Avoid Building Around Model Limits That Are About to Disappear

▶ Watch (07:40)

Paige Bailey’s rule: when everyone sprints toward the same workaround, that’s a signal the model is about to make it unnecessary. Context windows sat at 8,000 tokens, so teams built vector databases. Models struggled with non-English text, so teams built fine-tunes. Both were absorbed into base Gemini. MedPalM and MedLM customers now use Gemini out of the box with a custom prompt, no fine-tune required.

The same pattern is repeating with agent frameworks and MCP servers. Build around the constraint that won’t be fixed, not the one that will.

Analyzing Video and Images in AI Studio

▶ Watch (13:09)

Gemini 3.1 Flash Light costs $0.25 per million tokens and handles video natively. In the AI Studio video analysis demo (14:30), Paige pasted a YouTube URL, set an end time of 300 seconds, and asked for a timestamped dinosaur table. The model processed 30,900 tokens of frame-by-frame footage and returned a species list with fun facts.

With code execution enabled, the same model drew Python bounding boxes around green Lego bricks in a sandboxed environment. Click “get code” and the prompt, model settings, and tool calls export directly to TypeScript or Python.

Building Full-Stack Apps with AI Studio Build

▶ Watch (21:07)

AI Studio Build generates full-stack web apps from a plain-text description. In the AI Studio bookshelf app demo (21:35), Paige described a bookshelf scanner with Google Auth, Firestore persistence, and Google Search grounding to fill in titles and authors. The model planned, coded, and configured Firestore security rules without manual intervention, then fixed a permissions error on the next prompt.

Gemini Live, shown in the same session, carries screen video and a live camera feed into a single speech-to-LLM-to-speech pipeline. Paige switched the response language to Castilian Spanish mid-conversation without modifying system instructions.

A Multimodal Pipeline for Illustrating a Book

▶ Watch (51:25)

Guillaume Vernade’s workshop chains Gemini, Imagen 3, Veo, and LyRia through a single Python notebook against Wind in the Willows. In the Wind in the Willows illustration pipeline (58:12), Gemini generates per-character and per-chapter image prompts; Imagen 3 renders them in a dark fantasy style, passing prior character images in context so the toad stays consistent across scenes.

The LyRia 3 music generation demo (01:06:08) scores each chapter from the same chat session. The picnic gets orchestral folk guitar, the road gets fiddle and bass, the forest gets suspenseful pizzicato strings. Guillaume ran it without pre-checking the results.

Multi-Speaker TTS and API Priority Tiers

▶ Watch (01:08:03)

Gemini’s TTS API supports two named voices per call, but Guillaume shows a way to produce more. He asked Gemini to reformat a Wind in the Willows dialogue as a play script, tagging each character with a style note like “fast-paced British accent.” The model varies delivery from those inline cues even when both characters share one base voice.

A service_tier parameter, added the week of the talk, sets request priority. flex queues cheaply; priority costs double but improves throughput. For a batch notebook run, remove it and save money.

Running Agentic Workloads Locally with Gemma 4

▶ Watch (01:26:40)

Gemma 4 shipped the previous Thursday under Apache 2. The E2B pages embeddings from flash storage so the 2B-core model runs on a Pixel phone; the 26B is a mixture-of-experts with only 4 billion parameters activated per forward pass; the 31B dense model fits on a MacBook M4 or single GPU.

In the local 10-agent SVG generation demo (01:34:56), Ian served the 26B model on port 1234 via LM Studio, then pointed OpenCode at that endpoint. One orchestrator spawned 10 sub-agents, each drawing a different SVG, all offline. The output was a compiled page of animated SVGs.

Q&A

Can LyRia-generated music be passed into Veo as an audio track? No — Veo’s generation model was not designed to ingest audio, so combining LyRia output with Veo video is not currently supported; Guillaume said future model generations will likely add cross-modal input. ▶ 1:18:41

Are there plans to release open-weight generative media models for image and video? Unlikely for image and video due to content safety controls that become unenforceable with open weights; Guillaume said open-weight music models are more plausible, especially for real-time on-device use. ▶ 1:22:47

Notable Quotes

had context windows of like 8,000 tokens Paige Bailey · ▶ 5:07

cost something like $20 to run so I Guillaume Vernade · ▶ 53:17

This may not be impressive if you’ve used any model in 2026 other than to know that it’s Ian Valentine · ▶ 1:46:24

Key Takeaways

  • Gemini 3.1 Flash Light costs $0.25 per million tokens and handles video, images, and code execution in one API call.
  • Chain Gemini, Imagen 3, Veo, and LyRia through a single Python notebook to turn a plain text book into illustrated, animated, and scored chapters.
  • Gemma 4’s 26B mixture-of-experts model runs on a MacBook or single GPU and supports local multi-agent workloads via OpenAI-compatible endpoints.