エピソード

  • Airbnb — Turning Scientific Method Into AI Infrastructure

    Airbnb — Turning Scientific Method Into AI Infrastructure
    2026/09/17
    Airbnb needed to understand the rare, risky edge cases a new AI customer-service assistant might run into — the kind of thing that used to take a data scientist months of manually reading through support transcripts, one judgment call at a time. So Airbnb built Insight Miner: an agent harness that doesn't just run an LLM over the data, but wraps it in an explicit scientific-methodology layer — extract, embed, cluster, then classify — so the process of investigation, not just the answer, is reproducible and auditable by someone other than the original analyst. We walk through how the harness is built, why "methodology as infrastructure" is the real idea here rather than any single new technique, and how investigations that used to take months now take days — with usage spreading from a handful of data scientists to dozens of non-technical teams across the company. Source article: "Beyond the Model: Engineering AI Infra with Scientific Judgement" — Airbnb Tech Blog, https://airbnb.tech/ai-ml/beyond-the-model-engineering-ai-infra-with-scientific-judgement/ (published 2026-09).
    続きを読む 一部表示
    21 分
  • Spotify — Why Bayesian A/B Testing Doesn't Pay Off

    Spotify — Why Bayesian A/B Testing Doesn't Pay Off
    2026/09/17
    Bayesian A/B testing gets pitched as more intuitive, safer to peek at early, and automatically smarter about multiple metrics — so why didn't Spotify adopt it? This episode walks through Spotify's own math: under the flat priors most platforms actually ship, Bayesian and frequentist testing produce numerically identical results, and the "advantages" people cite either require infrastructure most teams don't have (calibrated priors, Bayes-factor stopping, hundreds of historical experiments) or quietly reduce to standard frequentist practice anyway. We cover the business context (why experimentation trustworthiness matters more than framework fashion), the statistical reasoning (posterior equivalence, the winner's curse, decision theory), and what it means for teams facing the same "should we go Bayesian?" question. Source article: "Why Spotify Is Not Using Bayesian A/B Testing" — Spotify Engineering, https://engineering.atspotify.com/2026/9/why-spotify-is-not-using-bayesian-a-b-testing (published 2026-09).
    続きを読む 一部表示
    24 分
  • Instacart — Can an AI Agent Out-Model Your Best Data Scientist?

    Instacart — Can an AI Agent Out-Model Your Best Data Scientist?
    2026/09/15
    Instacart's machine learning engineers ran an experiment most teams are quietly wondering about: what happens when you let a coding agent loose not just on your pipeline code, but on the actual modeling work — picking features, trying architectures, tuning hyperparameters — for models that have already been optimized by humans for months? The results were a genuine surprise in places: on some of Instacart's most mature, hardest-to-improve production models, agent-driven runs delivered 3-5% offline error reductions among their most promising attempts, holding up against a battery of randomization checks before anyone trusted them. But the team is just as candid about the other side — agents with broad permissions in a modeling environment are a real safety surface, not a hypothetical one, and they're recommending sandboxing and compliance awareness even after their own trial runs went mostly smoothly. Source article: "Agentic Machine Learning Modeling at Instacart" — Instacart Tech Blog, https://tech.instacart.com/agentic-machine-learning-modeling-at-instacart-fb3ecd295ee7 (published 2026-09).
    続きを読む 一部表示
    19 分
  • Netflix — Grading the Graders: Can an AI Judge Your Synopsis?

    Netflix — Grading the Graders: Can an AI Judge Your Synopsis?
    2026/09/14
    Every show on Netflix lives or dies a little by its synopsis — the couple of sentences that decide whether you press play. With a catalog too vast for humans to proofread line by line, Netflix built a team of AI "judges" to score synopses on tone, clarity, precision, and factual accuracy, each specializing in its own dimension rather than trying to do it all at once. To make sure those judges could actually be trusted, the team calibrated them against 600 hand-labeled examples through eight rounds of work with real creative writers — and didn't stop until the AI's agreement with expert writers matched how much the writers agreed with each other. The payoff: judged quality scores turn out to predict real viewer behavior, like whether people actually click play and how often they abandon a title early, letting Netflix catch a weak synopsis months before a show even launches. Source article: "Evaluating Netflix Show Synopses with LLM-as-a-Judge" — Netflix Technology Blog, https://netflixtechblog.com/evaluating-netflix-show-synopses-with-llm-as-a-judge-6269251e6f28 (published 2026-04).
    続きを読む 一部表示
    20 分
  • Google — Beating Your Own Teacher at Tool Use

