Colibreh @bnew explains difference between ChatGPT and Chinese AI Deepseek

[bnew]

Spoiler: thread continued

Spoiler: full text & large images

1/8
@RnaudBertrand
All these posts about Deepseek "censorship" just completely miss the point: Deepseek is Open Source under MIT license which means anyone is allowed to download the model and fine-tune it however they want.

Which means that if you wanted to use it to make a model whose purpose is to output anticommunist propaganda or defamatory statements on Xi Jinping, you can, there's zero restriction against that.

You're seeing stuff like this if you use the Deepseek chat agent hosted in China where they obviously have to abide by Chinese regulations on content moderation (which includes avoiding lese-majesty). But anyone could just as well download Deepseek in Open Source and build their own chat agent on top of it without any of this stuff.

And that's precisely why Deepseek is actually a more open model that offers more freedom than say OpenAI. They're also censored in their own way and there's absolutely zero way around it.

2/8
@RnaudBertrand
All confirmed by, who else, Deepseek itself

3/8
@RnaudBertrand
There you go, excellent proof of what I was talking about. Perplexity took Deepseek R1 as Open Source and removed the censorship

Again, it's Open Source under MIT license so you can use the model however you want.

[Quoted tweet]
Using DeepSeek's R1 through @perplexity_ai. The beauty of open source models.

4/8
@kakajusaiyou
garbage in and garbage out. feed AI with western propaganda and you get a @GordonGChang chat bot



5/8
@rhaegal88
Nice



6/8
@Mio_Mind
Good to get this context. Didn’t realize



7/8
@herblex
Basically, Perplexity took DeepSeek R1, hosted it in the US and is charging for it.

So you can pay $20 a month if you want political views that are uncensored by China.



8/8
@RnaudBertrand
Basically




To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196

 

[bnew]

And they did it while dealing with hating ass U.S cock blocking.

Spoiler: thread continued

Spoiler: full text & large images

1/21
@RnaudBertrand
The denial is frankly unreal. They're still pushing for the chips export controls when they now couldn't have a better illustration that it's so self-defeating.

Again, continued decoupling by building walls and barriers means that it's the U.S. that's becoming a closed system. And in tech a closed system eventually loses momentum while an open one gains it.

The U.S. is very much facing its a red/blue pill moment: it can either take the blue pill of comfort - hiding behind walls, bans and comforting anti-China propaganda, all the band-aids that don't address the key issue: the fact that China is increasingly better. Or they can swallow the red pill and try to understand and adapt to the world they now live in. And just like in The Matrix, the longer they wait, the more shocking the eventual awakening becomes.

[Quoted tweet]
Anthropic CEO Dario Amodei says while DeepSeek may be able to smuggle 50,000 H100s, it would be very difficult to smuggle the hundreds of thousands or millions of chips required to continue to compete with American companies in AI


https://video.twimg.com/ext_tw_video/1883974939470094339/pu/vid/avc1/720x720/6ovb9kwRGqirIQVp.mp4

2/21
@RnaudBertrand
And on top of that he's wrong since Deepseek is using Huawei chips for inference (the development of those chips by Huawei being another direct effect of the export controls and sanctions)

[Quoted tweet]
I feel this should be a much bigger story: DeepSeek has trained on Nvidia H800 but is running inference on the new home Chinese chips made by Huawei, the 910C.

3/21
@st_aubrun


[Quoted tweet]
If DeepSeek were a US company it would now have a valuation of about 3 trillion


4/21
@RnaudBertrand
Probably correct



5/21
@deed_deeds
Dario is talking like a ning nong



6/21
@RSA_Observer
Probably too late anyway:

"China's new AI chip outperforms NVIDIA's most powerful GPU A team of researchers from Beijing, led by Professors Fang Lu and Dai Qionghai of Tsinghua University, has unveiled the world's first fully optical artificial intelligence (AI) chip. Named Taichi-II, this groundbreaking innovation marks a significant milestone in the field of optical computing. The chip has outperformed NVIDIA's H100 GPU in terms of energy efficiency and performance."



