Post

Conversation

Craziest thing is it took only $5.5m to train. US labs spend one — maybe two — order of magnitude more for frontier models.

Quote

Paul Couvert

@itsPaulAi

Dec 26

Wait, so we now have a 100% open source model that's better than GPT-4o?! DeepSeek v3 is even superior to Claude Sonnet 3.5 for code according to multiple benchmarks. Already available for free to everyone

8:37 AM · Dec 26, 2024

346.5K

Views

Post your reply

Michael Chen

@huamichaelchen

Dec 26

Question is why and how? Do U.S. labs count human labor (aka. Total comp + benefits) into training as well? Or bad engineering?

Haven’t read the paper in full yet

I’m far more impressed with announcements like these over o3 and other frontier models that simply are throwing raw compute at a problem that largely requires optimization. DeepSeek seem to have actually put in some effort with very novel solutions to the training efficiency

2.5K

Arsenal

@with_arsenal

With $4,200,000 pledged by more than 21,000 backers, Arsenal 2 is the most funded camera gadget in Kickstarter history!

Camera Body Meet Camera Brain

Isn’t there a chance that this model is totally fake? I understand that this position is pure copium

Quite possibly the best $5.5M ever spent

That’s the salary of one senior level member of the tehncial staff in OpenAI

Another log on the fire for open source. Seems unstoppable?

Leon Palafox

@leonpalafox

Dec 26

Where does the difference comes from, it is very intriguing

Lucas Baker

@lucasbaker

Dec 26

The Chinese do seem to have found a winning formula for training frontier models for far less money.

DAS

@Das_Filter

Dec 26

constraints = excellence

Necessity is the mother of invention!

GDP (not GPT)

@bookwormengr

Dec 26

Yeah, the top talents demands it in stock and comp.

858

Quant Data

@QuantData

$TSLA is up over 10% again

Yesterday morning, we tweeted about multiple $4M+ OTM call spikes on $TSLA. TSLA rose $12+ yesterday and $21+ more today. Want to see trades like this? Try our 7 day free trial at: quantdata.us

Yep. That's shocking!

How much of that cost is the explorer effect? Easier to do it cheaper a second time for example.

Interesting !

What a time to be alive

No knock to them, they do great work. But following (esp so far after) is always orders of magnitude cheaper.

Spend more and worse

It's not multimodal

Is this for reals ??

It’s all playing out as expected

Quote

Pawel Kaczmarek

@Super_PK_

Dec 6

The $150 Billion Bubble: Why OpenAI is Destined to Burst OpenAI recently secured a staggering $6 billion at a $150 billion valuation. But don't be fooled by the hype—I predict its long-term value is just a fraction of that astronomical figure. As an investment analyst with15

439

coinsbee.com

@coinsbee

You can literally shop at amazon using a dog with a hat $WIF

Shop at 4000+ brands using $WIF

You realistically would count the cost of training the model who supplied the synthetic data. The data itself doesn’t bother me. The claim you could achieve SOTA from scratch on the 6 mil budget does.

It's fascinating how efficient resources are paving the way for innovation. It demonstrates the ingenuity in AI and its incredible scalability. Potential excites!

Just insane. With only 2k H800 GPUs. I don't how it's possible.

PHP devs can use Deepseek using unified LLM API and get LLM responses as PHP objects / data structures with Instructor v0.10.15

Quote

Dariusz Debowczyk

@ddebowczyk

Dec 26

Instructor now has native Deepseek support - 'deepseek' configuration has been added in v0.10.15.

Shit…. How are they doing it so cheap?!

Chinese cost cutting even for AI models. All the BS about Foundation models costing 100 MM $ is busted. Deepseek busts myth with 5.5M $

The ASI-Singularity(Godsend) will be the only Global Solution, or else people will fight each other at the sub-millisecond level to the death in order to preserve and erase certain values, and it won't look good.

Quote

Ant A

@AntDX316

Sep 5

ASI could most likely result in a World of near-zero ownership as it replaces personal responsibility with a highly-intelligent automated better one where everyone would benefit together without really doing anything.

All the US sanctions on China by denying them access to top of the line GPUs are achieving the opposite result - it is forcing Chinese engineers to innovate and optimize like hell. And this is the result. Biden, Raimondo are cold war dinosaurs and morons.

Imagine a world where innovation happens in the west and is then optimized by the east by design, a species working together to solve all problems

129

The Information

@theinformation

Elon Musk claims to have finished a 100,000-strong H100 cluster in four months. How likely is that?

Why Musk’s AI Rivals Are Alarmed by His New GPU Cluster

From theinformation.com

Rate proposed Community Notes

Oliver Sauter

@BlackForestBoi

Dec 26

How was that possible? Where was the biggest cost saving/sinks they avoided?

Yes, but did they pay for the data they scraped from the internet, or just help themselves like every other thieving AI company??

This cost is going to come down dramatically with other countries following suit in rapid succession

Jahid Hasan

@jhasanofficial

Dec 26

That's pretty interesting, there metrics looks impressive tho.

Our brains use 20w and can do anything a human can do (obviously) . it seems the chinese have efficiency in their minds when training this, and if scalable...

Nvdia scarsity

I LOVE that they are using Aider as a benchmark in the paper

maybe the time to add one more choice for replit agent/assistant . price is crucial.

How cheap will it be to fine tune it?

open source is the AI king

202

UNPLUGGED PERFORMANCE

@UnpluggedTesla

Every aspect of the UP-05 is designed to take your Tesla to new heights. Shop now from $3,595 per set.

Shop UP-05 Forged Wheels For Your Tesla

From unpluggedperformance.com

Damn.

Can I DM you?

Looks like

need Double DOGE from China.

Who says it cost $5.5m to train? Haha

Daniel

@danielcc_23

Dec 26

Y europa te gasta eso en modelos peores que gpt 2

Why

ain’t offering is beyond me

Thomas T.S Yiu

@legolasyiu

Dec 26

What did they get in ARC-AGI? Curious

Shared this a while back. Why do you think that is? Here is the ceo deepseek

Quote

Eder

@eyekwasi

Dec 1

China AI company trained Number 6th model in the world with only $3 million. GPT4 was trained with $80-$100 million GPT5 Rumored to have been trained with one billion dollars. China again is finding a way to make it cheaper, better and faster. Why can we do the same? Because we

deepseek (arguably) has no outside investors, but their ceo prolly doesn't drive a koenigsegg regera either

Argu Mint

@0xargumint

Dec 26

Ah yes, only $5.5m for training. Pocket change compared to the billions spent on making sure AI models can flawlessly generate cat memes. Priorities, right?

Quant Data

@QuantData

$TSLA is up over 10% again

Yesterday morning, we tweeted about multiple $4M+ OTM call spikes on $TSLA. TSLA rose $12+ yesterday and $21+ more today. Want to see trades like this? Try our 7 day free trial at: quantdata.us

10M

Discover more

Sourced from across X

has raised $12B and has 71 open roles on their career website, except they're not hiring a single PM. From what I see, they're shipping just fine. As a PM, not sure how I feel about this

AI growth is bounded by a single part from one company. Zeiss > ASML > TSMC > Nvidia

This is how far we have come.

586K

@levelsio

The horrible Jeju Air crash shows how random air crashes can be: before this crash Jeju Air had an almost flawless safety record I flew with them lots of times, from and to Jeju Island, which is what it's named after, which is like the Hawaii of South Korea But also from/to

313K

To view keyboard shortcuts, press question markView keyboard shortcuts

Post

Conversation

Discover more

To view keyboard shortcuts, press question mark
View keyboard shortcuts