One of the most thought-provoking talks at the Berkeley Haas AI Conference came from Idan Gazit, Senior Director of Research at GitHub Next — GitHub's internal R&D innovation lab within Microsoft. His central argument challenged an assumption most of us have internalized without questioning: that chat interfaces are the natural way to interact with AI.
The Television Analogy
When television was invented, what was the most popular form of serialized content? Radio dramas. So what was the first television show? A camera pointed at people performing a radio drama — complete with props that made sound effects, like popping a balloon to simulate a gunshot.
They didn't yet know the language of the new medium. Did they know that television would eventually produce MTV Cribs, or cinematic prestige TV? Of course not. Radio was what people knew, so radio was what they recreated.
Idan's argument: we're at exactly this moment with AI. Chat is our radio drama. We default to chat because it's the obvious interface, the one that made ChatGPT a phenomenon. But that doesn't mean it's the best — or even a particularly good — way to interact with AI for most use cases.
The Ladder of AI Value
Idan proposed a hierarchy of AI value that reframes how we should think about product design:
Highest value: Invisible. You walk into the kitchen wanting a salad, and the cutting board and vegetables are already there. Like an invisible butler. You never think about it.
Next level: Predictive. The AI guesses what you'll do next based on prior interactions. This is achievable today.
Current level: Conversational. You tell the AI what you want. This works, but it's not the ceiling — it's closer to the floor.
True "AI native" product design means finding the right moments and matching them with the right modalities. Sometimes that's chat. Sometimes it's proactive suggestions. Sometimes it's the AI silently doing something in the background you didn't even know you needed.
AI Makes Variants Cheap — And That Changes Everything
Here's a concrete product insight I haven't seen discussed elsewhere. With AI, generating multiple variants of anything is nearly free. Need three approaches to a design? Three drafts of copy? Three code implementations? The marginal cost is close to zero.
This changes how we measure success. If you generate three variants and one gets accepted, your acceptance rate is 33%. That sounds bad by traditional metrics. But the fact that you could explore three roads and pick the best one — that's actually the goal. That's AI-native thinking.
Traditional product metrics were built for a world where creating each variant was expensive. We need new measurement frameworks for a world of abundance.
The Innovation Lab Playbook
GitHub Next operates in a way that most corporate innovation labs talk about but few actually execute. Idan described it as "the department of high-leverage bets" — or more colorfully, "the department of f*** around and find out."
Their process: Every Thursday is demo day. People show off what they've been building. The first signal they look for: does this excite the other people in the room? If three people see a demo and say "yeah, let's do that" — that's the green light to continue.
Then they extend the project for a few more weeks. Build the next stage. Try it on other people at GitHub. Expand to design partners outside GitHub. Iterate. Gather evidence. Only then do they go to leadership and say: "Here's the bet. Here's why. Here's a working prototype."
Nothing speaks like a prototype. "As much as my official title is research," Idan said, "it's not research — it's prototyping. Our job is to make."
Hybrids Win
Idan's most personal advice was about career identity. Throughout his career, people asked: are you the best developer? No. The best designer? No either. So what's your value?
"I can glue those functions together because I live with one foot in either side. The ability to say I am self-capable of going from zero to one — today with AI, it's like wearing rocket boots. Hybridity was undervalued in the past. Now it's actually a superpower."
This resonated deeply. In an era where AI can help anyone execute in any domain, the people who can see across boundaries — who understand product AND engineering AND design — become exponentially more valuable. The specialist advantage is shrinking. The hybrid advantage is growing.
The Takeaway
We're in the radio drama phase of AI product design. The companies and builders who break out of the chat box — who discover the native language of this medium — will define the next era of software. The question isn't "how do we make a better chatbot?" It's "what does AI interaction look like when we stop assuming it has to be a conversation?"