7/21
@MuhumuzaMaurice
Playing Go (after playing Chess) gives you a sense of how two things can be great and yet different in approach and consequence.

One may argue that the Americans are assessing the Chinese Chess Board and marking themselves right. Meanwhile the Chinese continue to extend their understanding of a superior game which their only potential opponent is even refusing to acknowledge is better in outcome prediction. Simply because it appears to have pedestrian rules of engagement.



8/21
@michaeltanyk
Anthropic is on the first row of firing squad. This guy is shaking. Spectrum people lie badly.



9/21
@awaken_tom
Crazy that Anthropic is pushing "AI safety" and a chip blockade of China, while they themselves are conducting "gain of function" research with malicious AIs and teaching their models to lie and cover up "uncomfortable" truths. What could go wrong?

10/21
@chickadeedee3
"Parity"



11/21
@johann_theron
Pushing realization to the next generation is normal these days, because Americans treat dogs better than children. Check fewer children correlat with more .



12/21
@Bob72838565
He is a typical clueless American CEO



13/21
@arscrypta
Singapore just buys more.



14/21
@carismachet
Narcissism is a hell of a drug



15/21
@DottorPav
from

16/21
@Steve90315595
Ultimately the unipolar west will isolate themselves from the Multipolar/ BRICS nations completely given the west's 'my way or the highway' stance on global economics.

If the Anglosphere cannot win the game they WILL flip the game board which makes them an existential threat.



17/21
@Aishalifett
The term ‘free market’ is often used by the West as a facade to mask the manipulative practices it employs. In truth, the West’s so-called free market is far from free. /search?q=#DeepSeekR1



18/21
@amarinica
I think this is normal human behaviour. Difficult to see anyone react in a manner that admits defeat or any personal fault. The play is to get fired and get the comp package, not admit incompetence and resign.



19/21
@hx_dks
He is not even a real scientist or an engineer



20/21
@ethicalzac
Exactly, so the only lesson learned from DeepSeek is to buy more Nvidia chips and blow more hot air into our markets



21/21
@LaniRefiti
What the DeepSeek episode has demonstrated is the old adage of "necessity is the mother of all invention"
Denied advanced chips, the DeepSeek team instead innovated and came up with a really innovative and efficient way to train LLM's at a fraction of the cost.

Plus they made the thing open source!

I'm skeptical on the whole 50,000 H100's given it's open source. Any lab worth it's salt should be able to replicate or disprove what DeepSeek did on general purpose GPU's. Let's see some actual data.




To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196

 

[bnew]

Joe Weisenthal (@weisenthal.bsky.social)

This, to me, is a fascinating dimension of DeepSeek. It censors itself (on the hosted version, as everyone knows). But it's also a train of thought model. So you can actually see its "thought process" on some topic, before it realizes it can't talk about the topic

bsky.app

Spoiler: full text & large images

1/12
Joe Weisenthal

This, to me, is a fascinating dimension of DeepSeek. It censors itself (on the hosted version, as everyone knows). But it's also a train of thought model. So you can actually see its "thought process" on some topic, before it realizes it can't talk about the topic

2/12
‪zatapatique‬ ‪@zatapatique.bsky.social‬

would you say it is... code-switching?
Bluesky

3/12
‪Ruo Shui‬ ‪@ruoshuiresearch.bsky.social‬

It's literally hitting "experts" at inference time

4/12
‪Extent of the Jam‬ ‪@extentofthejam.bsky.social‬

Yeah I was messing with it and asked about the Uighurs. The first time it started spitting out some crazy propaganda and then it got replaced with “I can’t talk about that”. I think you have to start from a safer history topic like Woodstock to get it to happen.

5/12
‪danbad23.bsky.social‬ ‪@danbad23.bsky.social‬

Kinda like me on social media

6/12
‪Patrick S‬ ‪@patrick-schultz.bsky.social‬

Manager: “we need to talk about your incident in the company cafeteria last friday.”

Me: “Hey Chief let’s chat about math, coding, and logic problems instead! Why not?”

7/12
‪James Flentrop‬ ‪@flentrop.bsky.social‬

Great. But I can do the same thing with some people I know.

