Academia
The findings the whole industry is standing on
Vendor pages tell you what shipped. These tell you why it works. Every entry below is a concrete, checkable result from a university lab — with a plain reading of what it was, and where it surfaces in the products this site tracks.
23 institutions · 69 findings
Stanford University
Stanford, California · 5 findings · 3 sources
Stanford's edge is measurement and framing: it names the category (foundation models), then publishes the numbers everyone else argues over.
- 2021
“Foundation models” — naming the shift
Bommasani, Liang et al., CRFM
A 200-page report arguing that one pretrained model, adapted downstream, was becoming the substrate of the whole field — and that homogenisation concentrates risk.
Why it matters here · This is the framing behind Geek Mode: nearly every product in the tracker is a thin adaptation layer over four or five base models, so one vendor's change propagates everywhere.
- 2009
ImageNet
Fei-Fei Li and collaborators
14 million hand-labelled images across 20,000 categories, plus the annual challenge that made progress comparable.
Why it matters here · The 2012 AlexNet result on ImageNet is the moment deep learning stopped being a niche. Every image model in the catalog descends from that benchmark culture.
14M labelled imagesAI - 2022
HELM — holistic evaluation
Liang et al., CRFM
Evaluating models across accuracy, calibration, robustness, fairness, bias, toxicity and efficiency simultaneously, on the same scenarios.
Why it matters here · The reason we mark model claims as disclosed / inferred / unknown rather than repeating a vendor's single benchmark number.
AI - 2024
The inference cost collapse
HAI AI Index
Measured the cost of GPT-3.5-level performance falling by more than two orders of magnitude in under two years.
Why it matters here · Explains the pricing pattern in our change feed: entry prices fall or stay flat while capability jumps, because the underlying token cost keeps collapsing.
~280× cheaper in 18 monthsAIChips - 2023
Alpaca — cheap instruction tuning
Taori, Gulrajani et al.
Instruction-tuned a 7B LLaMA on 52K self-generated examples for a few hundred dollars of compute.
Why it matters here · Kicked off the open fine-tune ecosystem — the reason small vendors in the catalog can ship a credible assistant without training a base model.
<$600 of computeAI
UC Berkeley
Berkeley, California · 9 findings · 5 sources
Berkeley builds the plumbing. Most of what runs underneath the products in this tracker — the serving layer, the distributed runtime, the instruction set, the open evaluation — started in a Berkeley lab and left as open source.
- 2023
vLLM and PagedAttention
Kwon, Li, Zhuang, Sheng, Stoica et al., Sky Lab
Applied virtual-memory paging to the KV cache, eliminating the fragmentation that wasted most GPU memory during LLM serving.
Why it matters here · The default open serving engine. When a product in the catalog quotes tokens-per-second or self-hosts a model, this is usually the layer doing it.
- 2023
Vicuna and Chatbot Arena
LMSYS / Sky Lab
An open fine-tune plus a blind, crowd-sourced pairwise ranking that turned model comparison into an Elo leaderboard.
Why it matters here · The most-cited independent counterweight to vendor benchmark claims; a model's Arena placement now moves procurement decisions.
AI - 2017
Ray
Moritz, Nishihara, Stoica et al., RISELab
A distributed runtime for actor-style Python workloads: training, tuning, serving and RL on one substrate.
Why it matters here · Used to train and post-train frontier models, including RLHF pipelines at several labs in this catalog.
AIChips - 2010
Apache Spark
Matei Zaharia et al., AMPLab
In-memory resilient distributed datasets, an order of magnitude faster than disk-based MapReduce.
Why it matters here · Still the data-preparation layer under a large share of enterprise AI pipelines.
AIChips - 2010
RISC-V
Asanović, Patterson, Waterman, Lee
A free, open, extensible instruction set architecture with no licensing regime.
Why it matters here · The escape hatch from x86/Arm licensing — increasingly the control core inside AI accelerators and robotics controllers.
Chips - 1980 · 1988
Berkeley RISC and RAID
David Patterson, Randy Katz, Garth Gibson
Reduced instruction sets, and redundant arrays of inexpensive disks — commodity parts arranged to beat expensive monoliths.
Why it matters here · The architectural argument the entire hyperscale build-out rests on: scale cheap parts rather than buy big ones.
