Google TimesFM-3 zero-shot forecasting holds up, but new features show little gain

·2 views

Google released TimesFM-3, a 330-million-parameter time-series foundation model pretrained on over a trillion time points, on 31 August 2026. An independent data scientist tested it against classical statistical models and a LightGBM on public Wikipedia pageview and hourly temperature datasets. Using rolling forecast origins and a held-out period the model could not have trained on, the reviewer found no evidence of benchmark contamination and confirmed genuine zero-shot forecasting ability. However, two headline features of the new version — covariate support and multivariate forecasting — produced no meaningful improvement on the tested data. The reviewer also flagged a licensing concern: while the code is Apache-2.0, the model weights are restricted to non-commercial, non-production use under a separate license.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How clear repository conventions help AI coding agents avoid costly mistakes

AI coding agents can produce plausible-looking code changes that still violate architecture boundaries, duplicate components, or skip critical checks when repository conventions are unclear. The proposed solution is not better prompts but a codebase whose structure, ownership rules, and checks are explicitly documented and testable. Each directory should carry a defined ownership rule — for example, routes handle transport, services handle product logic, and data layers handle persistence — so any violation is easy to spot during review. A root README should list real, working commands for installation, testing, linting, and migrations, along with prerequisites, rather than generic or aspirational documentation. When boundaries are legible and each layer has a single reason to change, agents and human contributors alike can locate the right starting point quickly and limit the scope of any modification.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Browser-based document pipeline runs PDF OCR in 30 lines with no server needed

A developer tutorial published on September 3, 2026 demonstrates how to build a fully client-side document processing pipeline using the @stabrise/scaledp library at version 0.1.1. The pipeline converts PDF pages to images and runs optical character recognition entirely in the browser using WebAssembly and ONNX Runtime, requiring no server, API key, or file upload. Only one core package is a hard dependency, with ML engine packages installed optionally based on what the pipeline actually uses, keeping bundle sizes lean. Configuration requires a single configure() call where developers explicitly set asset URLs and caching preferences, with IndexedDB caching available to speed up repeat visits by storing model weights locally. The pipeline accepts various input formats including File, Blob, and URL, and returns one structured row per PDF page containing recognized text and word-level bounding boxes.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Agentic AI Workloads Force a Rethink of Platform Architecture Layers

A technical analysis argues that agentic AI workloads fundamentally challenge the core assumption of conventional integration platforms — that systems execute only what they are explicitly programmed to do. Unlike deterministic workflows where every route and rule is predefined, AI agents decide at runtime which tools to call and what actions to take, making behavior impossible to fully anticipate from code alone. This shift affects every architectural layer, from compute and messaging to data, governance, and observability. The compute model is particularly disrupted because agents run as long, stateful reasoning loops rather than clean request-response or event-driven bursts, making traditional autoscaling metrics less effective. The article does not argue against agentic workloads but calls for deliberately revisiting each infrastructure layer with this non-determinism explicitly in mind.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why AI-Generated Tests Passing Green Does Not Mean Your Code Actually Works

A common but flawed software development practice involves letting an AI model both write a feature and generate its own tests, then treating a passing result as proof the code works. Since the model authors both the implementation and the assertions, any misunderstanding it has will be consistently reflected in both files, making tests pass without validating correctness. A reliable evaluation gate requires a frozen oracle — test files and fixtures that the code generator cannot modify — kept in a separate, read-only directory. High test coverage also does not signal quality when the tests themselves were generated against the same logic they are supposed to verify. True verification demands an independent script with a logged exit code, human-owned fixtures, and a strict separation between what the AI can write and what it is judged against.

0 comments Read more at DEV Community