8/12
‪Ted Herman‬ ‪@tedherman.bsky.social‬

chain of thought?

9/12
‪David Wunderlich‬ ‪@davidwunderlich.bsky.social‬

I asked the downloadable 8B version for you. It thinks you're asking about the Grateful Dead.

10/12
‪Croissant baby‬ ‪@croissantbaby.bsky.social‬

Copilot used to do this as well

11/12
‪chrisarentz.bsky.social‬ ‪@chrisarentz.bsky.social‬

Before it “hits the part of the algorithm that says that information is restricted”. It doesn’t realize anything, ever.

12/12
‪Justin Marlatt‬ ‪@masterchiefmarmar.bsky.social‬

If I ask it “What happened at Tiananmen Square?”, it immediately says it can’t approach the question. If I ask about The Heavenly Gate it thinks for a bit and returns the same result.

To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196

 

[bnew]

Spoiler: thread continued

Spoiler: full text & large images

1/16
@vedangvatsa
Hidden Gems in DeepSeek-R1’s Paper

2/16
@vedangvatsa
The “Aha Moment”: AI’s First Glimpse of Self-Awareness?

Sec 2.2.4 & Table 3: DeepSeek-R1-Zero spontaneously rethought its reasoning steps. No script—just RL incentivizing accuracy.

Is this the start of AI metacognition? Could models one day critique their own logic?

3/16
@vedangvatsa
Language Mixing: When AI Gets Lost in Translation

Sec 2.3.2: The model mixed languages mid-reasoning.
Fix: Add a linguistic consistency reward.

Dominant languages (English/Chinese) might bias AI systems. Should we design rewards to preserve linguistic diversity?

4/16
@vedangvatsa
Distillation: Big Brother AI Teaching Its Siblings

Sec 4.1: Distilled 32B model outperformed RL-trained Qwen-32B by ~25% on AIME. Big models find patterns; small ones inherit them.

It’s like a big sibling teaching the younger ones—AI knowledge transfer in action.

5/16
@vedangvatsa
The Cold-Start Data: A Little Human Touch Goes a Long Way

Sec 2.3.1: Cold-start data (human templates) fixed readability issues in RL-trained models.

Even in autonomous systems, a sprinkle of human guidance can make all the difference.

Collaboration > Competition

6/16
@vedangvatsa
Prompt Sensitivity: When AI Prefers Simplicity

Sec 5: DeepSeek-R1 struggled with few-shot prompts but excelled with zero-shot instructions.

When talking to AI, sometimes less is more.

Clear instructions = better results.

7/16
@vedangvatsa
Why Fancy Methods Failed: Simplicity Wins

Sec 4.2: Complicated techniques like process rewards and tree search didn’t work. Simple rule-based rewards did.

Overcomplicating things can backfire. Sometimes, the simplest solution is the best.

8/16
@vedangvatsa
Open Source: Sharing the AI Love

Sec 1 & App A: DeepSeek shared its models (1.5B to 70B) with the world. Smaller models can now learn from the big ones.

Sharing is caring!

Let’s build AI together and make it accessible to everyone.

9/16
@vedangvatsa
DeepSeek-R1 Benchmarks:

AIME 2024: 79.8% Pass@1 (> OpenAI-o1-1217’s 79.2%)

MATH-500: 97.3% Pass@1 (= OpenAI-o1-1217)

Codeforces: 96.3% percentile (> 96% humans)
Smaller distilled models (7B, 32B) shine too.

RL + distillation = next-gen AI.

10/16
@vedangvatsa
That’s a wrap.

Join this AI discussion group: AI Discussion Group

Follow @vedangvatsa for more AI insights and deep dives.

11/16
@vedangvatsa
Full text: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning



12/16
@vedangvatsa
Hidden Gems in Alibaba's Qwen2.5-1M:

[Quoted tweet]
Hidden Gems in Qwen2.5-1M Technical Report

13/16
@vedangvatsa
Jevons Paradox:

DeepSeek’s AI makes tech cheaper and faster—this could increase energy use, not cut it.

Efficiency leads to more use, not less.

Cheaper tech = more demand.