Chips - 2015
TRPO / GAE, then visuomotor policies
Schulman, Levine, Abbeel, Jordan
Stable policy-gradient methods, and the first end-to-end training of raw pixels to robot torques.
Why it matters here · TRPO's successor PPO is the algorithm behind RLHF; the visuomotor line is the direct ancestor of today's robot foundation models.
AIRobotics - 1985 · Nobel 2025
Macroscopic quantum tunnelling in Josephson junctions
John Clarke, Michel Devoret, John Martinis
Demonstrated that an electrical circuit can behave as a single quantum object with discrete energy levels.
Why it matters here · The experimental basis for every superconducting qubit shipping today — IBM, Google Willow, Rigetti. Awarded the 2025 Nobel Prize in Physics.
Quantum - 2014
Caffe
Yangqing Jia, BAIR
The first widely-used fast CNN framework with a model zoo of pretrained weights.
Why it matters here · Established pretrained-weight sharing as normal practice — the cultural precondition for open model hubs.
AI
MIT
Cambridge, Massachusetts · 4 findings · 3 sources
MIT holds both ends: the founding theory of the field, and the journalism that tells you which of today's claims will still stand next year.
- 1969
Perceptrons
Marvin Minsky and Seymour Papert
Proved a single-layer perceptron cannot represent XOR, and questioned whether deeper nets could be trained.
Why it matters here · Triggered the first AI winter — and the exact limitation that backpropagation later dissolved. The clearest case study in how a proof about one architecture can freeze a field.
AI - 1948
A Mathematical Theory of Communication
Claude Shannon
Defined information in bits and the entropy of a source: H = −Σ p log p.
Why it matters here · Cross-entropy loss — the objective every language model in this catalog is trained against — is a direct application.
H = −Σ p log pAI - 1966
ELIZA
Joseph Weizenbaum
A pattern-matching script that convincingly imitated a Rogerian therapist in a few hundred lines.
Why it matters here · Named the ELIZA effect: people attribute understanding to fluent text. Still the single biggest source of overstated AI capability claims.
AI - 2020
Liquid neural networks
Hasani, Rus et al., CSAIL
Continuous-time networks whose neuron dynamics adapt after training, robust with very few units.
Why it matters here · An active alternative line for edge robotics and drones where a transformer will not fit in the power budget.
AIRobotics
Carnegie Mellon University
Pittsburgh, Pennsylvania · 5 findings · 4 sources
CMU is where robotics became an engineering discipline rather than a demo: navigation, manipulation and autonomy with error bars attached.
- 1986 – 1995
Navlab and ALVINN — the first neural self-driving
Chuck Thorpe, Dean Pomerleau, Todd Jochem
ALVINN trained a small neural network to steer from camera input in 1989; in 1995 'No Hands Across America' drove 2,797 of 2,849 miles autonomously steered.
Why it matters here · End-to-end learned steering — the approach Tesla FSD returned to three decades later — was demonstrated at CMU first.
98.2% autonomous, coast to coast, 1995RoboticsAI - 1968
Heuristic search and the Logic Theorist lineage
Newell, Simon and successors
The Logic Theorist (1956) proved theorems from Principia Mathematica; the search tradition it started produced the planners robots still run.
Why it matters here · Every motion planner and agent task-decomposition loop in the catalog is a descendant of this search tradition.
AI - 2007
Boss wins the DARPA Urban Challenge
Red Whittaker, Chris Urmson and team
An autonomous Chevy Tahoe completed a 60-mile urban course with traffic, merges and intersections.
Why it matters here · Its team leadership went on to found or lead Waymo, Aurora, Argo and Uber ATG — the entire AV industry traces to this team.
Robotics - 2017
Libratus beats poker pros
Tuomas Sandholm and Noam Brown
Defeated top heads-up no-limit hold'em professionals in an imperfect-information game.
Why it matters here · The search-at-inference-time idea proven here is the direct ancestor of today's reasoning models that spend more compute per answer.
AI - 1987
Capability Maturity Model
Watts Humphrey, SEI
A five-level staged model for assessing software process maturity.
Why it matters here · The template for every AI-maturity and AI-assurance framework enterprises are now using to gate adoption.
