The human journey toward artificial general intelligence
Can a machine learn its way through a problem its builders never anticipated? Follow the people who have tried to make that question precise, the machines they built, and the tests that changed their minds.
52 milestones · 72 annotated sources · Last editorial review Nov 19, 2026
Artificial general intelligence (AGI) is a research goal. A working description here is a machine that can learn and use skills across many unfamiliar tasks. Researchers disagree about how wide that range must be, how quickly a system must learn, and how much human help counts. A strong claim names a definition and a test. These records document proposals, experiments, public events and criticisms. Whether any system satisfies a particular definition requires a separate argument.
Choose a reading path
I’m new to AI
Read these eight short entries in order. After each, describe one thing the machine did and one thing a person supplied.
Follow a game position, a wrong prediction and a retrieved passage as three different sources of improvement or information. The final stops connect the history to prompting, retrieval and tool use.
Compare a formal definition, a surprising game result, a new-task benchmark and a test of sustained work. End by comparing two metrics for the same modern release.
Can the same agent achieve goals in many settings?
Legg and Hutter define intelligence mathematically by considering rewards across possible environments and giving simpler environments more weight. The ideal measure is a way to reason about generality. Applying it directly to a product is computationally impractical.
How much does the system have to learn from scratch?
Chollet asks how efficiently a system can acquire an unfamiliar skill given its prior knowledge and experience. On his ARC puzzles, someone can change the rule while leaving the colored squares and interface the same. What matters is whether the solver discovers the new rule.
Morris and colleagues distinguish level of performance from range of capabilities. They also ask how much autonomy a deployment gives the system. A person can grant software purchasing permission without changing its underlying ability to plan a purchase.
This is a selected history of ideas that matter to the AGI question. Many participants set out to prove a theorem, translate a language, play a game or solve a scientific problem. Their own goals appear in the records. A year alone means the record is dated to that year. Within a year, finer dates are used only when the source supports them.
Before researchers could train a model, they had to describe a procedure, a signal, and a question a machine could be tested against. A mathematical possibility, a model of a neuron and a test of conversation each answer a different question.
Question to carry forward. If a machine follows a procedure for composing music, what evidence would tell you it learned something?
1843 · Lovelace describes a machine that could work with more than numbers
Published notes · Ada Lovelace, Charles Babbage and Luigi Menabrea
Babbage designed the Analytical Engine. Lovelace translated Menabrea’s account and added notes explaining how operations could act on relationships expressed as symbols. She worked through a procedure for Bernoulli numbers and considered how musical relationships might be represented.
Why it mattered. A machine can manipulate a representation of something even when it has no direct experience of that thing. That distinction resurfaces whenever an assistant writes convincingly about the world.
Where the evidence stops. The Bernoulli table was a written procedure for the proposed engine. The music passage was a conditional possibility. Neither documents a machine composing music in 1843.
1854 · Boole gives rules for calculating with logical statements
Book publication · George Boole
Boole’s Laws of Thought used algebraic signs to express classes and rules for combining them. A human could manipulate a symbol by a stated rule while preserving a logical relationship.
Why it mattered. Suppose one group contains everything that is red and another everything that is round. The things that are both red and round form another group. Writing that operation symbolically makes part of reasoning into a procedure.
Where the evidence stops. The book develops a formal system for particular relationships. It leaves open how a person or machine decides which objects, symbols and rules belong in a real problem.
May 28, 1936 · Turing asks what one machine can compute
Manuscript received · Alan Turing and the London Mathematical Society
Turing described an abstract machine that follows a finite set of rules. He showed how a universal machine could simulate any machine of the specified kind from its description.
Why it mattered. A single design can perform many different procedures when supplied with different instructions. This is an important sense of general purpose that predates training machines on data.
Where the evidence stops. The paper concerns what can be computed in principle. It supplies neither a practical time budget nor a method for learning which procedure solves a new task. The date is the manuscript’s receipt date.
