@a16z - Sinofsky on Jev
@a16z on X Steven Sinofsky argues that natural language is a poor interface for AI and that the real opportunity is treating model outputs as probabilities inside software logic, which revives decades of computer science work. The post is a clip from an a16z podcast episode where he, Aaron Levie, and Martin Casado also discuss AI regulation and agent security.
- Most people struggle to ask good questions, and chat-style AI makes users pay for long answers they often find frustrating.
- Jev treats the model as a probabilistic if statement: a prompt returns a percentage like "80% customer service," which then drives routing logic.
- Sinofsky says this is not new, since 1960s and 70s computer science was built on probabilistic ideas that are now worth revisiting.
Steven Sinofsky's main point is that natural language was never a good interface for getting work done, and his company Jev is built around that idea. He says most people can't ask a good question even in a meeting, let alone craft a prompt that yields a useful answer. He also complains that you pay for every token while watching seven paragraphs stream out, often to find the answer was wrong or off-target. Jev's fix is to treat the model as a decision engine rather than a chatbot. Instead of asking a question and reading prose, you get back a probability, and that number feeds directly into ordinary control flow.
His example makes this concrete. Rather than writing "is this a customer service question, then route it there, otherwise send it to general help," the system returns "this is 80% customer service" and the if-statement branches on that percentage. Sinofsky argues this is not new: there is decades of computer science research on probabilistic programming, much of it from the 1960s and 70s, that people can now dust off. The programming model he describes is almost a small custom language, where you write a prompt, get back a number, and compare it inside a conditional. The coolest work in computer science, in his view, may shift back toward this older world.
The rest of the a16z conversation, which features Aaron Levie, Sinofsky, and Martin Casado, covers a separate track on regulation and security. The panel argues that most current AI regulation is being drafted before anyone knows how these systems will actually fail, which is how aviation and computer-virus rules were built after failures happened. They also flag practical security issues: agents run at huge scale, do not tire, and can probe systems in ways human staff never could, so permissions, authentication, and the whole security stack may need rethinking. Their closing argument is that much of the innovation may end up in the software layer built around models, since frontier labs are mostly building the models themselves and software builders need tools rather than beings.