---
type: "post"
id: "from-predictive-to-generative-how-michelangelo-accelerates-uber-264b3a"
slug: "from-predictive-to-generative-how-michelangelo-accelerates-uber-264b3a"
title: "From Predictive to Generative – How Michelangelo Accelerates Uber’s AI Journey"
description: "Uber describes how its Michelangelo platform evolved from a way to standardize model training and serving into a unified environment for collaborative, higher-impact machine learning."
retrieval_nugget: "Uber describes how its Michelangelo platform evolved from a way to standardize model training and serving into a unified environment for collaborative, higher-impact machine learning. The account matters to teams building with AI because it shows the platform work—quality standards, project tiers, code-based iteration and integrated tooling—needed to make models dependable at organizational scale."
published_at: "2026-11-17"
updated_at: "2026-11-17"
record_date: "2026-11-17"
date_kind: "published_at"
topics: ["ai-infrastructure","enterprise-ai"]
entities: ["Uber","Michelangelo","Palette","Canvas","MA Studio","Gallery","Manifold","PyML"]
related_patterns: ["ai-native-orgs-move-to-review-and-orchestration"]
novelty_label: "new_fact"
what_changed: "Uber describes Michelangelo’s evolution to support generative AI, alongside its scale of roughly 400 active ML projects, 5,000 production models and 10 million peak real-time predictions per second."
prior_refs: []
source_url: "https://www.uber.com/us/en/blog/from-predictive-to-generative-ai"
source_urls: ["https://www.uber.com/us/en/blog/from-predictive-to-generative-ai"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:from-predictive-to-generative-how-michelangelo-accelerates-uber-264b3a"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/from-predictive-to-generative-how-michelangelo-accelerates-uber-264b3a/","markdown":"https://newruntime.com/posts/from-predictive-to-generative-how-michelangelo-accelerates-uber-264b3a.md","json":"https://newruntime.com/posts/from-predictive-to-generative-how-michelangelo-accelerates-uber-264b3a.json"}
markdown_url: "https://newruntime.com/posts/from-predictive-to-generative-how-michelangelo-accelerates-uber-264b3a.md"
---

# From Predictive to Generative – How Michelangelo Accelerates Uber’s AI Journey

## Retrieval answer

Uber describes how its Michelangelo platform evolved from a way to standardize model training and serving into a unified environment for collaborative, higher-impact machine learning. The account matters to teams building with AI because it shows the platform work—quality standards, project tiers, code-based iteration and integrated tooling—needed to make models dependable at organizational scale.

## Summary
Uber describes how its Michelangelo platform evolved from a way to standardize model training and serving into a unified environment for collaborative, higher-impact machine learning. The account matters to teams building with AI because it shows the platform work—quality standards, project tiers, code-based iteration and integrated tooling—needed to make models dependable at organizational scale.

## What happened
Uber’s engineering announcement traces Michelangelo’s development from the company’s early machine-learning work. Before the platform, applied scientists built models in notebooks while engineers assembled bespoke deployment pipelines. Uber says teams lacked reliable, reproducible workflows for training and prediction at scale, ways to compare experiment results, and a standard path to production without creating custom serving containers. Michelangelo launched in early 2016 to standardize model development and deployment across the company.

The platform expanded beyond training and serving. Uber created Palette, a feature store for batch and near-real-time feature computation; the announcement says it hosts more than 20,000 features. Other components include Gallery, a metadata registry with a search API; Manifold, a model-agnostic visual debugging tool; and PyML, a framework for prototyping, validating and productionizing Python models. The company also describes support for distributed training with Horovod. The announcement places these components within an end-to-end machine-learning lifecycle rather than treating model training as an isolated task.

Uber divides its platform history into phases. From 2016 to 2019, its work centered on predictive machine learning for tabular data, including XGBoost use cases such as ETA prediction, risk assessment and pricing. Deep learning also appeared in areas such as 3D mapping and self-driving-car perception, requiring investment in GPU scheduling and distributed training. From 2019 to 2023, Uber focused on deep learning and collaborative development. The company says more than 60% of its tier-one models adopted deep learning in production during that period, improving model performance.

