PLUS: The 22-year-old AI billionaires, Meta's unsolvable flaw, and Hinton's warning
Good morning
Alibaba is making a major play in the open-source community with its new web agent, Tongyi DeepResearch. The company claims the agent's performance is comparable to leading proprietary systems.
By releasing its complete end-to-end training methodology, Alibaba is giving developers a full playbook for creating advanced agents. Could this comprehensive approach provide the open-source world with the tools it needs to truly compete with closed-off AI development?
In today’s Next in AI:
Alibaba's new open-source agent
The 22-year-old AI billionaires
Meta’s unsolvable AI security flaw
The AI attacks bypassing modern defenses
Alibaba's Open-Source Agent
Next in AI: Alibaba has released Tongyi DeepResearch, a new open-source web agent that it claims performs on par with top proprietary models. The team also shared its entire end-to-end training methodology, giving developers a full playbook for creating advanced agents.
Decoded:
Instead of just releasing a model, Alibaba shared its entire training pipeline, which uses fully synthetic data to train agents from pre-training through reinforcement learning.
The agent operates in two powerful modes: a simple ReAct mode for direct tasks and an advanced "Heavy Mode" that uses an iterative process to tackle complex, long-horizon research without information overload.
This technology is already powering applications inside Alibaba, including a travel-planning agent for its mapping service and a legal research assistant that automates attorney workflows.
Why It Matters: By open-sourcing a complete and battle-tested methodology, Alibaba provides a powerful alternative to closed-off agent development. This release equips the open-source community with the tools to build its own high-performance agents, potentially accelerating innovation across the industry.
The 22-Year-Old AI Billionaires
Next in AI: Three 22-year-old co-founders of AI recruiting platform Mercor just became the world's youngest self-made billionaires. A $350 million funding round valued their company at $10 billion.
Decoded:
Mercor pivoted from a basic freelance marketplace to providing human-in-the-loop services that help AI labs refine their models through data labeling and quality control.
The platform now serves major AI players including OpenAI, Anthropic, and Google DeepMind, connecting them with vetted contractors for critical AI training tasks.
The company's valuation skyrocketed from $250 million in late 2024 to $10 billion today, driven partly by Meta's acquisition of competitor Scale AI.
Why It Matters: Mercor's meteoric rise proves that the biggest AI opportunities extend beyond building models. The infrastructure supporting AI development, particularly human-powered data refinement, represents a massive and growing market.
AI's Unsolvable Flaw
Next in AI: Meta AI has acknowledged that prompt injection attacks remain an unsolved problem, and has proposed a new framework called the "Rule of Two" to help developers build safer AI agents.
Decoded:
The framework states an AI agent should never combine all three high-risk capabilities in one session: processing untrustworthy inputs, accessing private data, and taking external actions.
A separate research paper from top labs including OpenAI and DeepMind confirms the need for this, showing that adaptive attacks bypass most current defenses.
The study found that these advanced attacks achieved a success rate above 90% against 12 different security tools, with human red-teaming defeating every single one.
Why It Matters: This signals a major shift in AI security, moving from trying to block attacks to designing systems that limit their potential damage. For developers, this framework provides a practical guide to building safer AI applications today, even as the core vulnerability remains.
AI Pulse
Alibaba's chairman warned of a potential bubble in the AI datacentre market, pointing to speculative projects raising funds without customer commitments amid a projected $3T spending spree.
Geoffrey Hinton argued that to profit from their astronomical AI investments, tech giants will have to replace human labor, casting doubt on whether AI will create as many jobs as it destroys.
Anthropic launched Claude for Excel, joining a wave of new AI agents designed to automate complex spreadsheet tasks and transform workflows for non-technical users.
A developer reproduced the critical race condition behind a recent AWS outage using a model checker, providing a step-by-step technical breakdown of the subtle concurrency bug.