k__

My 2 weeks with DeepSeek V4:

Pro is ~50% more expensive than Flash.

Both need babysitting.

Plan, split in small tasks, give it docs, types, tests, linter, best practice examples, etc.

Always start a new session when starting a task.

Do regular manual sanity checks, and tell it to find issues in the codebase.

I pay like $1,50 per day for Pro.

Thanks!

What is the parento frontier?

Half-OT: can anyone recommend a LLM cost calculator that's up to date?

Effect's use of generators is also nice. While looking a bit off, it allows writing imperative business logic.

Are they still using eink?

"Fully open source model" with "Long-term, license-free availability for industry"

Nice trick to be not comparable with most other LLMs on the market, open weight or proprietary.

But I think, that's the right way.

I tried Hy3 today and liked it. It's a (small) step up from DSV4P.

Something on that level but multi-modal would be quite nice!

While it's true that unsupervised AIs build a heap of trash, I have to admit, it doesn't look much worse than some of the code I saw powering products that sold for millions.

The only thing that changed is the scale.

Strange, there aren't even 20 at a time above Germany.

How many uplinks can one satellite handle?

Pi

Yeah, DeepSeek V4 (Flash and Pro) is below $0.004 for 1M cache hits.

Even with usage based billing I'm below $1 writing code all day.

"Europe's rapid warming is partly the result of [...] a drop in the number of tiny polluting particles in the air. This means that less of the Sun's energy is reflected back into space, leaving more energy to heat the Earth's surface."

Hmmm...

I tried DS4 Pro yesterday for one particular task that Flash struggled with.

That one Pro task cost $1.50 and the whole day of using Flash cost me $5.

Thanks!

I'll try that.

Yeah, v4 flash is dirt cheap, but it's running in circles quite often.

Might very well be that a better model is cheaper if it gets things right the first try.

Maybe I should route to a better model when v4flash hasn't solved after a specific number of tokens.

Nice.

I paid $6 yesterday for DeepSeek V4 Flash on OpenRouter. That's like $120 dollar for a month, and it's not even a good model.

Llama3.1 instruct seems to be doing okay on that page, mostly because it's dirt cheap.

So, their incentive is to promote AI music, since they don't have to pay royalties for them.

Did Nim 3 go all-in on compile time memory management?

I'm curious about Denuvo's opinion on that.

Maybe, you could pipe it through T5 or something.

Training DeepSeek was magnitudes cheaper than training the SOTA models it relied upon.

In theory, other countries should be able to replicate that effort and improve it.

As I understood it, they were only a US thing anyway. Because of healthcare issues etc.

I tried some smaller Gemma4 and Qwen3.6 quants on my MBA with M5/16GB and had like 20-60 tokens per second. At 60 it felt pretty okay and that hardware is on the lower end.

I'd assume a Mac with 32-64GB memory would get some reasonable results.

I thought the joke was that people aren't paying enough money.

100K seems quite much.

I had the impression, models would get inconsistent after just 3000 words.

Why doesn't Mistral distill?

I tried it with embedded programming, and failed miserably.

200ml every waking hour?

Seems excessive.

Ah, so it a "smart" retry mechanism?