Hybrid thinking is the end of the line? Former Qwen leader explains why agents are the future of AI

Junyang Lin, former technical leader of Alibaba's Qwen project, spared no criticism of the so-called 'hybrid thinking' — an approach that combines symbolic and connexionist reasoning into language models. In a recent lecture and a detailed post, he stated that this direction is a strategic error. For Lin, the natural evolution of artificial intelligence is not to improve generalist models, but to build agents capable of performing complex tasks autonomously.

Why did hybrid thinking fail, according to Qwen's former leader?

Junyang Lin argues that hybrid thinking attempts to unite two fundamentally different paradigms without solving each other's limitations. While purely connecting models already demonstrate emerging reasoning on a large scale, the addition of symbolic components creates engineering bottlenecks and hinders scalability. "It's like trying to put wheels on a boat -- you end up with something that doesn't navigate well or walk well," illustrated the former leader during his presentation.

The central criticism is that the hybrid focuses on improving the quality of the internal reasoning of the model, but neglects the most important step: the ability to act in the real world. For Lin, the real jump will come when models stop being just 'brains' and gain 'hands' — that is, when they are trained to interact with tools, APIs and dynamic environments.

What are AI agents and how do they differ from traditional models?

AI agents are systems designed to perceive the environment, make decisions and perform actions autonomously to achieve objectives. Unlike traditional language models, which only generate text in response to prompts, an agent can, for example, research information on the web, send emails, control IoT devices or execute code — all in sequence, without human intervention.

Lin argues that agent training should be based on long-term goals and reinforcement from environmental feedback rather than just predicting the next word. This view is aligned with recent movements by large technology companies. To learn more about how other companies are investing in infrastructure for AI, check out the article on building America from AI: NVIDIA leads local chip production and infrastructure.

How does Lin’s vision impact the ecosystem of AI agents in Brazil?

The criticism of one of the technical leaders of one of the largest open models in the world has direct weight for Brazilian developers and companies using Qwen or similar models. If the trend confirms the bet on agents, the focus of the AI teams in Brazil will need to migrate from fine tuning of models to agent engineering — that is, orchestration of flows, integration with APIs and design of autonomous systems.

In addition, Lin's position reinforces an already present debate in startups such as Mistral AI, which challenges OpenAI and symbolizes European technological independence: smaller and specialized models combined with active architectures can offer more practical value than a giant monolithic model.

Future of AI agents: promising progress with caution

Despite the enthusiasm, Junyang Lin recognizes that training reliable agents is still a challenge. Security issues, target alignment and computational cost are real barriers. However, he believes the path is inevitable. "It is not about abandoning language models, but putting them at the service of systems that actually do things," he concluded.

Frequently Asked Questions about AI agents

Will hybrid thinking be completely abandoned?

Lin believes that yes, in the long term, because the approach does not solve the fundamental problems of scalability and practical action that agents address more naturally.

Which companies are already adopting the agent approach?

Companies like OpenAI (with GPT-4 and plugins), Anthropic and Alibaba itself (via Qwen) already experience active architectures, although still incipiently.

How can a Brazilian developer start working with agents?

The first step is to study frameworks such as LangChain, AutoGPT or Qwen's own ecosystem for agents, and to experience the integration of models with external APIs and tools.

---