Google's Logan Kilpatrick on Gemini 2.5 Pro, the four converging AI curves, and why builders have never had it better

Summary

Logan Kilpatrick, who leads developer relations for Gemini at Google, makes a straightforward bull case for building AI products right now: four curves are moving simultaneously in builders' favor. AI costs are down roughly 99% over the past two years. Model intelligence keeps rising. Consumer understanding of AI is growing. And willingness to pay for AI products is increasing. All four at once, Kilpatrick argues, is historically unprecedented for anyone building a product.

The Gemini 2.5 Pro moment

Kilpatrick describes Gemini 2.5 Pro as the first time Google has simultaneously held a top-tier intelligence ranking and kept pricing competitive. He attributes this to Google's vertical integration — controlling silicon through TPUs, training infrastructure, and model delivery — which gives the company degrees of freedom on cost that most AI companies lack. The inference team is currently running around the clock to keep up with demand, which Kilpatrick frames as a forcing function for additional optimization work.

On pre-training, he pushes back on the idea that scaling is exhausted. His argument is that pre-training improvements compound: a 3% gain at the pre-training level can yield multiples of that benefit once reasoning and post-training are applied on top. Gemini 2.5 Pro, he says, improved through a combination of RL, pre-training, and post-training innovation — not RL alone.

Where the application opportunity sits

Kilpatrick sees multimodal as the most underbuilt area relative to what the models can already do. He points out that tasks requiring dedicated computer vision pipelines two years ago can now be handled with a prompt and an image or video, often at near-parity or better accuracy. Real-time audio and video products are similarly underdeveloped — he thinks the foundation is in place but the startup wave hasn't arrived yet, estimating roughly a 12-month lag between capability and product.

Coding is the category with the most current momentum, driven by what Kilpatrick calls the genuine vibe coding phenomenon.

API flywheel and the consumer-enterprise question

On the question of whether AI labs should focus purely on consumer products or maintain API businesses, Kilpatrick argues the two are complementary. His read on ChatGPT's rise is that the API business drove a proliferation of AI products that educated the market, which in turn brought users back to the best consumer products. He thinks companies with the capital and capability to do both should.

Product UX as the real bottleneck

Kilpatrick's sharpest product opinion is that the current UX model — where the user carries the entire burden of assembling context and crafting prompts — is a fundamental flaw, not just a design choice. He says he won't build a product that requires users to change their behavior to get value. The tools should extract context from users, not the other way around. Deep Research inside the Gemini app is his example of what this looks like working correctly: a simple question yields a 50-page report, which can be converted to a NotebookLM audio overview in one click.

His near-term vision is AI that integrates into existing communication patterns — texts, emails, calls, push notifications — rather than requiring users to adopt new interfaces. The live screen-sharing mode in AI Studio, where the model can see what's on a user's screen in real time, is where he thinks this ambient context model is most promising today.

Google Labs and internal innovation

Notebook LM, Whisk, and AI Studio all came out of Google Labs. Kilpatrick flags that Josh Woodward, who leads Google Labs, now also runs the Gemini app team — a structural change he expects to accelerate the pipeline from experimental product to consumer surface.

Hiring and evals

On whether to allow AI-assisted interviews, Kilpatrick says Google is actively evaluating both formats. His broader point is that the ability to work effectively with AI tools is itself a meaningful signal — and one that few interviewers currently test for directly.