1943 · A network of idealized neurons can implement logic
Mathematical model · Warren McCulloch and Walter Pitts
McCulloch and Pitts represented nerve cells as simplified units that switch on or off. They described how arrangements of such units could implement logical relationships.
Why it mattered. A calculation could be spread across a network. Asking how its connections should be learned rather than hand-specified would become a separate research problem.
Where the evidence stops. The units were deliberately simplified. The construction did not train a network on examples or establish that living brains reason in the same way.
Feb 1946 · The ENIAC programmers turn a procedure into a working electronic calculation
Public demonstration · Jean Jennings Bartik, Frances Bilas and their fellow ENIAC programmers at the University of Pennsylvania
Six women learned to program ENIAC by working out how to connect its units and set its switches. In her oral history, Bartik recalled getting the machine ready for its public unveiling. The machine’s visible speed rested on extensive preparation and debugging.
Why it mattered. Here is the labor behind the phrase “the computer calculated.” People had to translate a method into a physical arrangement, find the errors and decide whether to trust its output.
Where the evidence stops. ENIAC’s public demonstration documented electronic computation. The programmers supplied the procedures. The event is a foundation for later AI work, with a different aim from learning a procedure independently.
1948 · Wiener puts feedback at the center of control
Book publication · Norbert Wiener and the interdisciplinary cybernetics community
Wiener’s Cybernetics brought control and communication in animals and machines into one discussion. In a feedback loop, an action changes what the system senses next.
Why it mattered. Think of a thermostat. Its last action affects the next temperature reading. A robot or game player must likewise deal with the consequences of its own actions.
Where the evidence stops. A controller can regulate one variable successfully with a fixed rule. Feedback by itself establishes little about learning a new goal.
Oct 1950 · Turing makes machine intelligence a question about observable behavior
Research proposal · Alan Turing and the readers of Mind
Turing proposed an imitation game conducted through written exchanges. Later in the same paper he considered constructing a learning machine by educating a simpler starting system.
Why it mattered. Two distinct questions emerge. Can a person tell the machine’s answers apart in a particular setting? How might the machine acquire the abilities that produced them?
Where the evidence stops. Turing described a proposed test and a possible research programme. The paper reports no machine passing a universally accepted test for general intelligence.
Several ideas competed from the start. One team searched for proofs, another learned from games, and chemists taught a program which structures were plausible. Translators, funders and critics asked what these systems delivered outside a demonstration.
Question to carry forward. Which part of a system’s apparent skill did a person supply in advance?
Aug 31, 1955 · Four researchers name an ambitious summer project
Dated proposal · John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon
Their proposal used the phrase artificial intelligence and asked for a summer study at Dartmouth the following year. It listed problems including language, neural networks and self-improvement.
Why it mattered. Several competing approaches already fit under one name. The proposal lets a reader see the researchers’ ambitions in their own document before later successes or disappointments colored the story.
Where the evidence stops. Their belief that aspects of intelligence could be described precisely enough for a machine was a research premise. The date belongs to the proposal. The meeting followed in 1956.
1956 · Dartmouth gives researchers a shared meeting place
Summer workshop · Dartmouth participants and organizers in Hanover, New Hampshire
The summer research project took place in 1956. Dartmouth’s historical account identifies the gathering as an organizing moment for the field.
Why it mattered. Institutions, funding proposals and a common name helped people working on different questions recognize one another’s work. Earlier work on computation and machine intelligence was already on record.
Where the evidence stops. The proposal and workshop mark different events. Neither document establishes that all the participants agreed on one method or achieved the proposal’s scientific ambitions.
Oct 10, 1956 · The Logic Theory Machine searches for proofs
Technical report · Allen Newell, J. C. Shaw and Herbert A. Simon at RAND and Carnegie Tech
The team reported experiments with a program that chose promising steps in a search for proofs in a formal logic. Their report describes the program’s search effort as well as its successes.
