to play doesn’t work on Linux / bazzite.

In fact Atlas and their push to be a C-Corp is a bad play for us small timers

A C corporation can still make a lot of sense for a small business depending on the nature of the business and the interests and needs of the owners.

C corps can offer a lot of advantages in terms of shareholder and employee benefits, particularly around healthcare, and QSBS is a huge benefit for qualified businesses and owners who might one day sell the business.

...unless you're based in a jurisdiction like the UK, where an LLC may not get the same tax treatment as a C-corp (Ltd companies being the norm here).

The memory bandwidth and size seems to be there, but what is the tokens per sec on like a qwen model? And you can basically do 3x opus 4.5 on the $100 a month claude plan. Your payback will be near infinity years after electricity.

you are assuming the idea, the data, the code being touched by opus worths nothing.

how about measuring returns in the sense that I no longer need to share my flagship idea with some random 3rd party just because it hosts the LLM I am using?

At the price points they are hitting, can’t afford over 32 GB on the cheaper model.

256 memory gets you to like 11k. So like 15-20k.