Open AI Challenges the Limits of Online Search | Large language models tutorial pdf free download | What is llm | Llm dataset format | Turtles AI

Hugging Face challenges OpenAI with an open source initiative for advanced web research. The Open Deep Research project aims to replicate the capabilities of Deep Research, integrating advanced models with an agentic framework. However, without an alternative to OpenAI’s proprietary LLM o3, the performance gap remains stark.

Key Points:

• Open Deep Research: Hugging Face’s initiative to create an open source alternative to OpenAI’s Deep Research.
• Agentic Framework: Using an advanced architecture to improve search and analysis capabilities.
• GAIA Benchmark: The project scored 55.15% in the test, lower than OpenAI’s 67.36%.
• Open Source Challenge: The lack of an equivalent of the o3 model limits competitiveness against proprietary solutions.

The recent competition between OpenAI and Hugging Face in the AI ​​research space has seen the emergence of an ambitious project: Open Deep Research, an initiative aimed at providing an open alternative to OpenAI’s powerful Deep Research system. The latter, revealed at a dedicated event, stands out for its ability to explore the web and generate detailed reports on a variety of topics, but is only accessible to users of the $200 monthly ChatGPT Pro plan. In response, a team of researchers from Hugging Face, including Thomas Wolf, has developed an open source agentic model capable of interacting with the web, reading files, and performing advanced analytics. However, while it leverages OpenAI’s o1 model for its initial phases, the project is limited by the lack of an open alternative to o3, the LLM underlying Deep Research, which stands out for its superior performance in managing complex information.

Open Deep Research is based on an agentic framework that guides the AI ​​model in the use of advanced tools, including a text browser and a file analysis module. This approach, combined with the use of code agents to optimize task execution, has improved performance compared to previous open source implementations. On the GAIA benchmark platform, used to evaluate advanced AI assistants, the project scored 55.15%, lower but still competitive with OpenAI’s 67.36%. The performance gap is largely attributable to the superiority of o3, which currently has no open source rivals capable of matching it.

The integration of code agents represents a significant step forward: using code to orchestrate actions allows for greater efficiency compared to traditional JSON-based formats, reducing the steps required and optimizing computational costs. Furthermore, the use of modular tools such as Microsoft Research’s Magentic-One has facilitated the development of a system capable of autonomously navigating the web, extracting information, and performing detailed analysis. However, to achieve parity with Deep Research, the project needs improvements, including the adoption of a visual browser and more advanced interaction with external sources.

In parallel, Hugging Face launched Open-R1, an initiative to create an open-source alternative to DeepSeek’s R1 reasoning model. This project has already received widespread support from the community, collecting over 10,000 stars on GitHub in just a few days. Unlike R1, Open-R1 aims to make the training process and datasets used completely transparent, ensuring greater control and reducing the risk of bias. Funded by a Chinese fund, R1 has proven to compete with OpenAI’s o1 model in several benchmarks, offering solid performance in areas such as physics and mathematics thanks to its fact-checking capabilities.

As open-source AI research continues to evolve, the main challenge remains to bridge the gap with OpenAI’s proprietary models.

While Open Deep Research is an important step towards greater accessibility of advanced research tools, the need for an open LLM that can compete with o3 is important to ensure a real alternative to closed solutions.