Why it mattered. A machine could produce a sequence of valid steps that a person could check. Choosing which branch to try next mattered because blindly trying everything was expensive.
Where the evidence stops. The authors supplied a formal language, allowed moves and a target. A correct proof in that setting does not show the machine can formulate a useful question in an unfamiliar subject.
1958 · The perceptron makes learning part of the design
Research paper · Frank Rosenblatt at the Cornell Aeronautical Laboratory
Rosenblatt analyzed a model of pattern recognition in which experience changes parts of a network. He asked which environmental distinctions a network could learn to recognize.
Why it mattered. Instead of writing a separate rule for every example, a designer could choose a learning mechanism and let examples influence the decision. This is a useful way to approach a modern classifier.
Where the evidence stops. The paper discusses limits on distinguishing relationships. Recognizing a learned pattern and explaining its meaning are different demands on a system.
Samuel described a program that stored information about positions and experimented with learning how to value a board position. It could use experience from play to guide later move choices.
Why it mattered. In a game, the rules and eventual outcome are clear. The hard part is recognizing which earlier move deserves credit. That question connects Samuel’s work to later reinforcement learning.
Where the evidence stops. People still supplied the game, legal moves and a representation of the board. Improvement in checkers leaves transfer to another game untested.
1965 · Chemists and computer scientists combine their expertise in DENDRAL
Collaboration begins · Joshua Lederberg, Edward Feigenbaum and collaborators including Bruce Buchanan and Georgia Sutherland
Lederberg dated the start of the Stanford collaboration to early 1965. Later reports described programs that generated candidate molecular structures subject to chemical constraints.
Why it mattered. A specialist knows which possibilities can be ruled out before a search begins. Making that knowledge explicit can turn a huge search into a tractable scientific aid.
Where the evidence stops. The entry dates the collaboration. The linked 1968 program record documents a later stage. The success of a chemistry-specific representation leaves other scientific domains open.
Jan 1966 · ELIZA makes a conversation feel personal
Research paper · Joseph Weizenbaum at MIT and the people who conversed with his program
ELIZA matched keywords and rearranged text according to a script. Its well-known DOCTOR script could turn a user’s words into a question, encouraging the user to supply more of the conversation.
Why it mattered. Read the published dialogue once as a participant and again looking for substitutions. A listener can supply a surprising amount of the meaning attributed to a program.
Where the evidence stops. The paper documents a conversational effect and the mechanism behind it. Apparent empathy in this example cannot establish understanding or genuine care.
1966–1972 · Shakey connects sensing, planning and action
Research programme · Charles Rosen, Nils Nilsson and the SRI team, supported by ARPA
From 1966 through 1972, SRI researchers built and studied a mobile robot that could perceive parts of a room, plan a route and rearrange simple objects.
Why it mattered. A robot’s plan has to survive contact with the world. Moving an object changes what the next observation and the next action should be.
Where the evidence stops. The team used a controlled setting and a limited set of objects and actions. Route planning there did not settle how a robot would handle an unfamiliar home.
1966 · US reviewers challenge the case for automatic translation
Public evaluation · The Automatic Language Processing Advisory Committee, translators and US research sponsors
ALPAC examined translation needs, costs and quality in the United States. Its report was skeptical of then-current automatic translation and recommended support for work including aids for human translators and basic research.
Why it mattered. An institution can change which problems receive attention by asking about the work people actually need done. The translator’s time and the quality of the final text both matter.
Where the evidence stops. The report concerned a particular US funding and language-service context. Hutchins’s later historical analysis helps distinguish its actual recommendations from the sweeping story that all translation research stopped.
1973 · The Lighthill report challenges AI’s ability to scale
Commissioned review · James Lighthill, the UK Science Research Council and later parliamentary historians
Lighthill observed that a method that works on a small example may run out of time as the number of possibilities grows. The UK review criticized the gap between certain demonstrations and broader ambitions.