Policy
Georgia Institute of Technology
Atlanta, Georgia · 4 findings · 3 sources
Georgia Tech's contribution is behaviour and scale: how robots act in the world, and how machine learning gets taught and deployed at volume.
- 1998
Behavior-Based Robotics
Ronald Arkin
Formalised reactive and hybrid architectures — motor schemas composed into behaviour, rather than a single deliberative planner.
Why it matters here · The subsumption-to-hybrid architecture pattern still shapes how humanoid stacks split reflexes from planning.
Robotics - 2014
OMSCS — the $7k accredited CS master's
Zvi Galil and Georgia Tech
An online MS in Computer Science at roughly a tenth of on-campus cost, with the same degree and admission standards.
Why it matters here · The largest single supply expansion of formally-trained ML engineers in the US — a structural input to every hiring plan in the catalog.
10,000+ enrolledAIPolicy - 2018
Embodied Question Answering and Habitat
Dhruv Batra, Devi Parikh and collaborators
Agents that must navigate a simulated 3D home to answer a question, plus the high-throughput simulator to train them.
Why it matters here · Established simulation-first training for embodied agents — how humanoid and household robot policies in the tracker are actually trained.
AIRobotics - 2017
Robotarium — remote-access swarm testbed
Magnus Egerstedt and GRITS Lab
A free, remotely-accessible multi-robot lab where anyone can upload and run swarm control code on real hardware.
Why it matters here · Made multi-agent control results reproducible on physical robots — the discipline drone-swarm vendors are now held to.
Robotics
California Institute of Technology
Pasadena, California · 4 findings · 3 sources
Caltech supplies the theory that tells you what is physically possible — and the vocabulary the rest of the industry then markets with.
- 2018
NISQ — naming the era we are actually in
John Preskill
Defined the Noisy Intermediate-Scale Quantum regime: 50–few-hundred noisy qubits, useful for experiments, not yet for fault-tolerant computation.
Why it matters here · The honest yardstick for every quantum product in the tracker — logical qubits and error rates matter, raw qubit counts do not.
Quantum - 1981
Simulating Physics with Computers
Richard Feynman
Argued that simulating quantum systems requires a computer that is itself quantum.
Why it matters here · The founding argument of the entire quantum computing industry, and still its clearest use case.
Quantum - 2020
Fourier Neural Operators
Anima Anandkumar, Zongyi Li and collaborators
Learn mappings between function spaces, solving families of PDEs orders of magnitude faster than numerical solvers.
Why it matters here · The technical basis for AI weather and physics-simulation products now shipping from several vendors in the catalog.
~1000× faster than FEM solvers on benchmark PDEsAIBiotech - 1989
Analog VLSI and Neural Systems
Carver Mead
Coined 'neuromorphic engineering' — silicon that mimics neural structure with analog physics.
Why it matters here · The ancestral line for event cameras and neuromorphic accelerators appearing in edge robotics stacks.
Chips
Virginia Tech
Blacksburg & Alexandria, Virginia · 4 findings · 4 sources
Virginia Tech is where autonomy meets the regulator: FAA-designated test ranges, naturalistic driving data, and the certification evidence products need before they can fly or drive.
- 2006
The 100-Car Naturalistic Driving Study
VTTI
Instrumented 100 vehicles continuously for a year, capturing 82 crashes and 761 near-crashes with video and sensor context.
Why it matters here · Produced the driver-inattention baseline that autonomous-vehicle safety claims are still measured against.
43,000 hours of real drivingRoboticsPolicy - 2016 – 2024
First FAA drone-delivery and BVLOS approvals
MAAP with Google Wing, Flytrex and partners
Flew the first FAA-approved package delivery in the US (2016, Wise County), then the BVLOS waivers that let operators scale beyond visual range.
Why it matters here · Zipline, Wing and Flytrex operate commercially today because this test range generated the safety case for them.
RoboticsPolicy - 2007
Odin — third place, DARPA Urban Challenge
Virginia Tech / TORC Robotics
A student-led autonomous vehicle that finished the urban course, spinning out TORC Robotics.
Why it matters here · TORC is now Daimler Truck's autonomous-freight arm — the trucking side of the AV market in the catalog.
Robotics - 2022
Quantum networking and Virginia's quantum corridor
VT Center for Quantum Information Science and Engineering
Entanglement distribution and quantum-network testbeds run with regional and federal partners.