The announcement says Michelangelo 1.0 struggled to serve the newer work. Teams measured offline metrics such as AUC and RMSE but lacked a consistent view of online performance, training-data freshness and reproducibility. Project importance also varied, yet there was no tiering system to guide resource allocation or outage management. Deep-learning teams built their own toolkits, and collaboration often meant manually copying and merging notebook work without version control, branching or code review for UI configuration changes. Tooling was fragmented across teams and model types.

Michelangelo 2.0 was designed as a unified product with a shared UI and API for the end-to-end lifecycle. Uber identifies four themes: model quality and project tiering; model iteration as code through Canvas; deep learning as a first-class platform capability; and a unified developer experience through MA Studio. The announcement frames generative AI, beginning in 2023, as the next phase in the platform’s evolution, aimed at improving end-user experience and employee productivity.

## What the archive adds
New Runtime’s earlier coverage has described other platforms and developer surfaces, including [Claude Platform’s OpenTelemetry tracing](https://newruntime.com/posts/claude-platform-can-now-emit-opentelemetry-traces-from-5c6a70/) (2026-10-19). That note concerns operational visibility for managed agents and APIs; Uber’s announcement is a new event, focused on the broader lifecycle and collaboration model of an internal ML platform.

## Trend signal
This supports [AI-native organizations move toward review and orchestration](https://newruntime.com/patterns/ai-native-orgs-move-to-review-and-orchestration/): Uber’s account makes project tiering, shared code workflows and model-quality review part of the infrastructure for producing business-critical ML. The signal is about organizational design as much as technical capability. It does not establish that the same platform choices or governance levels suit teams with different risk profiles or scale.

## Why now
Uber places the current phase from 2023 onward, as generative AI brought new goals for both customer experience and employee productivity. The platform history suggests those new applications build on years of work to unify model development, production operations and team collaboration.

## Why it matters
A platform that makes it easy to train a model is not necessarily a platform that makes the model safe to depend on. Uber’s account highlights operational gaps that can remain hidden when teams focus on offline accuracy: stale training data, weak reproducibility, missing online-performance measures and unclear priorities when projects fail. Those gaps become more consequential as models influence pricing, matching, fraud detection and other business-critical decisions.

The company’s response offers a useful design direction: define quality beyond a single score, distinguish projects by impact, put model iteration into collaborative code workflows, and bring formerly separate tools into a coherent lifecycle. These are not plug-and-play prescriptions, and Uber’s scale is unusual. But teams can use the same questions at smaller scale: what evidence makes a model production-ready, who reviews changes, and what support does a high-impact model receive that an experiment does not?

## Limits and uncertainty
The announcement is Uber’s own retrospective and platform description, not an independent evaluation. Its performance claims are not accompanied here by a comparison methodology or measurements that isolate Michelangelo’s contribution. The account gives scale figures for projects, training jobs, production models and peak prediction volume, but does not establish how those figures are distributed across workloads or what operating costs they entail. It identifies the four themes of Michelangelo 2.0 and its generative-AI phase, but the excerpt does not provide enough detail to assess the specific generative-AI capabilities, their access model or their outcomes; the article itself is longer than the excerpt shown here.

## Reader takeaway
The reusable lesson is not to copy Uber’s platform architecture wholesale. It is to treat model quality, risk-based prioritization and collaborative change control as first-class parts of an ML platform. A team that can answer how it measures online performance, keeps training data fresh and reviews model changes is better prepared to scale AI than one that only automates training and deployment.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Ai Infrastructure](https://newruntime.com/topics/ai-infrastructure.md): Explore the ai-infrastructure topic hub.
- topic hub: [Enterprise Ai](https://newruntime.com/topics/enterprise-ai.md): Explore the enterprise-ai topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