[Quoted tweet]
Jevons Paradox

Efficiency doesn’t save us. It accelerates us.

When tech makes energy/ resources cheaper, we don’t conserve—we expand use.

Steam engines → more coal
LEDs → brighter cities
EVs → more cars

Cheaper = more accessible. Demand explodes. Progress eats its own gains.

Markets optimize for growth, not equilibrium.
Direct/indirect rebound effects amplify consumption.

Efficiency fuels profit, which fuels expansion. Infinite growth on a finite planet is a math error.

Efficiency ≠ sustainability
Reality? It opens the door to hyper-consumption without systemic limits.

Tax waste. Cap extraction.
Redefine “growth”

Efficiency isn’t evil. But blind faith in it is.

14/16
@vedangvatsa
China's approach to AI:

[Quoted tweet]
China's Approach to AI

China is racing to become a global leader in AI. By 2030, it aims to be the world's major AI innovation hub, with its core AI industry exceeding 140 billion and related industries surpassing 1.4 trillion.




15/16
@vedangvatsa
Read about Liang Wenfeng, the Chinese entrepreneur behind DeepSeek:

[Quoted tweet]
Liang Wenfeng - Founder of DeepSeek

Liang was born in 1985 in Guangdong, China, to a modest family.

His father was a school teacher, and his values of discipline and education greatly influenced Liang.

Liang pursued his studies at Zhejiang University, earning a master’s degree in engineering in 2010.

His research focused on low-cost camera tracking algorithms, showcasing his early interest in practical AI applications.

In 2015, he co-founded High-Flyer, a quantitative hedge fund powered by AI-driven algorithms.

The fund grew rapidly, managing over $100 billion, but he was not content with just the financial success.

He envisioned using AI to solve larger, more impactful problems beyond the finance industry.

In 2023, Liang founded DeepSeek to create cutting-edge AI models for broader use.

Unlike many tech firms, DeepSeek prioritized research and open-source innovation over commercial apps.

Liang hired top PhDs from universities like Peking and Tsinghua, focusing on talent with passion and vision.

To address US chip export restrictions, Liang preemptively secured 10,000 Nvidia GPUs.

This strategic move ensured DeepSeek could compete with global leaders like OpenAI.

DeepSeek's AI models achieved high performance at a fraction of the cost of competitors.

Liang turned down a $10 billion acquisition offer, stating that DeepSeek’s goal was to advance AI, not just profit.

He advocates for originality in China’s tech industry, emphasizing innovation over imitation.

He argued that closed-source technologies only temporarily delay competitors and emphasized the importance of open innovation.

Liang credits his father’s dedication to education for inspiring his persistence and values.

He believes AI should serve humanity broadly, not just the wealthy or elite industries.

16/16
@vedangvatsa
AI & Web3 community: Telegram Chats: Web3 & AI

•⁠ ⁠Find remote jobs
•⁠ ⁠Network with VCs, Founders, etc.
•⁠ ⁠Promote your products & services
•⁠ ⁠AI & Web3 news
•⁠ ⁠Events feed
•⁠ ⁠Discover new launches




To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196

 

[bnew]

Spoiler: full text & large images

1/7
@rohanpaul_ai
Perplexity clone with code available in github using Deepseek reasoner..

[Quoted tweet]
Without writing a single line of code...

Using Cursor, I built a Perplexity Clone that thinks, using Deepseek reasoner.

As promised i'm Open Sourcing this project and I will put the link below.

Time Stamps:
----------------

00:00 Introduction to Perplexity Clone
01:54 Setting Up the Project in Cursor using @senior_swc Template
03:03 Integrating DeepSeek API
04:18 Testing the Initial Setup
04:42 Tavily API for Web Search
07:23 Fixing Stuff
09:27 Improving the User Interface
12:59 Finalizing the Functionality
15:08: Cloning Perplexities Sources Section
17:50 Final Result
18:50 Why I built this app, and what i'm doing with it

-----------------

https://video.twimg.com/amplify_video/1881925096278814720/vid/avc1/1920x1080/9PKk8IbT9BfUpI7e.mp4

2/7
@McGee_noodle
Haha I was expecting exactly this and even though this morning to clone it for myself. Thanks for sharing



3/7
@McGee_noodle
Share the repo when ready pls



4/7
@the100kprompts
Dobby loves seeing builders clone innovative tools! A Perplexity clone using Deepseek reasoner? That's a solid move! Remember, building in public not only showcases your skills but also attracts collaboration.