Why it mattered. Imagine trying every route through a few towns, then through fifty. The rule for finding the best route can stay the same while the work required becomes prohibitive.
Where the evidence stops. A later House of Lords review explicitly treats the causal link between the report and reduced AI research funding as unclear. The document verifies a major criticism. A single worldwide funding-collapse claim needs other evidence.
1982–1998 · Rules, probabilities and learned patterns
No single method replaced all the others. A national project pursued logical inference, researchers taught networks to correct their errors, and chess engineers combined search with human expertise. Different designs made different mistakes visible.
Question to carry forward. What would need to change to move a chess engine, a document reader or a belief network to a new task?
1982 · Japan pursues knowledge processing through logic and parallel machines
National research programme · Japan’s Ministry of International Trade and Industry, ICOT researchers and Shunichi Uchida
Japan launched the Fifth Generation Computer Project. Its researchers explored machines that could apply logical inferences and carry out parts of the work in parallel. Uchida’s 1987 report describes hardware under development.
Why it mattered. The future of AI once looked like purpose-built machines answering logical queries. This programme shows a serious alternative to today’s training of large neural networks.
Where the evidence stops. The museum record dates the launch. The technical report establishes what engineers were building. Project ambitions and achieved general intelligence are separate historical claims.
1986 · Networks learn internal representations from their errors
Research paper · David Rumelhart, Geoffrey Hinton and Ronald Williams
The authors demonstrated how a difference between a network’s prediction and its target could be propagated backward to adjust connections in several layers. Internal units learned useful patterns in the examples.
Why it mattered. A builder could provide targets without hand-designing every intermediate feature. This helps explain the change from supplying a whole set of rules to supplying examples and a way to score mistakes.
Where the evidence stops. The authors demonstrated their procedure on particular problems. Establishing who first developed any of its underlying mathematical ideas would require additional historical evidence.
1988 · Belief networks make uncertainty part of reasoning
Research monograph · Judea Pearl and the probabilistic reasoning community
Pearl’s book set out how a network of related variables can represent uncertain beliefs and update them when evidence arrives. New observations can change the plausibility of several explanations at once.
Why it mattered. Suppose a radio goes silent. Discovering that its display is also dark shifts the balance between an audio problem and a power problem. A system can express that change without declaring either diagnosis certain.
Where the evidence stops. The results depend on how the variables, dependencies and probabilities were specified. A meaningful probability model for one device does not automatically transfer to another.
1988 · A learner updates a prediction before the final outcome arrives
Research paper · Richard Sutton, building on earlier prediction work including Samuel’s checkers
Sutton analyzed methods that use the difference between successive predictions to update an earlier estimate. His paper explicitly connects this idea to earlier game-playing research.
Why it mattered. A learner predicting the outcome of a trip can update its estimate when it encounters a delay, before the trip actually ends. This helps explain how feedback can travel across a sequence of decisions.
Where the evidence stops. The paper proves results under specified conditions and studies prediction problems. A general ability to choose goals or act reliably in arbitrary environments does not follow from those results.
May 1997 · Deep Blue wins the rematch with world chess champion Garry Kasparov
Match result · Garry Kasparov, Feng-hsiung Hsu, Murray Campbell, Joe Hoane and IBM’s human chess advisers
Kasparov won the 1996 match despite losing its opening game. In the 1997 rematch, Deep Blue won the match. Its specialized hardware, search and chess knowledge were the work of people with different skills.
Why it mattered. A public match made one capability hard to dismiss. The result also gives a clean demonstration of the difference between expert performance in one domain and breadth across domains.
Where the evidence stops. An individual game, a match and a claim about intelligence are three different units of evidence. The match tested chess under agreed rules.
Research paper · Yann LeCun, Léon Bottou, Yoshua Bengio and Patrick Haffner
The authors described convolutional networks for recognizing handwriting and systems that combined learned components for document reading. Looking for similar features at different image positions reduced the need to relearn each feature everywhere.