Why it matters here · Networking is the unsolved half of quantum: useful machines will be linked, not monolithic.
Quantum
Harvard University
Cambridge, Massachusetts · 3 findings · 2 sources
Harvard's Belfer Center converts research capability into geopolitics — the country-level view of who can actually build this technology.
- 2023
Ranking national technology capability
Belfer Center
A composite index across five critical technologies showing the US leading overall while China leads or closes fast in specific layers.
Why it matters here · The context behind the China badge in our catalog: origin is a supply-chain and export-control fact, not a label.
PolicyChips - 2011
Soft robotics
George Whitesides, Rob Wood and the Wyss Institute
Pneumatic elastomer actuators — robots that are compliant by material rather than by control loop.
Why it matters here · The grasping approach used where rigid humanoid hands fail: produce, textiles, human contact.
Robotics - 2013
RoboBee
Rob Wood and team
An insect-scale flapping-wing robot weighing 80 milligrams achieving controlled flight.
Why it matters here · The extreme end of the micro-drone curve, and the source of much of the actuator work below it.
Robotics
Princeton University
Princeton, New Jersey · 3 findings · 2 sources
Princeton is the field's fact-checker: rigorous work on what AI systems can and cannot do, and on the evaluations that mislead.
- 2023
SWE-bench
Carlos Jimenez, Ofir Press, Karthik Narasimhan et al.
2,294 real GitHub issues from twelve Python repos; a model must produce a patch that passes the project's own tests.
Why it matters here · The benchmark every coding agent in this catalog now quotes. When a vendor claims an agent 'resolves issues autonomously', this is the number to ask for.
2,294 real-world issuesAI - 2024
AI Snake Oil
Arvind Narayanan and Sayash Kapoor
Separated genuinely capable generative systems from predictive-AI products that cannot work, and documented widespread evaluation leakage.
Why it matters here · Why claims in this tracker carry a confidence level and a source rather than being restated as fact.
AIPolicy - 2024
AI agents that matter
Kapoor, Stroebl, Narayanan et al.
Showed agent leaderboards ignore cost, so trivially expensive baselines can top them; proposed joint accuracy-cost evaluation.
Why it matters here · The reason we pair capability with entry price everywhere in the catalog instead of ranking on capability alone.
AI
University of Washington
Seattle, Washington · 3 findings · 1 sources
UW, with AI2 next door, is the strongest institutional advocate for genuinely open models — weights, data and training recipe together.
- 2023
QLoRA
Tim Dettmers, Luke Zettlemoyer et al.
4-bit quantised backpropagation through frozen weights into low-rank adapters — fine-tune a 65B model on one 48GB GPU.
Why it matters here · The reason a small vendor can ship a domain-tuned model without a training cluster.
65B fine-tune on a single GPUAIChips - 2024
OLMo — fully open models
AI2 with UW
Released weights, training data (Dolma), code, logs and checkpoints together.
Why it matters here · The only reference point for what 'open' should mean when a vendor in the catalog claims it.
AI - 2018
ELMo — deep contextual embeddings
Matthew Peters et al., AI2/UW
Word representations conditioned on the whole sentence, from a bidirectional language model.
Why it matters here · The bridge between static word vectors and the pretrain-then-finetune era that BERT and GPT industrialised.
AI
University of Toronto & Vector Institute
Toronto, Canada · 3 findings · 1 sources
Toronto kept neural networks alive through the winter, then detonated the current era with one ImageNet result.
- 2012
AlexNet
Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton
A deep CNN trained on two consumer GPUs cut ImageNet top-5 error from 26% to 15.3%.
Why it matters here · The result that made GPUs the substrate of AI and started the compute build-out we now track as its own segment.
26% → 15.3% top-5 errorAIChips - 1986
Learning representations by back-propagating errors
Rumelhart, Hinton, Williams
Popularised gradient descent through multi-layer networks via the chain rule.
Why it matters here · Still the training algorithm for every model in this catalog, at every scale.
AI - 2012
Dropout
Srivastava, Hinton et al.
Randomly zeroing units during training as an implicit ensemble, sharply reducing overfitting.
Why it matters here · Made deep nets trainable on modest data — standard in almost every architecture that followed.
