Jensen Huang and Marc Benioff on AI Agents and the Future of Computing

NVIDIA CEO Jensen Huang and Salesforce CEO Marc Benioff discuss AI agents at Salesforce’s Dreamforce conference. I am enthusiastic about agents and see this as an important direction for AI. They predict a future with large numbers of agents working for people; this is a vision for future systems, not evidence that those deployments already exist at that scale.

Huang says the computing industry has long provided tools, and envisions an industry that also supplies skills through agents using those tools. Benioff responds with interest as Salesforce introduces Agentforce. Huang imagines agents that can interpret requests, use tools, find collaborators, and work together on a mission. Both executives have business interests in the technologies they are discussing.

My own vision is that agents will be able to create or select tools, test them, retain useful results, and collaborate. I prefer to define an agent as a language model with tools, memory, and access to other agents. That is my preferred definition rather than a universal technical requirement; existing agents vary in autonomy, memory, and implementation.

Huang argues that learning without requiring humans to label every example removes a major bottleneck, and discusses using language-based knowledge to learn from other modalities. I expand on data constraints and the possibility of using private or synthetic data. My assertion that all public data has already been used is too sweeping: datasets and training access differ, and the clip provides no inventory proving that claim.

I use AlphaGo to illustrate reinforcement learning and self-play, but the history needs a correction. [DeepMind explains](https://deepmind.google/research/alphago/) that the original AlphaGo initially learned from human expert games and then improved through self-play. The later [AlphaGo Zero](https://deepmind.google/blog/alphago-zero-starting-from-scratch/) skipped human game records and learned through self-play; it still relied on human-designed rules, algorithms, and evaluation.

I describe two ways to extend capability: making better use of available data, including synthetic data, and allocating more computation at inference time. I speculate about OpenAI’s Q* project and whether it became o1 or generates training data; those connections are not established by this discussion. More test-time compute may help on some tasks, but it brings cost and latency and does not guarantee a better answer.

Huang says AI progress has moved much faster than the historical pace associated with Moore’s law, attributing the acceleration to changes across hardware and software. His numerical comparison is a broad claim whose metric and timeframe matter. I describe Moore’s law as chip compute doubling every 18 months, but it is conventionally an observation about transistor counts, not a guarantee that general-purpose computing power doubles at that rate.

I contrast CPU and GPU computing and say transistor shrinking has effectively stopped. That is too absolute: CPUs also use parallelism, and chip design continues to change through multiple approaches, even as conventional scaling becomes harder. GPUs are especially suited to many parallel workloads, but a claim that chip compute doubles every six months needs a defined workload and measurement rather than a universal rule.

Huang also describes a feedback loop in which AI helps create software and computer systems. I expect coding assistants and tools such as Cursor and Replit to make programming more accessible, potentially increasing the number of people who build software. Those outcomes are predictions; their scale and effects on work are not settled by the interview.

I go further and speculate that future systems might generate interfaces or model weights instead of human-readable code. Model weights are difficult to interpret, but interpretability is an active research area rather than a solved impossibility. The idea of replacing most software with generated weights is a possible future, not an established technical trajectory.

Huang lists safety work including supervised fine-tuning, curated data, guardrails, and reflection on generated answers. He describes models checking whether responses are safe and appropriate. I connect this to inference-time computation and OpenAI’s o1, while emphasizing its potential for reasoning. Reflection and longer reasoning are techniques to evaluate; they do not by themselves ensure truth or safety.

Benioff says people should be able to understand and use the technology in practical work, rather than treat agent construction as a difficult science project. I interpret his remarks in the context of his public criticism of Microsoft Copilot, though that comparison is my commentary rather than something established by the excerpt itself. I disagree with dismissing Copilot and see potential in its integration into workplace tools.

Benioff compares deploying agents to onboarding employees. I find the analogy useful: a human colleague becomes more effective with training, context, and clear expectations, and an agent likewise needs relevant documentation, instructions, appropriate access, and a way to retain useful context. The analogy does not mean an agent learns, remembers, or assumes responsibility in the same way as a person.

Huang reflects on the long effort to reinvent computing and the possibility that AI can help tackle other hard problems. I imagine a shift toward systems that generate software or interfaces for the task at hand rather than relying only on prewritten programs. That is an exciting possibility, but its form and timeline remain uncertain.