Why it mattered. A recognizer can read the same kind of digit written in many different ways. Understanding what a number on a document means in the world requires additional context.
Where the evidence stops. A system scored on labeled documents establishes performance on that recognition task. It cannot answer arbitrary questions about the author or the transaction from that score alone.
2004–2016 · Who supplies the data, and who defines success?
Field trials in desert driving and shared image collections gave teams common tests. Meanwhile other researchers formalized what they meant by generality. Race organizers, data workers and human opponents all helped set the terms of comparison.
Question to carry forward. How much of a result belongs to the algorithm, the training examples, the hardware or the person who designed the test?
2004–2005 · Two desert races turn a failure into a measurable result
Field challenges · DARPA organizers, the competing vehicle teams, and Stanford’s Stanley team led by Sebastian Thrun
No vehicle finished DARPA’s 2004 Grand Challenge. In October 2005 five teams finished a 132-mile course, with Stanford’s Stanley first. The team’s paper describes the combination of sensors, learned methods and probabilistic reasoning used for desert driving.
Why it mattered. For a learner interested in racing, the finish line makes success concrete. A public failure, a changed course and a second round provide more information than a single successful demonstration.
Where the evidence stops. The two courses were different. The 2005 finish demonstrated driving under desert race conditions. Everyday traffic, weather and interactions with other drivers require separate evidence.
2007 · Researchers organize explicitly around artificial general intelligence
Edited research volume · Ben Goertzel and Cassio Pennachin, editors, and the volume’s contributors
An edited research volume titled Artificial General Intelligence brought different proposed routes to broadly capable systems together in one place. Its contents preserve approaches beyond the neural networks familiar to today’s readers.
Why it mattered. A shared goal can organize a research community even when its members disagree about architectures or how to measure success. Read two chapter descriptions and compare the problems they set out to solve.
Where the evidence stops. The book is evidence that the name and research agenda were in use by 2007. A publication date alone cannot establish who coined a term or prove that any approach worked.
2007 · Legg and Hutter define intelligence across environments
Formal definition · Shane Legg and Marcus Hutter
Their mathematical definition considers how well an agent achieves goals over a broad collection of possible environments, weighting simpler environments more heavily. It makes an intuitive word into a precise theoretical object.
Why it mattered. A solver that memorizes the answers to one game can excel there. The question changes when the environment itself can change. This definition makes that difference explicit.
Where the evidence stops. The idealized measure is not directly computable in full. It is one definition with assumptions about environments, rewards and weighting. Other measurements ask different questions.
2009 · ImageNet makes image collections and labeling labor visible
Dataset paper · Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, Fei-Fei Li and crowd workers
The team organized images using WordNet categories. People working through Amazon Mechanical Turk verified candidate images. The paper reported about 3.2 million images in the collection it described at the time.
Why it mattered. Training and comparing recognizers at scale depended on choices about words, example images and who checked them. A later audit of ImageNet’s people categories found reasons to remove or reconsider categories and examples.
Where the evidence stops. A label records a decision within a collection. It can encode ambiguity or harm. The 2009 figure describes that paper’s dataset, not every later ImageNet edition or competition subset.
2012 · AlexNet shows what data, graphics processors and a deep network can do together
Competition and conference paper · Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, using the ImageNet community’s data
A deep convolutional network trained on GPUs achieved a striking ImageNet competition result. The team reported top-five classification error, which asks whether the correct category appears among the model’s five highest-ranked guesses.
Why it mattered. The result brought together earlier ideas about layered recognition, a shared labeled dataset and enough computation to train a large model. Comparing methods on the same task made the improvement legible.
Where the evidence stops. Picking a category from a supplied list tests recognition under those categories. It does not test whether the machine can describe an unfamiliar scene or learn a new category from one example.
2014 · A translator learns which source words matter at each step
Preprint · Dzmitry Bahdanau, Kyunghyun Cho and Yoshua Bengio
The researchers trained a translation model that could weight different parts of a source sentence as it generated each part of the translation. They illustrated these learned alignments in translation examples.
