Open-source or not? DeepSeek’s R1 chatbot fuels debate in AI innovation

Open-source or not? DeepSeek’s R1 chatbot fuels debate in AI innovation


PARIS, Jan 31 — Chinese AI shooting star DeepSeek has made headlines for its R1 chatbot’s supposed low cost and high performance, but also its claim to be a public-spirited “open-source” project in contrast to closed alternatives from OpenAI and Google.

Here are things to know about a decades-old software development philosophy that remains relevant in today’s race for the most powerful AI.

What is open source?

Open source refers to the practice of programmers revealing the source code of their software, rather than just the “compiled” programme ready to run on a computer.

“It’s having all the elements to reproduce, modify, reconstruct and adapt” a programme or system, said Michel-Marie Maudet, chief operating officer of French open-source AI firm Linagora.

The California-based Open Source Initiative (OSI) says truly open-source software should also be free to redistribute, allow derived works and be free of a slew of potential licensing restrictions.

Over the years, this has inevitably clashed with private companies’ pursuit of revenue and intellectual property protection.

Nevertheless, many fundamental internet technologies such as the Linux operating system and Apache web server are open-source projects.

Consumer offerings include Mozilla’s popular browser Firefox and VLC media player.

Navigating some open-source software can be challenging for non-specialist users.

And keeping projects up to date across a loose network of collaborators can be tricky, especially with funding hard to come by.

Which AIs are open-source?

DeepSeek’s R1 offers similar text-generating functionality to western rivals including OpenAI’s ChatGPT, Anthropic’s Claude or Meta’s Llama.

Like them, R1 is based on an AI large language model (LLM) — trained by ingesting vast quantities of text to ape natural human language.

But Meta, DeepSeek and France-based AI developer Mistral claim to set themselves apart by allowing developers free access to their tools.

DeepSeek, like other startups, used open-source tech to kick off its own development of R1.

“DeepSeek has profited from open research and open source… they came up with new ideas and built them on top of other people’s work,” Meta’s AI research chief Yann LeCun posted on Threads.

Is AI really open?

But the complexity of AI systems means the old model of open-source software cannot be mapped directly onto them.

In the AI field, “open source” is “often… more aspiration or marketing than technical descriptor,” researchers David Gray Widder and Sarah Myers West and Signal chief Meredith Whittaker argued in a 2023 paper.

“Code and documentation can’t tell you exactly how a model will perform in a given context, or enable you to predict the system’s emergent properties,” they added.

AI’s ingredients also include vast quantities of training data and the “weights” that emerge from the training process — which define how the system will respond to users’ prompts.

“The release of DeepSeek-R1 is an amazing opportunity for the open-source community, (but) they didn’t publish everything,” said Thomas Wolf, co-founder of open-access AI platform Hugging Face.

Outsiders were especially missing “the datasets and the code used to train” the AI model, he added.

Some lesser-known, fully open AI platforms do exist, such as Bloom, born from a collaborative effort dubbed BigScience, or Linagora’s Lucie chatbot, which stumbled with obviously wrong answers in a recent public release.

‘Inspire the next generation’

Even without the full details, “DeepSeek may enable other researchers to create AI applications and deliver innovations previously only possible in labs run by large companies,” said Wolfgang Meyer, associate professor at the University of South Australia.

The model showed off “the power of open source in action”, said Wolf from Hugging Face.

He pointed to 700 models already built atop the Chinese one and “a host of companies, researchers and users getting used to this model… adapting it and testing it on new use cases”.

DeepSeek’s accompanying research paper offers “several very smart, relevant recipes that will inspire the coming generation of our models,” Maudet said.

It will become “one of the fundamental elements” for creating more widely-used AI applications that will generate a true “generative AI economy”, he predicted.

But he added that the model was also inevitably “yet another tool of geopolitical, geostrategic influence” between the United States and China.

Even limited openness helped “create buzz” and ensure global publicity, Maudet pointed out. — AFP



Source link

OR

Scroll to Top