Why do you use so many different agents if I may ask?

I mainly use Claude Code, but I had an idea to build an orchestrator for coding agents, so I experimented with several

The robots will love this.

There are already some products for arms that do this.

We are so used to these numbers being thrown around in the AI era that something one needs a reminder that this is 13 billion, not million. Insane exit by HF.

80x its yearly revenue. Clearly they aren't buying HF for margins.

In fact, has HF ever wished to make money? They probably pay AWS more infra cost than they earn. The exit was planned all along.

Was looking for this comment, the tech world has gotten completely insane with ”valuations”. I would love to hear why it was 13 and not 5. It would still be completely insane at 5 billion, but someone though they should add another 8…

Opportunity cost for Nvidia to prevent HF from going public or being sold to someone else. It's an insight into Nvidia's market prediction and what HF told them they'd do if left alone.

My second brain is 500 open tabs of Notepad++

Excellent work!

Forgejo/Gitea is lightweight and easy to setup.

There is no reason to be on github anymore.

Your code will used for training and the platform enshitification has been accelerated by Microsoft. It would be wise to move to another platform or even better, host your own Gitlab/Forgejo/Gitea on a vps or at home.

Seems like that's the tradeoff with this model. Close to 27b intelligence while using less vram.

Won't the Chinese providers have to raise their prices as well due to the economics of serving inference at scale?

Yes. People keep thinking that Chinese models are free or near free for some reason.

In reality, they're raising prices too. I was looking at Kimi K3 prices on OpenRouter and it's nearly the same as Anthropic and OpenAI.

Buy an M5 max for $4k and you have a portable Deepseek 0731 for life.

But it won't last for life. It will probably last about 4 years so that equates to about 100/mo.

M1 is finishing it's 6th year of life and going strong ...

Have you been trashing the non-replaceabe SSD with constant AI workloads? Do you have any idea how much electricity it uses over it those 4 years?

You cannot wear out an SSD through AI inference alone. LLM weights are only read from the SSD, and read operations do not contribute to SSD wear. SSD wear is primarily caused by write and erase operations.

> Have you been trashing the non-replaceabe SSD with constant AI workloads?

Your other point is valid, but you're either vastly underestimating how many reads/writes a modern SSD can take or vastly overestimating how much output a typical LLM is capable of.

models are stuck in time. eventually the world moves forward and it is trained on too much obsolete data.

training cannot end for LLMs intrinsically. it's not some fixed cost. it's an ongoing one.

> eventually the world moves forward and it is trained on too much obsolete data.

This is why LLMs are never going to be AGI. Humans don’t become obsolete just because they age.

Humans do die eventually though, which is an even more severe form of obsolescense. I don't think they'll be AGI either, but that argument doesn't seem super compelling.

On the other hand you have the old saying "science advances one funeral at a time", indicating death solves at least as many problems as it causes.

(the point being that old, powerful professors have more than once blocked progress in their fields for decades. And only a funeral, eventually, solves the problem ...)