What is the impact of DeepSeek on the AI sector? đŸ”„

Can only imagine how stressful their production engineers have it at the moment. The totally unprecedented overnight success bolstered by a massive hype train must have multiplied the demands on their infrastructure by several orders of magnitude. Very uncomfortable time for the rank engineers and all the pressure being applied by some very senior people and the expectation of customers.

Definitely a good reminder that it’s not just the capability of the model but of the infrastructure around it.

4 Likes

I don’t think they are that concerned. Their goal is AGI as they’ve stated several times. If their goal was competing with OAI, probably wouldn’t have released OS

He missed two points:

  • Measuring progress to AGI by the sudden drops of breakthroughs (in math, comp sci first)

  • AI’s first expert task should be the AI programmer, use the tool to build the tool

lol: " The 2B model outperforms the 72B model in OOD tests within just 100 training steps." GitHub - Deep-Agent/R1-V: Witness the aha moment of VLM with less than $3.

1 Like

@anon10827405

Lol what? Open source isn’t “right behind”, it’s literally LEADING the charge in most cases. Maybe do your homework before making such claims?

Linux runs most of the internet, powers 97% of smartphones (Android is Linux-based), dominates supercomputers (like, literally 100% of top 500), runs your smart TVs, powers most IoT devices, and even runs the Mars helicopters! Git revolutionized code collaboration to the point where even Microsoft (yeah, the Windows guys) had to embrace it and bought GitHub. The most cutting-edge AI research is happening in open source right now with projects like Stable Diffusion, LLaMA, and Whisper leading innovation. Kubernetes? Open source. Docker? Open source. PostgreSQL? Open source. The whole web runs on open-source tech like Apache, Nginx, and Node.js.

The whole “right behind” thing is laughably wrong when literally every piece of modern tech infrastructure is built on open source foundations. Even “closed” companies like Meta and Google rely heavily on open source and have to contribute back to stay competitive.

And “despite the walls put up”? Please. Open source BREAKS walls. That’s literally the point. While big tech is busy building their walled gardens, open source communities are out there sharing knowledge and pushing boundaries.

2 Likes

Context matters.
This topic is about language models and in this category it is perfectly reasonably to claim that the proprietary models are still in the lead.
Deepseek R1 is a super interesting release for the open source community but it has nothing to do with other types of software and AI models you mentioned.

1 Like

@vb

Actually, you missed the context. I was responding to RonaldGRuckus claiming “Open source has constantly shown that it’s right behind despite the walls put up” which is a general statement about open source, not specifically about LLMs.

But since you brought up language models even there your claim about proprietary models leading is debatable. Have you seen the recent benchmarks? Deepseek R1 outperforms GPT-4o in several key areas including coding, math, and reasoning tasks.

The open source community has gone from “no viable models” to “competitive with state-of-the-art” in less than a year. That’s not “being behind”, that’s catching up and overtaking at an unprecedented pace.

And again to respond to @anon10827405 where he mention “being able to see reasoning tokens is huge”.

Tokens is basically a chains of random characters, not readable text. So first, what you see is generated text, not token :roll_eyes:

And Deepseek actually show the reasoning text on their own client for showcase. But using API, it literally takes one line of code to hide the reasoning which is already neatly packaged in [reasoning] tags. That’s the whole point of structured output.

You can’t compare model if you don’t use models with API. That it.

And don’t forget that OpenAI have been existing for the last 10 years while Deepseek was founded 18 months ago.

1 Like

I am fully aware but decided to reply anyways to keep the discussion on-topic in a constructive manner.

Not agreeing with the opinion of other community members is perfectly fine. Putting other developers down because of differences in opinion is not.

My suggestion is to go back and edit the first post you made to remove the ad-hominem statements.

1 Like

Sure, open source is fundamental to a lot.

I’m not sure what you’re arguing here. It’s quite common for companies to take an open source project (or many), repackage it/them and then try to profit off it by making sure what they’ve made can’t be surpassed.

This is what I was referring to, bring on an OpenAI forum.

I’m not arguing that open source is useless. It truly is fundamental to a lot of the world

It’s a good thing we can map it out

Also, R1 is not better than SOTA OpenAI models. You really can’t be serious by comparing R1 to gpt4o.

It’s very strange, you’re making it seem like I’m bashing open source when all I’ve said is that despite the barriers put up, like hiding the reasoning tokens, open source has still managed to keep up with it

3 Likes

DeepSeek launched a tactical Nuke:


CODEI/O: Condensing Reasoning Patterns via Code Input-Output Prediction:

DeepSeek AI Introduces CODEI/O: A Novel Approach that Transforms Code-based Reasoning Patterns into Natural Language Formats to Enhance LLMs’ Reasoning Capabilities


Read full article: DeepSeek AI Introduces CODEI/O: A Novel Approach that Transforms Code-based Reasoning Patterns into Natural Language Formats to Enhance LLMs' Reasoning Capabilities - MarkTechPost

Paper: [2502.07316] CodeI/O: Condensing Reasoning Patterns via Code Input-Output Prediction

GitHub Page: GitHub - hkust-nlp/CodeIO: CodeI/O: Condensing Reasoning Patterns via Code Input-Output Prediction

1 Like