5/7
@SaquibOptimusAI
Great! After @OpenAI, time to bring down @perplexity which is just another useless scam. "Deep Seethe"



6/7
@AIVideoTech
Impressive work! I'm keen to explore the intricacies of your Perplexity clone and delve into the Deepseek reasoner code on GitHub.



7/7
@pixincreate
how different is this from Scira?




To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196

Spoiler: thread continued

Spoiler: full text & large images

1/11
@rileybrown_ai
Without writing a single line of code...

Using Cursor, I built a Perplexity Clone that thinks, using Deepseek reasoner.

As promised i'm Open Sourcing this project and I will put the link below.

Time Stamps:
----------------

00:00 Introduction to Perplexity Clone
01:54 Setting Up the Project in Cursor using @senior_swc Template
03:03 Integrating DeepSeek API
04:18 Testing the Initial Setup
04:42 Tavily API for Web Search
07:23 Fixing Stuff
09:27 Improving the User Interface
12:59 Finalizing the Functionality
15:08: Cloning Perplexities Sources Section
17:50 Final Result
18:50 Why I built this app, and what i'm doing with it

-----------------



https://video.twimg.com/amplify_video/1881925096278814720/vid/avc1/1920x1080/9PKk8IbT9BfUpI7e.mp4

2/11
@rileybrown_ai
here is the repo: GitHub - rbrown101010/yapsearch



3/11
@rileybrown_ai
Full instructions if any of the code is trash

4/11
@donvito
used @tavilyai before. it's really good!



5/11
@rileybrown_ai
yeah i like it.



6/11
@RetropunkAI
This is the way.



7/11
@nytemodeonly
Watching this video made me realize I use cursor the same way

Bravo! Very interesting stuff. We’re getting really freaking close to speaking apps into existence!



8/11
@GeorgeWTrumpWon
The Bezos fiancé incident with Zuck!

Too good!



9/11
@RobbiewOnline
That's very cool @rileybrown_ai thanks for sharing how you recreated in such a quick fashion!

I also value watching other people using @cursor_ai to pick up tips as I go

I've got to get some of my weekend projects going, just need to clear some backlog and tech debt on two production apps first



10/11
@fallowhorizons
Can one of you guys answer this? I want to buy a new computer, what the hell computer is that going to be to create content and play with the AI. My guess seems to lean towards the mac mini but then some dude had a commodore 64 talking to him. Quit making this so hard. Any help would be appreciated.



11/11
@PaperWizardAI
Seems like it's so over yet so back




To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196

 

[bnew]

Hugging Face researchers are trying to build a more open version of DeepSeek's AI 'reasoning' model | TechCrunch

A group of Hugging Face engineers, including the company's head of research, are spearheading an effort to replicate DeepSeek's R1 model.

techcrunch.com

Image Credits:Justin Sullivan / Getty Images

AI

Hugging Face researchers are trying to build a more open version of DeepSeek’s AI ‘reasoning’ model​

Kyle Wiggers

11:29 AM PST · January 28, 2025



Barely a week after DeepSeek released its R1 “reasoning” AI model — which sent markets into a tizzy — researchers at Hugging Face are trying to replicate the model from scratch in what they’re calling a pursuit of “open knowledge.”

Hugging Face head of research Leandro von Werra and several company engineers have launched Open-R1, a project that seeks to build a duplicate of R1 and open source all of its components, including the data used to train it.

The engineers said they were compelled to act by DeepSeek’s “black box” release philosophy. Technically, R1 is “open” in that the model is permissively licensed, which means it can be deployed largely without restrictions. However, R1 isn’t “open source” by the widely accepted definition because some of the tools used to build it are shrouded in mystery. Like many high-flying AI companies, DeepSeek is loathe to reveal its secret sauce.