    Google — Beating Your Own Teacher at Tool Use
    2026/09/14
    A deep dive into how Google Research flipped the standard recipe for training AI models to use tools and APIs. Instead of writing a user question first and then struggling to find a matching chain of tool calls that answers it — which fails constantly once you're picking from thousands of real APIs — their ToolGrad system builds a verified, working tool-call workflow first, using an iterative loop that treats an AI critic's plain-language feedback as a steering signal (borrowed from a technique called "textual gradients"), and only then writes the question that workflow answers. We cover the business problem (query-first data generation wastes enormous effort on unanswerable or unverifiable examples), the technical approach (a four-step propose-execute-select-update loop building tool-call chains from a catalog of 16,000+ real APIs), and the standout result: a fine-tuned 12-billion-parameter model trained on this data scored competitively with Gemini 2.5 Pro and Claude 4.5 Opus on a leading function-calling benchmark — and notably outperformed the very model that generated its own training data. Source article: "ToolGrad: Efficient Tool-Use Dataset Generation with Textual 'Gradients'" — Google Research, https://research.google/blog/toolgrad-efficient-tool-use-dataset-generation-with-textual-gradients/ (published 2026-09).
    続きを読む 一部表示
    17 分
  • Uber — Who's Really Behind Your AI Agent's Actions?

    Uber — Who's Really Behind Your AI Agent's Actions?
    2026/09/12
    A deep dive into how Uber built a cryptographic identity system for AI agents that solves a problem traditional identity models were never designed for: agents that "act for or in the place of another." We cover the business problem (an on-call Investigation Agent modifies a monitoring configuration, but when something goes wrong, nobody can trace the action back to the human engineer ultimately responsible, because the chain of who-asked-whom disappears across hops), the technical approach (a Security Token Service issuing short-lived JWTs that embed the *entire* actor chain — not just the last caller — using an OAuth 2.0 Token Exchange–inspired pattern layered on SPIFFE/SPIRE cryptographic workload identity, enforced at an MCP Gateway), and the realized payoff: sub-40ms p99 token-exchange latency at production scale across thousands of internal agents, with a roadmap toward dynamic access control and a unified enforcement plane. Source article: "Solving the Identity Crisis for AI Agents" — Uber Engineering Blog, https://www.uber.com/us/en/blog/solving-the-agent-identity-crisis/ (published 2026-05).
    続きを読む 一部表示
    22 分
  • Netflix — One Model, One Pass, Your Whole Homepage

    Netflix — One Model, One Pass, Your Whole Homepage
    2026/09/11
    A deep dive into how Netflix replaced its traditional multi-stage homepage recommender — separate systems for candidate retrieval, ranking, and row layout — with a single transformer that generates the entire structured, multi-row homepage in one pass. We cover the business problem (stitching together multiple separately-trained models to build one coherent homepage adds engineering complexity and serving latency, while still needing to handle cold start, freshness, and business-rule placement constraints in production), the technical approach (treating the user's history and request context as a prompt and autoregressively generating the whole homepage as the response, trained with an LLM-style recipe of pretraining on historical homepages followed by post-training via either weighted binary classification or reinforcement learning), and what the reported results actually show — a substantial lift on the core engagement metric alongside a 20% cut in end-to-end serving latency, plus the surprising offline finding that enriching the prompt beat simply scaling up the model, and that reinforcement learning increased homepage diversity as a side effect even though diversity was never an explicit training objective. Source article: "GenPage: Towards End-to-End Generative Homepage Construction at Netflix" — Netflix / arXiv, accepted RecSys 2026, https://arxiv.org/abs/2606.31031 (published 2026-06).
    続きを読む 一部表示
    21 分
  • NVIDIA — One Checkpoint, Three Jobs for Robots

    NVIDIA — One Checkpoint, Three Jobs for Robots
    2026/09/10
    A deep dive into how AWS and NVIDIA built a reference architecture for training physical AI — robots and autonomous vehicles — as a continuous, closed-loop "model factory" instead of a one-time training run split across fragmented GPU clusters. We cover the business problem (physical AI needs a perpetual loop of real-world data ingestion, synthetic data generation, post-training, and simulation evaluation, and running each stage on its own dedicated GPU cluster wastes capacity and adds expensive data movement), the technical approach (NVIDIA Cosmos 3's unified token stream carrying video, images, actions, audio, and text through one sequence; a Mixture-of-Transformers design pairing a reasoner and generator tower; and the elegant trick that the same checkpoint becomes a world model, an action-labeling model, or a deployed robot policy purely by changing which tokens start as noise, all running on one shared, persistent SageMaker HyperPod cluster), and what the reported scaling numbers actually show — near-linear multi-node training efficiency and roughly 1% checkpoint overhead, reported as reproducible methodology rather than leaderboard bragging rights. Source article: "Build a Physical AI Model Factory with NVIDIA Cosmos 3 on SageMaker HyperPod" — AWS Machine Learning Blog / NVIDIA, https://aws.amazon.com/blogs/machine-learning/build-a-physical-ai-model-factory-with-nvidia-cosmos-3-on-sagemaker-hyperpod/ (published 2026).
    続きを読む 一部表示
    21 分