Industries / Technology / Software and SaaS
AI and machine learning
Companies and labs that gather data, train computer programs that learn from it (models), test them and sell their use, along with the tools and computing power that work depends on.
1956
In the summer of 1956 a small group of researchers met at Dartmouth College in Hanover, New Hampshire, for a study proposed the year before by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon; their proposal gave the field its name, artificial intelligence.
48
Milestones achieved across 7 domains.
17
To watch for: announced, not yet confirmed outside.
15
Open deliverables. Nobody has shipped them.
In one minute
Where it started
In 1955 four researchers proposed a summer study at Dartmouth College of "how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves", and the 1956 meeting gave the field its name. For decades the work stayed mostly in university and company labs, with spells of commercial use such as the rule-based "expert systems" of the 1980s. Today's business took shape after 2012, when a neural network trained on two gaming graphics cards won the ImageNet image-recognition contest by a wide margin.
Where it stands
By Stanford's 2026 AI Index, private investment in US AI firms reached $285.9 billion in 2025, and industry, not universities, made over 90% of that year's notable frontier models. Epoch AI, a research group that tracks AI, counted over 30 models trained with at least 10^25 FLOP (a FLOP is one arithmetic step) from 12 developers by June 2025. Most people meet the industry through chat apps: in a February 2026 Pew Research Center survey, 44% of US adults said they use ChatGPT and 6% said they use Claude.
What is open
Much of the open work is checking what the industry says about itself: whether published test scores hold up, whether a model behaves the same at every company that rents it out, whether web crawlers keep their promises, and whether AI-made images carry the marks EU law now requires. Other jobs make everyday work measurable, such as how long a rented model takes to wake up, which compressed model files lose the least, and what the same text costs in each language. Most need accounts, scripts and patience, not a supercomputer.
Domains
AchievedWatch forOpen
Domain
Progress
Last achieved
Watch for
Open
Data
Data work gathers, cleans, labels and licenses the text, images and other material that models learn from.
7 · 3 · 2
Last achievedA required list of sources2025
Watch forPaid when content is used
2
openTraining
Training turns data and computing power into a working model: choosing the model's design, running it through the data until its errors shrink, and then teaching it to follow instructions.
7 · 2 · 2
Last achievedReasoning learned by trial2025
Watch forTraining for repeatable answers
2
openTesting models
Testing measures what a model can do, and how reliably, before and after it is sold.
6 · 2 · 2
Last achievedMeasured in hours of work2025
Watch forNear experts on office work
2
openRunning models for users
This is the work of answering people's requests with a trained model, through an app, through a paid connection that other programs call (an API), or on the user's own computer, quickly and at low cost.
7 · 2 · 3
Last achievedSame model, different answers2026
Watch forEnergy per prompt, self-measured
3
openSafety and misuse
This work keeps AI models doing what their makers intend: testing them for dangerous abilities before release, catching customers who misuse them, labeling what they make, and meeting the rules governments set for the companies that build them.
7 · 3 · 2
Last achievedA pause pledge loosened2026
Watch forAI-run spying, as claimed
2
openTools 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 · 3 · 2
Last achievedA shared plug for tools2024
Watch forAgents defined as code
2
openSupply of computing power
This is how AI firms get the chips and machines they train and run models on: the chips sold for AI, the software that ties firms to them, renting by the hour, large purchase deals and the rules on who may buy.
7 · 2 · 2
Last achievedPaper speed versus real speed2024
Watch forGigawatt-scale chip deals
2
open
Advances to build with
2 research notes

Cursor Agent SDK
Findings that cite notes
Cursor released a kit for building AI agents, and its example, a code reviewer that cites the note behind each finding, is known only from Cursor's account.

Microsoft ThinkingBox
All twenty times
By Microsoft's own count, 50 steps of training took an AI model from passing 38 of 507 test tasks on all 20 tries to passing 107.
Open in ai and machine learning
3 of 15 open deliverables
Open · Data
Canary count in open datasets
A public count of "canary strings" left in the main open training datasets. A canary is a unique code that test makers put in their files so model builders can find and remove test material before training.
Built with what exists
infini-gram's free count API, which indexes DCLM-baseline, Dolma v1.7, OLMo-mix-1124, RedPajama, the Pile and C4; Hugging Face's dataset streaming plus ripgrep for FineWeb, which infini-gram does not index; canary strings copied from the repositories of BIG-bench and other test sets.
Open · Training
Spike warnings on real runs
A labeled public set of loss spikes taken from large training runs whose logs are published, and a warning tool scored on it: how many steps before each spike it raises an alarm, and how often it raises a false one.
Built with what exists
the Allen Institute for AI's published Weights & Biases training logs for its open OLMo 2 models (1, 7 and 13 billion parameters); the OPT-175B logbook; Python with pandas; simple statistics or a small sequence model over the loss and the other measures each log records.
Open · Testing models
Explain the biggest claim gaps
For the 25 largest gaps between a lab's published benchmark score and an independent score for the same model and test, rerun the test under the lab's stated settings and under a neutral default, and label each gap by cause: settings, test harness (the code that runs the test), grading, leaked test data, or unexplained.
Built with what exists
whatmodel's MIT-licensed data; Inspect, the open evaluation framework from the UK AI Security Institute, or EleutherAI's lm-evaluation-harness; the models' public APIs; the settings each lab published with its score.
Where it sits
Broader