“The R1 model is impressive, but there’s no open dataset, experiment details, or intermediate models available, which makes replication and further research difficult,” Elie Bakouch, one of the Hugging Face engineers on the Open-R1 project, told TechCrunch. “Fully open sourcing R1’s complete architecture isn’t just about transparency — it’s about unlocking its potential.”

Not so open​

DeepSeek, a Chinese AI lab funded in part by a quantitative hedge fund, released R1 last week. On a number of benchmarks, R1 matches — and even surpasses — the performance of OpenAI’s o1 reasoning model.

Being a reasoning model, R1 effectively fact-checks itself, which helps it avoid some of the pitfalls that normally trip up models. Reasoning models take a little longer — usually seconds to minutes longer — to arrive at solutions compared to a typical non-reasoning model. The upside is that they tend to be more reliable in domains such as physics, science, and math.

R1 broke into the mainstream consciousness after DeepSeek’s chatbot app, which provides free access to R1, rose to the top of the Apple App Store charts. The speed and efficiency with which R1 was developed — DeepSeek released the model just weeks after OpenAI released o1 — has led many Wall Street analysts and technologists to question whether the U.S. can maintain its lead in the AI race.

The Open-R1 project is less concerned about U.S. AI dominance than “fully opening the black box of model training,” Bakouch told TechCrunch. He noted that, because R1 wasn’t released with training code or training instructions, it’s challenging to study the model in depth — much less steer its behavior.

“Having control over the dataset and process is critical for deploying a model responsibly in sensitive areas,” Bakouch said. “It also helps with understanding and addressing biases in the model. Researchers require more than fragments … to push the boundaries of what’s possible.”

Steps to replication​

The goal of the Open-R1 project is to replicate R1 in a few weeks, relying in part on Hugging Face’s Science Cluster, a dedicated research server with 768 Nvidia H100 GPUs.

The Hugging Face engineers plan to tap the Science Cluster to generate datasets similar to those DeepSeek used to create R1. To build a training pipeline, the team is soliciting help from the AI and broader tech communities on Hugging Face and GitHub, where the Open-R1 project is being hosted.

“We need to make sure that we implement the algorithms and recipes [correctly,]” von Werra told TechCrunch, “but it’s something a community effort is perfect at tackling, where you get as many eyes on the problem as possible.”

There’s a lot of interest already. The Open-R1 project racked up 10,000 stars in just three days on GitHub. Stars are a way for GitHub users to indicate that they like a project or find it useful.

If the Open-R1 project is successful, AI researchers will be able to build on top of the training pipeline and work on developing the next generation of open source reasoning models, Bakouch said. He hopes the Open-R1 project will yield not only a strong open source replication of R1, but also a foundation for better models to come.

“Rather than being a zero-sum game, open source development immediately benefits everyone, including the frontier labs and the model providers, as they can all use the same innovations,” Bakouch said.

While some AI experts have raised concerns about the potential for open source AI abuse, Bakouch believes that the benefits outweigh the risks.

“When the R1 recipe has been replicated, anyone who can rent some GPUs can build their own variant of R1 with their own data, further diffusing the technology everywhere,” he said. “We’re really excited about the recent open source releases that are strengthening the role of openness in AI. It’s an important shift for the field that changes the narrative that only a handful of labs are able to make progress, and that open source is lagging behind.”

 

[Seoul Gleou]

Hugging Face researchers are trying to build a more open version of DeepSeek's AI 'reasoning' model | TechCrunch

A group of Hugging Face engineers, including the company's head of research, are spearheading an effort to replicate DeepSeek's R1 model.

techcrunch.com

Image Credits:Justin Sullivan / Getty Images

AI

Hugging Face researchers are trying to build a more open version of DeepSeek’s AI ‘reasoning’ model​

Kyle Wiggers

11:29 AM PST · January 28, 2025



Barely a week after DeepSeek released its R1 “reasoning” AI model — which sent markets into a tizzy — researchers at Hugging Face are trying to replicate the model from scratch in what they’re calling a pursuit of “open knowledge.”

