This year, AI agents have been on fire, with startups popping up like mushrooms, and Y Combinator alone has invested in 94 related startups this year.
It's no wonder that Andrew Ng said: "AI agent workflows are set to drive tremendous advances in AI this year – perhaps even more than the next generation of foundational models." This is an important trend, and I call on everyone working in AI to pay attention to it. ”
Leifeng.com noticed that another company recently made a high-profile appearance - this time to do things in the semiconductor field.
Founded by William Wang, the startup, called Alpha Design AI, aims to disrupt the traditional chip design and verification process and create a native AI-powered electronic design automation (EDA) system.
To date, Alpha Design AI has raised $3.09 million in a seed funding round led by ScOp Venture Capital, with other investors including Impact Assets, Amino Capital, and a number of semiconductor executives and angel investors.
With this investment, Alpha Design AI's flagship product, ChipAgents, is the world's first AI agent for chip design and verification.
With state-of-the-art generative AI technology, ChipAgents can seamlessly analyze and generate RTL design specs and code, write Verilogs, and automate the setup of test beds.
In addition, ChipAgents can learn from simulation in real time, autonomously verify and debug code, ensuring comprehensive and quality testing, making it a one-stop solution.
ChipAgents accelerates time-to-market and lowers development costs by dramatically reducing design and verification cycles, with the ultimate goal of improving RTL design and verification efficiency by 10x and driving cross-industry innovation with smarter, more efficient chip designs.
As Founder and CEO William Wang said, "ChipAgents marks a critical shift in the way the semiconductor industry designs and verifies. We believe AI agents are the key to solving complex EDA challenges and bringing innovations to market faster. ”
Technical team
The technical background of the Alpha Design AI team is equally impressive.
Not to mention the founder William Wang, a pioneer in the field of large language models and generative artificial intelligence, and currently the Melikamp Professor of Artificial Intelligence at the University of California, Santa Barbara, and was also named one of the IEEE Top 10 People to Watch in Artificial Intelligence.
Board member John Bowers is also a big bull, Fred Kavli Professor of Nanotechnology, Director of the Institute for Energy Efficiency, and a member of the United States National Academy of Engineering. He has extensive entrepreneurial experience, having served as a co-founder of several companies such as Quintessent and Aurrion.
Another board member, Ivan Bercovich, served as VP of Engineering at Graphiq, which was previously acquired by Amazon, where he focused on natural language understanding and generative technologies, specializing in structured data, dialogue systems, and human-computer interaction models, and is now a board member of several technology companies.
The rest of the team is also a veteran of industry giants such as NVIDIA, Microsoft, Meta, Yahoo, Snowflake, Salesforce, and NEC.
Alpha Design AI 核心技术解读：Gödel Agent
Recently, arxiv included a paper on AI agents that proposes a framework called Gödel Agent that can achieve autonomous recursive self-improvement through large language models, significantly improving task performance without relying on human-designed components.
William Wang's name appears prominently in the author column of the paper, and one of the major selling points of ChipAgents is that it can independently verify and debug the design code, and Leifeng.com speculates that Gödel Agent may be related to the basic framework of ChipAgents.
And in contrast to existing agents, only the Göodel Agent can recursively improve itself without any restrictions. Hand-designed agents rely on limited and labor-intensive human expertise, while meta-learning optimization agents are constrained by fixed meta-learning algorithms.
Judging from the content of the paper, Gödel Agent uses the "monkey patch" method.
Monkey patching technology allows agents to directly read and modify their own code in runtime memory during execution, enabling its functionality to be updated in real-time without being limited by fixed algorithms. This flexibility is essential for the agent's ability to improve itself.
Core features
The paper also mentions several other core features of Gödel Agent, including:
For example, self-awareness by checking runtime memory, dynamic code modification to improve its own logic, and its interaction with the environment. These features allow Gödel Agents to learn, improve, and even get smarter with each recursive iteration, just like a human.
- Self-awareness by inspecting runtime memory: The Gödel Agent is self-aware by inspecting runtime memory, specifically local and global variables in Python. This capability enables the agent to extract and interpret the variables, functions, and classes that make up the environment and the agent itself, in line with the modular structure of the system. By introspecting these elements, the agent gains an understanding of its own state of operation so that it can adjust accordingly.
- Self-improvement through dynamic code modifications: Gödel Agent is able to reason and plan to determine if its own logic needs to be modified. If a change is deemed necessary, the Gödel Agent generates new code, writes it to runtime memory dynamically, and integrates it into its operating logic. This dynamic modification allows it to evolve by adding, replacing, or removing logical components in the face of new challenges, enabling it to improve itself.
- Environmental interaction: In order to evaluate performance and gather feedback, Gödel Agents are equipped with interfaces to interact with their environment. Each task provides a tailored environment interface that enables it to assess performance and adjust policies accordingly. This interaction is an important part of the feedback loop in the recursive improvement process.
- Recursive improvement mechanism: At each time step, the Gödel Agent determines the sequence of actions to be performed, including reasoning, decision-making, and action execution. When the operation is complete, the Gödel Agent evaluates whether its logic has improved and decides whether to proceed to the next recursive iteration. As it iterates, Gödel Agent's logic evolves, with each step potentially improving its decision-making capabilities.
- Target prompts and task handling: Target prompts inform the Gödel Agent that it has the necessary permissions to elevate the logic and describes the tools available for improvement. The tip encourages Gödel Agent to fully explore its potential and use the tools to optimize itself. To ensure effectiveness in different tasks, an initial strategy is provided for the Gödel Agent, which will begin to explore different strategies to analyze their efficiency in optimizing performance.
Combining the above techniques, Gödel Agent theoretically allows for unlimited self-improvement, but current large language models (LLMs) have some limitations. To address these challenges, the paper integrates several support mechanisms to improve performance:
- Think before you act: The Gödel Agent infers before taking action, which results in an inference path and analysis instead of taking action immediately. This approach improves the quality of decision-making, prioritizing planning over reckless action.
- Error Handling Mechanism: An error during execution can cause the agent process to terminate unexpectedly. To alleviate this problem, if an operation is incorrect, the Gödel Agent stops the current sequence, moves on to the next time step, and saves the error information to improve future decisions.
- Additional tools: The Gödel Agent is also equipped with other potentially useful tools, such as the ability to execute Python or Bash code and call LLM APIs. The addition of these additional tools accelerates the convergence of the Gödel Agent recursive optimization process.