Logan Kilpatrick, who leads AI Studio and Gemini API at Google DeepMind, revealed something wild when we interviewed him during a recent episode of The Neuron podcast, which I co-host with Corey Noles.<\/p>\n
Kilpatrick says you can take screenshots of websites, drop them into AI Studio, type \u201cclone this,\u201d and get a working version in seconds<\/em>. In fact, that\u2019s how Google\u2019s AI team prototypes its own products. He literally takes a screenshot of AI Studio, spins up a clone, then iterates on it.\u00a0<\/p>\n This is something you could do, too,\u00a0right now!\u00a0That\u2019s because\u00a0AI Studio\u00a0is completely free and already available to you to experiment with.<\/p>\n In this Neuron podcast episode<\/a>, Kilpatrick pulled back the curtain on Google\u2019s entire AI ecosystem and showed us exactly how to use it.<\/p>\n (<\/strong>3:31<\/strong><\/a>)<\/strong>\u00a0How AI Studio has evolved:<\/strong>\u00a0It started as a playground tool, but\u2026 \u201cwe\u2019ve moved on beyond that\u2026 we released this build tab where you can go in and actually have the model build you an entire app.\u201d<\/p>\n (<\/strong>4:42<\/strong><\/a>)<\/strong> His personal AI Studio workflow:<\/strong>\u00a0\u201cPrototyping of AI experiences is is how I spend most of my time. It\u2019s like, ah, wouldn\u2019t it be cool if you could do this thing with AI in a product?\u2026I can just put in a really simple prompt and it just like is connected to my Gemini API keys and it\u2019s connected to Gemini and all of our other models by default out of the box\u2026It\u2019s just like baked in.\u201d<\/p>\n (<\/strong>46:03<\/strong><\/a>)<\/strong>\u00a0His go-to first project:<\/strong>\u00a0\u201cClone your favorite website. I take a screenshot of AI Studio, put it back in AI Studio, and I say clone this.\u201d<\/p>\n (<\/strong>16:59<\/strong><\/a>)<\/strong>\u00a0AI Studio\u2019s sweet spot for your workflow:<\/strong>\u00a0\u201cWe\u2019re not trying to solve every problem\u2026 with all the vibe coding stuff we\u2019re doing and with build mode, we\u2019ll get you a working prototype and then get you out into a full-fledged professional developer product.\u201d<\/p>\n (<\/strong>21:43<\/strong><\/a>)<\/strong>\u00a0Text-to-speech applications:<\/strong>\u00a0\u201cThere\u2019s been a huge amount of demand\u2026 for very native human sounding text-to-speech video generation.\u201d<\/p>\n (<\/strong>22:18<\/a><\/strong>)<\/strong>\u00a0Make your own AI podcast in your voice:<\/strong>\u00a0\u201cYou could make your own AI podcast in your voice\u2026 there\u2019s actually not a big gap between being able to do that and what Notebook LM is doing.\u201d<\/p>\n (<\/strong>26:43<\/strong><\/a>) AI tutor for learning<\/strong>: \u201cHaving an AI tutor that could help me\u2026 would have been so valuable and not spend my time getting yelled at on Stack Overflow.\u201d<\/p>\n (<\/strong>30:50<\/strong><\/a>) A surprising lawyer adoption stat:<\/strong>\u00a0\u201c99% of lawyers had tried Gemini, Claude, or ChatGPT.\u201d Why Gemini? \u201c\u2026they just want long context to bring more documents.<\/a>\u201d<\/p>\n (<\/strong>36:04<\/a><\/strong>)<\/strong>\u00a0Build a food tracking that actually works:<\/strong>\u00a0\u201cSnap a picture and have the model deeply understand\u2026 using bounding boxes and size proportional understanding.\u201d<\/p>\n (<\/strong>42:04<\/a><\/strong>)<\/strong>\u00a0Medical applications with<\/strong>\u00a0MedGemma<\/a><\/strong>:<\/strong>\u00a0\u201cI talk to customers all the time who are blown away with MedGemma\u2026 one of the highest traction query use cases in ChatGPT was this medical use case.\u201d<\/p>\n AI Studio \/ Gemini\u2019s most underrated features?<\/strong><\/p>\n Deep Research (<\/strong>47:04<\/strong><\/a>):<\/strong>\u00a0\u201cThe model will go visit between tens and hundreds or even thousands of different web pages\u2026 it\u2019s a mind-blowing experience.\u201d (P.S.: In case you forgot, Google invented Deep Research.)<\/p>\n The Live API (<\/strong>20:54<\/strong><\/a>):<\/strong>\u00a0\u201cAll the context you need to solve the problem is visible on my screen\u2026 the AI system doesn\u2019t need access to other tools.\u201d\u00a0(It looks like OpenAI is competing on this vector now too<\/a>.)<\/p>\n (<\/strong>5:45<\/strong><\/a>)<\/strong>:\u00a0Google Cloud integration:<\/strong>\u00a0\u201cAI Studio is built on all the Google cloud infrastructure\u2026 deploy uses Cloud Run behind the scenes.\u201d<\/p>\n (<\/strong>17:31<\/strong><\/a>):<\/strong>\u00a0Large codebase challenge:<\/strong>\u00a0\u201cThere\u2019s lots of context buried in different places\u2026 it\u2019s really difficult to orchestrate this together.\u201d<\/p>\n (<\/strong>18:08<\/strong><\/a>) On context engineering:<\/strong>\u00a0\u201cNot based on the user prompt\u2026 it\u2019s trying to do optimization on getting the relevant data to actually help answer.\u201d<\/p>\n (<\/strong>39:28<\/strong><\/a>)<\/strong>\u00a0Firebase Studio<\/strong><\/a>\u00a0as another solution:<\/strong>\u00a0\u201cFirebase Studio\u2026 you get the batteries included with compute and storage\u2026 should be able to get it to do image understanding with Gemini through Firebase, get a database, all that stuff.\u201d<\/p>\n (<\/strong>40:20<\/strong><\/a>)<\/strong>:\u00a0Google\u2019s open source strategy and why\u00a0<\/strong>Gemma<\/strong><\/a>\u00a0(their small open models) exists:<\/strong>\u00a0\u201cIt\u2019s us meeting developers and customers where they are.\u201d<\/p>\n (<\/strong>8:26<\/strong><\/a>)<\/strong>\u00a0On Genie 3 and world models:<\/strong>\u00a0\u201cFully functional world models that have temporal consistency where you can make edits to the world\u2026 it\u2019s not like there\u2019s a file somewhere being referenced\u2026 it is actually on the fly generating the actual ecosystem.\u201d<\/p>\n (<\/strong>13:03<\/a><\/strong>)<\/strong>\u00a0Logan\u2019s AGI reality check:<\/strong>\u00a0\u201cThe models can\u2019t follow basic instructions of chess\u2026 they want to make all these illegal moves\u2026I could probably teach a 12-year-old how to play chess in 30 minutes and they\u2019d be able to pick it up.\u201d<\/p>\nActionable gold from this episode about secret AI tools<\/h2>\n
Getting started with AI Studio<\/h3>\n
Real-world applications of AI Studio you can try today<\/h3>\n
Developer and technical insights<\/h2>\n
Where all this goes from here<\/h2>\n