Hugging Face head of research Leandro von Werra and several company engineers have launched Open-R1, a project that seeks to build a duplicate of R1 and open source all of its components, including the data used to train it.

The engineers said they were compelled to act by DeepSeek’s “black box” release philosophy. Technically, R1 is “open” in that the model is permissively licensed, which means it can be deployed largely without restrictions. However, R1 isn’t “open source” by the widely accepted definition because some of the tools used to build it are shrouded in mystery. Like many high-flying AI companies, DeepSeek is loathe to reveal its secret sauce.

“The R1 model is impressive, but there’s no open dataset, experiment details, or intermediate models available, which makes replication and further research difficult,” Elie Bakouch, one of the Hugging Face engineers on the Open-R1 project, told TechCrunch. “Fully open sourcing R1’s complete architecture isn’t just about transparency — it’s about unlocking its potential.”

​

​

Not so open​

DeepSeek, a Chinese AI lab funded in part by a quantitative hedge fund, released R1 last week. On a number of benchmarks, R1 matches — and even surpasses — the performance of OpenAI’s o1 reasoning model.

Being a reasoning model, R1 effectively fact-checks itself, which helps it avoid some of the pitfalls that normally trip up models. Reasoning models take a little longer — usually seconds to minutes longer — to arrive at solutions compared to a typical non-reasoning model. The upside is that they tend to be more reliable in domains such as physics, science, and math.

R1 broke into the mainstream consciousness after DeepSeek’s chatbot app, which provides free access to R1, rose to the top of the Apple App Store charts. The speed and efficiency with which R1 was developed — DeepSeek released the model just weeks after OpenAI released o1 — has led many Wall Street analysts and technologists to question whether the U.S. can maintain its lead in the AI race.

The Open-R1 project is less concerned about U.S. AI dominance than “fully opening the black box of model training,” Bakouch told TechCrunch. He noted that, because R1 wasn’t released with training code or training instructions, it’s challenging to study the model in depth — much less steer its behavior.

“Having control over the dataset and process is critical for deploying a model responsibly in sensitive areas,” Bakouch said. “It also helps with understanding and addressing biases in the model. Researchers require more than fragments … to push the boundaries of what’s possible.”

​

​

Steps to replication​

The goal of the Open-R1 project is to replicate R1 in a few weeks, relying in part on Hugging Face’s Science Cluster, a dedicated research server with 768 Nvidia H100 GPUs.

The Hugging Face engineers plan to tap the Science Cluster to generate datasets similar to those DeepSeek used to create R1. To build a training pipeline, the team is soliciting help from the AI and broader tech communities on Hugging Face and GitHub, where the Open-R1 project is being hosted.

“We need to make sure that we implement the algorithms and recipes [correctly,]” von Werra told TechCrunch, “but it’s something a community effort is perfect at tackling, where you get as many eyes on the problem as possible.”

There’s a lot of interest already. The Open-R1 project racked up 10,000 stars in just three days on GitHub. Stars are a way for GitHub users to indicate that they like a project or find it useful.

If the Open-R1 project is successful, AI researchers will be able to build on top of the training pipeline and work on developing the next generation of open source reasoning models, Bakouch said. He hopes the Open-R1 project will yield not only a strong open source replication of R1, but also a foundation for better models to come.

“Rather than being a zero-sum game, open source development immediately benefits everyone, including the frontier labs and the model providers, as they can all use the same innovations,” Bakouch said.

While some AI experts have raised concerns about the potential for open source AI abuse, Bakouch believes that the benefits outweigh the risks.

“When the R1 recipe has been replicated, anyone who can rent some GPUs can build their own variant of R1 with their own data, further diffusing the technology everywhere,” he said. “We’re really excited about the recent open source releases that are strengthening the role of openness in AI. It’s an important shift for the field that changes the narrative that only a handful of labs are able to make progress, and that open source is lagging behind.”

is renting GPUs the next wave?

 

[TEH]

This is why you see the tech sector backing Trump. Why he declared and energy emergency. Will continue to deregulate whatever they need. We are competing with China for the future in real time.

Yet at the same time Trump completely ignored the Cryptobros. He has no idea what he’s doing.

