Industries / Technology / Software and SaaS / AI and machine learning
Tools for builders
This is the work of making the free libraries, model-sharing sites and connection standards that people use to build AI models, share them and plug them into other software.
7
Milestones achieved, from 2010 to 2024.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2010
A shared toolbox in Python
scikit-learn began in 2007 as a Google Summer of Code project by David Cournapeau. On 1 February 2010 four researchers at INRIA, the French national computing institute, made its first public release, a free Python library of standard machine-learning methods that has shipped new versions about every three months since.
Credit
David Cournapeau; Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort and Vincent Michel, INRIA
2015
Google gives away TensorFlow
On 9 November 2015 Google released TensorFlow, the system it used for speech recognition, Smart Reply and Google Photos search, as open-source software anyone could use. Google said it trained neural networks up to five times faster than its first-generation system.
Credit
Google; announced by Sundar Pichai
2017
A pioneer library stops
On 28 September 2017 MILA, the Montreal lab behind Theano, one of the first widely used libraries for training neural networks, said it would stop developing it; version 1.0, released on 15 November 2017, was the last main release. Developers of PyMC, a statistics library that depended on it, later forked the code as PyTensor.
LessonA tool kept by one research lab can lose its maintainers, so projects built on it need a plan for that day.
Credit
MILA; the PyMC developers
2019
One library, many models
Hugging Face described its Transformers library, which gives one programming interface to many models of the Transformer design (the design behind most current language models), backed by a collection of trained models that users share. By August 2025 a study by Benjamin Laufer, Hamidah Oderinwale and Jon Kleinberg counted 1.86 million models on Hugging Face, many of them adapted (fine-tuned) from one another.
Credit
Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond and colleagues, Hugging Face
2022
PyTorch leaves Meta
On 12 September 2022 Meta moved PyTorch, its library for building and training neural networks, to the Linux Foundation as the PyTorch Foundation, governed by a board from AMD, AWS, Google Cloud, Meta, Microsoft Azure and NVIDIA. The project said it had more than 2,400 contributors and nearly 154,000 projects built on it. In May 2025 the foundation began hosting other projects, starting with vLLM.
Credit
Meta; the Linux Foundation
2024
Model files that run code
PyTorch's usual file format uses Python's "pickle" method, which can run code when a file is opened. On 27 February 2024 the security firm JFrog reported about 100 models on Hugging Face with real harmful payloads; one opened a hidden connection that gave an outside computer control of the user's machine as soon as the model was loaded. Hugging Face's safetensors format holds only numbers; a May 2023 audit by Trail of Bits, paid for by Hugging Face, EleutherAI and Stability AI, found "No critical security flaw leading to arbitrary code execution".
LessonLoading a model should never run code, so shared models belong in a format that holds only data.
Credit
David Cohen, JFrog; Hugging Face; Trail of Bits
2024
A shared plug for tools
On 25 November 2024 Anthropic published the Model Context Protocol (MCP), an open standard for connecting AI assistants to data and tools, "replacing fragmented integrations with a single protocol". On 9 December 2025 it gave MCP to the Linux Foundation's new Agentic AI Foundation, alongside Block's goose agent and OpenAI's AGENTS.md convention; the foundation counted more than 10,000 published MCP servers, and its top members include AWS, Google, Microsoft and OpenAI. When the MCP Verified newsletter installed 750 servers from the official registry on 19 October 2026 without supplying any passwords or keys, 427 (57%) started and listed their tools; most of the rest needed keys or settings, crashed or were offline.
Credit
Anthropic; the Linux Foundation
Watch for
Agents defined as code
Cursor says its Agent SDK is the framework it uses to build its own agents, such as Bugbot and Security Reviewer, and that one of its teams built a reviewer in about a day that checks each proposed code change against the team's notes on past mistakes and cites the note behind each finding (see the Note "Findings that cite notes"). Cursor gives no count of findings or false alarms.
Achieved whena team outside Cursor publishes the findings and false alarms of an agent built on the SDK.
