Colab

Agentic Evaluations at Scale, For Everybody — Nicholas Kang & Michael Aaron, Google DeepMind

May 25, 2026 · 20:03

Nicholas Kang and Michael Aaron from Google DeepMind's Kaggle team argue that AI evaluations are broken due to being scattered, stale, and lacking transparency, citing a competing lab publishing inflated results by using custom compaction settings. They introduce four solutions: hackathons to channel community expertise, a standardized agent exam that returned 500+ submissions in its first week without promotion, a Game Arena where models play poker, chess, and werewolf for an ELO rating that cannot saturate, and an open benchmarks platform. A wastewater treatment plant engineer in Turkey built a novel safety benchmark from 20 years of field experience. They note that on SWE-Bench Pro, six frontier models land within a couple of percentage points, but the harness shifts performance by 22%, complicating comparisons. Challenges include high cost (400,000 poker hands for statistical significance), maintaining community engagement, and dealing with fast model deprecation cycles.

Let's go Bananas with GenMedia — Guillaume Vernade, Google DeepMind

May 18, 2026 · 1:17:14

Guillaume Vernade from Google DeepMind demonstrates the full GenMedia stack—Nano Banana for images, Veo for video, Lyria for music, and TTS—by illustrating an open-source book live. The core insight is that Gemini acts as the prompt engineer for every other model, and this works partly because the gen media models were trained on prompts written by Gemini. He shows how to generate character portraits, chapter scenes, animated video clips using images as first frames, and distinct music per chapter, with Lyria Realtime allowing continuous music that responds to new prompts like a DJ. A new interactions API caches context server-side to make chained multi-turn calls cheaper. For TTS, he uses a trick: assigning different speaking styles (e.g., long poetic pauses vs. breathless stutter) to the same two voices to create four distinct-sounding characters.

Ship Real Agents: Hands-On Evals for Agentic Applications — Laurie Voss, Arize

May 14, 2026 · 2:04:18

Laurie Voss, head of developer experience at Arize AI, delivers a hands-on workshop on evaluating agentic applications using Arize Phoenix, demonstrating that choosing the right eval matters more than tuning it: a correctness eval scored 0 out of 13 on the same financial analysis agent that a faithfulness eval scored 13 out of 13, because the model doesn't know the current year and cannot verify forward-looking data. He walks through building a complete eval pipeline from scratch—starting with tracing a Claude Haiku-based financial agent, reading and categorizing traces to identify root causes, then implementing code evals, built-in LLM-as-a-judge evals, and a custom actionability rubric with labeled examples. Voss emphasizes the importance of meta-evaluation to validate judge accuracy and introduces Phoenix experiments to prove prompt changes actually improve scores, not just vibes. Practical tips include using the impact hierarchy (data quality > prompting > model selection > hyperparameters) and the value of regression evals for safe model upgrades. The workshop closes with cost-aware evaluation, pairwise evaluation, and reliability scoring as next steps beyond the foundations…

[Full Workshop] Reinforcement Learning, Kernels, Reasoning, Quantization & Agents — Daniel Han

Jul 19, 2025 · 2:42:28

Daniel Han of Unsloth presents a technical workshop covering reinforcement learning (RL), kernels, reasoning, quantization, and agents, arguing that RL with verifiable rewards (RLVR) is the key to unlocking LLM capabilities beyond supervised fine-tuning. He explains why open-source models plateaued after September 2024 until DeepSeek-R1 showed that RL can elicit reasoning, and breaks down PPO, GRPO, and the REINFORCE algorithm, emphasizing that GRPO removes the value model for efficiency. Han details how reward functions—not algorithms—are the hardest part, with examples like distance-based scoring for math. He demonstrates a free Colab notebook training a base model to reason, and shows that dynamic quantization can shrink models like DeepSeek-R1 from 730 GB to 140 GB with only ~1% accuracy loss, arguing that GPUs may stop getting faster after FP4 precision.

AI Engineering with the Google Gemini 2.5 Model Family - Philipp Schmid, Google DeepMind

Jul 11, 2025 · 1:44:51

In this workshop, Philipp Schmid from Google DeepMind demonstrates AI Engineering with the Gemini 2.5 model family, focusing on using Gemini 2.5 Flash via a free API tier for hands-on coding tasks including text generation, multimodal processing of images, audio, and PDFs, function calling with structured outputs, and integration with MCP servers. The session covers setting up API keys in AI Studio, uploading files via the Files API (free for 1 day), and controlling thinking budgets (0–24,000 tokens) to manage cost and reasoning depth. Schmid shows how Gemini natively processes videos at 1 frame per second for accurate timestamp extraction and how PDFs are handled by combining OCR text with image understanding. He introduces native tools like Google Search with grounding metadata, code execution, and URL context, and explains how MCP servers can be used seamlessly with the Gemini SDK for tool calling. The workshop also covers parallel vs sequential function calling, the Agent Development Kit (ADK), and the upcoming asynchronous function calling for the Live API, providing a practical path from simple generation to agentic workflows.

Text-to-Speech Data Preparation and Fine-tuning Workshop - Ronan McGovern

Jun 3, 2025 · 34:00

Ronan McGovern walks through fine-tuning Sesame's CSM-1B text-to-speech model on a specific voice, using a YouTube video as the data source. He explains token-based TTS models, including how audio is represented via codebooks and how CSM-1B uses a main transformer for zeroth tokens and a secondary transformer for 31 hierarchical tokens. The workshop covers data preparation: downloading audio with yt-dlp, transcribing with Whisper Turbo, manually correcting the transcript, and splitting audio into 30-second chunks (41 clips from a 30-minute video). Fine-tuning uses Unsloth with LoRA adapters (rank 32, alpha 16) on linear layers, training for one epoch with a batch size of 2 and virtual size of 8, reducing loss from ~6.34 to ~3.72. Evaluation compares zero-shot inference (random speaker), voice cloning (closer but imperfect), and fine-tuned plus cloning (best result, producing an Irish-accented voice with natural errors). McGovern recommends 50+ 30-second clips for noticeable effect and notes that combining fine-tuning with voice cloning yields good performance even with limited data.

The Hidden Life of Embeddings: Linus Lee

Nov 7, 2023 · 18:15

Linus Lee, a Research Engineer at Notion, presents a tour of embedding visualization and manipulation at the AI Engineer Summit 2023. He demonstrates an encoder-decoder model fine-tuned from T5 that can reconstruct text from embeddings, allowing direct manipulation of features like length and sentiment by moving in latent space. Lee shows that mixing embeddings by splicing dimensions from two texts produces a semantic blend, and a linear adapter can decode text from OpenAI's text-embedding-ada-002 embedding space. Using CLIP, he interpolates between photographic and cartoon images and performs vector arithmetic to modify facial expressions. Lee releases these custom text embedding models on Hugging Face, enabling others to explore and interact with latent spaces. He argues that making model internals visible and manipulable fosters deeper understanding and more humane interfaces to generative AI.