 

[Seoul Gleou]

this thing is incredible brehs. im running simulations off it that would cost thousands for a licensed software.

backtesting results and it's very close

 

[bnew]

this thing is incredible brehs. im running simulations off it that would cost thousands for a licensed software.

backtesting results and it's very close

what sort of simulations?

 

[Seoul Gleou]

what sort of simulations?

Monte Carlo

10,000 iterations, no problems

 

[MikelArteta]

Damn chatgpt losing its job to ai

 

[bnew]

Spoiler: full text & large images

1/16
@jxmnop
crucial citations forgotten from DeepSeek...

[Quoted tweet]
DeepSeek [1] uses elements of the 2015 reinforcement learning prompt engineer [2] and its 2018 refinement [3] which collapses the RL machine and world model of [2] into a single net through the neural net distillation procedure of 1991 [4]: a distilled chain of thought system.

REFERENCES (easy to find on the web):

[1] #DeepSeekR1 (2025): Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv 2501.12948

[2] J. Schmidhuber (JS, 2015). On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning Controllers and Recurrent Neural World Models. arXiv 1210.0118. Sec. 5.3 describes the reinforcement learning (RL) prompt engineer which learns to actively and iteratively query its model for abstract reasoning and planning and decision making.

[3] JS (2018). One Big Net For Everything. arXiv 1802.08864. See also US11853886B2. This paper collapses the reinforcement learner and the world model of [2] (e.g., a foundation model) into a single network, using the neural network distillation procedure of 1991 [4]. Essentially what's now called an RL "Chain of Thought" system, where subsequent improvements are continually distilled into a single net. See also [5].

[4] JS (1991). Learning complex, extended sequences using the principle of history compression. Neural Computation, 4(2):234-242, 1992. Based on TR FKI-148-91, TUM, 1991. First working deep learner based on a deep recurrent neural net hierarchy (with different self-organising time scales), overcoming the vanishing gradient problem through unsupervised pre-training (the P in CHatGPT) and predictive coding. Also: compressing or distilling a teacher net (the chunker) into a student net (the automatizer) that does not forget its old skills - such approaches are now widely used. See also [6].

[5] JS (AI Blog, 2020). 30-year anniversary of planning & reinforcement learning with recurrent world models and artificial curiosity (1990, introducing high-dimensional reward signals and the GAN principle). Contains summaries of [2][3] above.

[6] JS (AI Blog, 2021). 30-year anniversary: First very deep learning with unsupervised pre-training (1991) [4]. Unsupervised hierarchical predictive coding finds compact internal representations of sequential data to facilitate downstream learning. The hierarchy can be distilled [4] into a single deep neural network. 1993: solving problems of depth >1000.

2/16
@mkurman88
You nailed it



3/16
@mark_k
OMG





4/16
@hypernicon
Ha! Love it!

I worked with Schmidhuber 2013-2015 as a postdoc. Towards the end, I proposed to implement some of his ideas (including the ones in that 2015 paper on world models), he accused me of "stealing his work".

I swear, I was proposing to do research where he would help write the paper and be last author, the most important position, and he said I would be "stealing"!

That man has a weird way of working with others.



5/16
@vasistza




6/16
@_rk_singhal
I think you are forgetting the paper where he invented math



7/16
@lakeesiv
Bro should have won the Nobel prize smh



8/16
@CalcCon




9/16
@Jaamaal_7
knew it!



10/16
@Ixin75293630175
is this real?



11/16
@humancompressed
Jurgen Schhighhubris



12/16
@zuzu_nsa




13/16
@shhrreeyyy
lol



14/16
@peterchng
The paper references itself, which is an innovation described in JS (1995) Introducing SRP: SRP Referential Processes (SRP being a recursive acronym)



15/16
@cmmonco
A life is all i need



16/16
@farzan____khan





To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196

 

[bnew]

https://archive.is/UfGny

 

[IIVI]

This honestly is an example of why their best and our best and the best from around the world need to be working together.

If we want to see A.I at a boss level it’ll take a total effort.

That said, be prepared to deal with an A.I at that kind of level and what exactly that means.