AI
University of Illinois Urbana-Champaign
Urbana, Illinois · 2 findings · 1 sources
Illinois builds public compute and the software that made the internet usable — the open counterweight to hyperscaler infrastructure.
- 1993
NCSA Mosaic
Marc Andreessen, Eric Bina
The first widely-used graphical web browser, distributed free.
Why it matters here · The distribution precedent every consumer AI product now follows: free client, network effects, monetise later.
Policy - 2024
DeltaAI
NCSA and NSF
A national GH200-class system dedicated to open AI research workloads.
Why it matters here · Sets the floor for what academic groups can train without renting frontier-lab compute.
ChipsAI
University of Michigan
Ann Arbor, Michigan · 2 findings · 1 sources
Michigan owns the proving ground: autonomy claims tested against a repeatable, instrumented city.
- 2015
Mcity — the first purpose-built AV proving ground
University of Michigan and industry partners
A closed urban environment with intersections, roundabouts and staged pedestrians for repeatable testing.
Why it matters here · Turned autonomy validation into a reproducible experiment instead of a public-road anecdote.
Robotics - 2023
Accelerated evaluation of AV safety
Henry Liu and team
A dense-learning method that concentrates rare safety-critical scenarios, cutting required test mileage by orders of magnitude.
Why it matters here · How AV vendors can substantiate safety claims without driving billions of real miles.
~10³–10⁵× fewer miles neededRoboticsPolicy
University of Texas at Austin
Austin, Texas · 2 findings · 1 sources
UT Austin pairs one of the strongest robot-learning groups with a grand-challenge program on AI ethics.
- 2003 – present
Learning agents and RoboCup
Peter Stone and the Learning Agents Research Group
Multi-agent reinforcement learning, layered learning and transfer, proved out in competitive robot soccer.
Why it matters here · The multi-agent coordination lineage behind today's drone swarms and multi-agent LLM systems.
AIRobotics - 2016
The One Hundred Year Study on AI (co-led)
Peter Stone chairing the first study panel
A standing, longitudinal assessment of AI's effect on society rather than a one-off forecast.
Why it matters here · The template for continuous rather than episodic technology assessment — the same posture this tracker takes.
Policy
ETH Zürich
Zürich, Switzerland · 2 findings · 2 sources
Europe's strongest legged-robotics lineage, plus a serious push on trustworthy and open models.
- 2019
Learning agile and dynamic locomotion
Hwangbo, Hutter et al.
Trained locomotion policies in simulation with a learned actuator model and transferred them to hardware without tuning.
Why it matters here · Sim-to-real with actuator modelling is now the standard recipe for quadrupeds and humanoids across the robotics segment.
Robotics - 2022
Perceptive locomotion in the wild
Miki, Hutter et al.
Fused proprioception with terrain perception so a robot degrades gracefully when vision fails.
Why it matters here · The robustness bar industrial inspection robots are sold against.
Robotics
Tsinghua University
Beijing, China · 2 findings · 2 sources
Most of China's open-weight model lineage runs through Tsinghua labs and their spinouts — the supply side of the China column in this catalog.
- 2021
GLM — a bilingual pretraining objective
Du, Tang et al., KEG
An autoregressive blank-infilling objective unifying understanding and generation, released with open weights.
Why it matters here · The base of the GLM/Zhipu product family tracked here, and a template for other Chinese open releases.
AI - 2019
Tianjic — hybrid neuromorphic chip
Luping Shi and team
A single chip running both spiking and artificial neural networks, demonstrated on an autonomous bicycle. Nature cover.
Why it matters here · The clearest signal that China's accelerator strategy includes non-GPU architectures, not just GPU substitution.
Chips
NIST
Gaithersburg, Maryland · 2 findings · 1 sources
Not a university, but the body that turns research consensus into deadlines enterprises must actually meet.
- 2024
Post-quantum cryptography standards
NIST, after an eight-year open competition
Published ML-KEM, ML-DSA and SLH-DSA as the first standardised quantum-resistant algorithms.
Why it matters here · Converts 'quantum someday' into a dated migration project — the concrete reason the post-quantum thesis appears on our future tab.