Claimed by
Cursor, October 2026
Idle agents restored fast
Google says Agent Substrate, open-source software for running many AI agents on Kubernetes (a common system for running programs across many servers), saves idle agents to storage and restores them "in under 500 milliseconds", packs 10 times as many agents per machine as standard containers, and handles more than 500 suspend-and-resume actions per second. Google gave it to the Cloud Native Computing Foundation's early-stage Sandbox (see the Note "Put away between turns").
Achieved whensomeone outside Google measures restore time and density on its own cluster.
Claimed by
Google, November 2026
Checking a model's parent
Cisco released the Model Provenance Kit, open-source software that judges whether two models share an origin, first from their settings files and then from five signals in their weights. On Cisco's own test of 111 model pairs it reports 96.4% accuracy, 98.1% precision and 94.6% recall. Its database of known models covers about 150 base models.
Achieved whenan outside group scores it on a test set the group built itself.
Claimed by
Cisco, April 2026
Open
01
Checking declared parent models
A measured answer to how often a model on Hugging Face names the wrong parent. Take a random sample of 2,000 models under 8 billion parameters that declare a base model in their metadata, compare each with its declared parent using an open provenance tool, check every mismatch by hand, and publish the rate. License checks and family trees on the site rest on these declarations.
The briefCredit · unclaimed02
Recording changes to tool lists
An open archive that once a day starts every MCP server in the official registry that needs no keys, records the tools it offers and their descriptions, and publishes every change as a readable before-and-after comparison. The AI model reads these descriptions, so a changed description changes what the model is told.
The briefCredit · unclaimed
Open 01 · the brief
Checking declared parent models
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A measured answer to how often a model on Hugging Face names the wrong parent. Take a random sample of 2,000 models under 8 billion parameters that declare a base model in their metadata, compare each with its declared parent using an open provenance tool, check every mismatch by hand, and publish the rate. License checks and family trees on the site rest on these declarations.
Why it is still open
The tools have been scored only on their makers' own test sets, and studies of the site take the declarations as given. Nikolic, Baluta and Saxena report 90% to 95% precision on benchmarks of over 600 models; Cisco scored its kit on 111 pairs it chose; LISCAN.AI checks licenses along declared family trees for 2,204,432 models and reports incompatibilities in 89.7% of derived models. No published rate of wrong declarations was found.
Built with what exists
Cisco's Model Provenance Kit in compare mode; the huggingface_hub Python library to sample models by their base_model field; an ordinary computer for the settings-file stage and a rented large-memory machine for the weight signals; a spreadsheet for hand checks.
First step
Pull 100 models that declare a Llama or Qwen parent, run compare mode, and check the ten lowest scores by hand.
Done looks like
2,000 pairs checked; a mismatch rate with a 95% confidence interval; every flagged pair checked by hand with the reason recorded; the owners of mismatched models told, and their replies counted.
Open 02 · the brief
Recording changes to tool lists
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open archive that once a day starts every MCP server in the official registry that needs no keys, records the tools it offers and their descriptions, and publishes every change as a readable before-and-after comparison. The AI model reads these descriptions, so a changed description changes what the model is told.
Why it is still open
Studies so far follow what the registry says about servers, or whether servers start, not what servers tell the model. Gautam Bharti tracked the official registry's entries for 19,099 servers over 88.6 days in 2026 and found that 8.6% of servers ever rewrote their description, but his study leaves out the tool lists that servers send once started. MCP Verified's October 2026 run checked the handshake, the tool list and one call, and keeps no history; Aguara, a security scanner, says it monitors five registries daily, but its listing mentions no public history.
Built with what exists
the official MCP registry's list of servers; Docker containers with no keys or passwords inside; the MCP Python or TypeScript SDK to call only the handshake and the tool list; Git to store each day's snapshot; GitHub Pages for the change feed.
First step
Snapshot the 209 npm servers that started in MCP Verified's October run, twice, one week apart, and compare the two.
Done looks like
90 days of daily snapshots for at least 400 servers; a public change feed; a monthly count of description changes; a hand review, published, of every change that adds instructions aimed at the model.