Ed Chi, Vice President of Research at Google DeepMind, opened this year’s COMPUTEX forum on generative AI and intelligent content applications with a speech tracing today’s trillion-dollar AI industry back to its roots in recommendation and search systems.
Reviewing key milestones in AI development, Chi highlighted the progression from deep learning–powered ranking systems to neural networks capable of learning through sequential transduction processes, and eventually to chain-of-thought prompting, which enabled the emergence of large language models — which argued that would be better understood as “large reasoning machines.”
“The point of AGI is not that it can speak our language, but that it can reason the way we do,” Chi said, describing this as the beginning of mimicking human intelligence.
He concluded by showcasing Google’s Project Astra, a prototype “universal agent” powered by Gemini 2.0. The system demonstrates agentic capabilities that can solve complex tasks by seeing through a device’s camera, controlling a phone, remembering user preferences, and interacting through natural voice commands. According to Chi, these advances will fundamentally change how people interact with computers.
Next, Thomas Anderson, Vice President of AI and Machine Learning at Synopsys, outlined the applications of AI for the semiconductor design process, and highlighted the growing role of AI agents in electronic design automation (EDA) tools used to develop integrated circuits.
“Whether it’s data centers, self-driving cars, or cellphones, you need EDA software,” Anderson said. As demand grows for more advanced chips, manufacturers face pressure to automate design processes and reduce development cycles, which currently take about a year for each new generation.
As a leading EDA software provider partnered with NVIDIA and Microsoft, Synopsys has adopted AI in stages: first as an optimization tool, then as a generative AI assistant. Today, Anderson said, the company is moving toward AI as a colleague, “where essentially we design virtual engineers to work alongside humans to do certain tasks.”
Its latest platform, AgentEngineer, is designed to provide a fully autonomous multi-agent workflow spanning front-end, physical, analog, and back-end chip design, as well as simulation analysis.
Demonstrating the company’s Spec2RTL Agent, Anderson showed how AI can generate RTL design abstractions from specifications, create tests, debug code, and perform lint checks and fixes. These capabilities, he said, will help shorten design cycles and accelerate chip development.
Paul Cunningham, Senior Vice President at Cadence Design Systems, expanded on the challenges posed by the semiconductor industry’s rapid growth. As chip complexity increases and development schedules tighten, he said, companies are being forced to “push complexity like never before.”
Agentic chip design presents a particular challenge because of the vast amount of information that must be processed. “The complexity of the task is very, very high,” Cunningham said. To manage this, developers must create structured workflows that break problems into smaller steps and provide guardrails for large language models.
Drawing on decades of EDA expertise, Cadence has developed AgentStack, a multi-agent platform that enables automated workflows within a single integrated environment where agents can share information and collaborate.
“The world is changing,” Cunningham said. “The semiconductor industry is exploding because of AI in data centers and the physical world. At the same time, we need to use AI to transform how chips are designed so that we can move at the pace required to deliver the next generation of AI systems.”