3 standards finalised, Aug 2024QuantumPolicy - 2023
AI Risk Management Framework
NIST
A voluntary govern-map-measure-manage framework for AI risk.
Why it matters here · The vocabulary enterprise buyers now use in procurement questionnaires for the products in this catalog.
AIPolicy
University of Chicago
Chicago, Illinois · 1 findings · 1 sources
The first mathematical model of a neuron was written here — the 1943 paper every neural network in this catalog inherits from.
- 1943
A Logical Calculus of the Ideas Immanent in Nervous Activity
Warren McCulloch and Walter Pitts
Modelled a neuron as a threshold logic unit and proved that networks of them can compute any propositional function.
Why it matters here · The origin point of the entire field: the artificial neuron, unchanged in principle, is still the unit inside every model tracked on this site.
Dartmouth College
Hanover, New Hampshire · 1 findings · 1 sources
The field got its name at a summer workshop here, along with its founding over-optimism.
- 1956
The Dartmouth Summer Research Project
John McCarthy, Marvin Minsky, Nathaniel Rochester, Claude Shannon
A two-month workshop proposing that 'every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it'. It coined the term artificial intelligence.
Why it matters here · It also set the template for the hype cycle: the proposal estimated significant progress in one summer. Read every vendor roadmap in this tracker against it.
AI
Cornell University
Ithaca & New York City · 1 findings · 2 sources
Cornell built the first machine that learned from examples rather than instructions — and then took the credibility hit when it was oversold.
- 1958
The Perceptron
Frank Rosenblatt, Cornell Aeronautical Laboratory
A trainable linear classifier implemented in hardware — the Mark I Perceptron — with a convergence proof for separable data.
Why it matters here · The first learning machine, and the first time press coverage ran far ahead of capability. Both patterns repeat continuously in this tracker.
w ← w + η(y − ŷ)xAI
University of Manchester
Manchester, United Kingdom · 2 findings · 1 sources
The first stored-program computer ran here, and the question of whether a machine can think was posed here in testable form.
- 1950
Computing Machinery and Intelligence
Alan Turing
Replaced 'can machines think?' with the imitation game — a behavioural, decidable test — and pre-answered the standard objections.
Why it matters here · Every claim about model 'understanding' in vendor marketing is still arguing with this paper, usually without knowing it.
- 2018
SpiNNaker
Steve Furber and team
A million ARM cores wired to model spiking neural networks in biological real time.
Why it matters here · The largest neuromorphic machine built — the reference point for event-driven, low-power AI hardware.
1,000,000 coresChipsAI
University of Oxford
Oxford, United Kingdom · 2 findings · 2 sources
Oxford pairs deep vision research with the field's most-cited work on long-run risk and governance.
- 2014
VGGNet — depth with 3×3 convolutions
Karen Simonyan and Andrew Zisserman
Showed that stacking small uniform filters to 16–19 layers beat wider, shallower designs.
Why it matters here · Made architectural depth the default lever, and VGG features are still a standard perceptual loss.
AI - 2014
Superintelligence
Nick Bostrom, Future of Humanity Institute
Formalised the alignment and control problem for systems that exceed human capability.
Why it matters here · Directly shaped the safety teams, evaluations and policy commitments that frontier labs in this catalog now publish.
PolicyAI
Université de Montréal & Mila
Montréal, Canada · 3 findings · 1 sources
Bengio's lab produced the attention mechanism and the generative-adversarial idea — two of the three pillars the current era stands on.
- 2014
Neural machine translation by jointly learning to align and translate
Bahdanau, Cho, Bengio
Introduced an attention mechanism letting a decoder look back at any encoder state instead of one fixed vector.
Why it matters here · Attention, three years before Transformers. Every model in the catalog runs a scaled variant of this idea.
AI - 2014
Generative Adversarial Networks
Ian Goodfellow et al.
Two networks in a minimax game: a generator producing samples, a discriminator judging them.
Why it matters here · Dominated image generation until diffusion took over, and adversarial training still underpins evaluation and red-teaming practice.
min_G max_D V(D,G)AI - 2003
A Neural Probabilistic Language Model
Bengio, Ducharme, Vincent, Jauvin
Learned distributed word representations and a language model jointly, beating n-gram baselines.
Why it matters here · The first neural language model — the direct ancestor of the whole LLM line.
AI