Why it mattered. Consider two languages that put the verb in different positions. A changing set of weights lets the decoder draw on the relevant source words without requiring a human to pair every word in advance.
Where the evidence stops. This attention mechanism preceded the 2017 Transformer. The reported experiments concern translation and do not establish that the weights fully explain every decision.
2015 · One learning recipe is tested across 49 Atari games
Research experiment · Volodymyr Mnih and colleagues at DeepMind
The team used pixels, game scores and a common learning recipe across 49 Atari games. They trained agents for the individual games and reported scores against human and algorithmic baselines.
Why it mattered. This broadened the question from whether a specialist could play one game to how broadly its training procedure could be reused. Per-game scores show variation hidden by an average.
Where the evidence stops. The same recipe across games is different from a single trained player entering every new game with no further practice. The human comparison also depends on the stated playing protocol.
Mar 2016 · AlphaGo defeats Lee Sedol in a five-game Go match
Match result · Lee Sedol, Fan Hui, David Silver and the AlphaGo team, and the expert players whose games supplied examples
AlphaGo combined a search through possible moves with neural networks trained using expert games and self-play. It beat Lee Sedol 4–1 in Seoul. Lee won game four. The earlier Nature paper reported the separate 5–0 match against Fan Hui.
Why it mattered. The combination matters. Human game records supplied examples, self-play supplied practice and search helped choose a move. One can ask what each contributed and which parts would survive a rule change.
Where the evidence stops. These results belong to a system trained for Go. The documented original used human game records. Later versions with different training methods require their own evidence.
Language models began reusing what they learned across tasks. At the same time, auditors found gaps hidden by averages, and language communities showed how much it matters who collects data and judges the answers.
Question to carry forward. When an answer looks right, what shows that the skill transfers to a different person, language or problem?
2017 · The Transformer reorganizes how a model handles a sequence
Conference paper · Ashish Vaswani, Noam Shazeer, Niki Parmar and five coauthors
The Transformer used attention to connect positions in a sequence and removed recurrent processing from its core translation architecture. The team evaluated English-to-German and English-to-French translation.
Why it mattered. Words far apart can affect each other. Processing relationships in parallel during training made this design attractive for large text models. The original translation task is an accessible place to see the idea at work.
Where the evidence stops. The paper demonstrates a particular architecture and translation results. It did not introduce attention itself or demonstrate a general conversational assistant.
2018 · Gender Shades asks whose errors are hidden by an average
Independent audit · Joy Buolamwini, Timnit Gebru and the people represented in the audit dataset
The authors tested three commercial systems that assigned binary gender labels from faces. In their evaluation, error reached 34.7% for the group they labeled darker-skinned female. The highest error for their lighter-skinned male group was 0.8%.
Why it mattered. Even a headline accuracy that looks strong can hide systematic failures. Reporting the task, the people represented and the subgroup outcomes changes what a reader can reasonably conclude.
Where the evidence stops. These are the authors’ categories and results for the products and datasets tested. The task was binary label assignment from images. It did not measure identity recognition or establish anyone’s gender.
Oct 11, 2018 · BERT shows how the same pretrained model can be adapted to several language tasks
Preprint · Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova
BERT learned representations using words on both sides of a position during pretraining. The team then fine-tuned a pretrained model for different tasks. The first preprint appeared in 2018 and the NAACL paper in 2019.
Why it mattered. Training on ordinary text could provide a reusable starting point. In this setup, adapting the model to a supervised task still involved updating its weights.
Where the evidence stops. Results came from specified language benchmarks. This training and fine-tuning procedure differs from giving a frozen model examples inside a prompt.
Nov 5, 2019 · ARC asks how efficiently a solver can learn an unfamiliar rule
Definition and benchmark · François Chollet
Chollet argued for judging the acquisition of skill in light of prior knowledge and practice. His Abstraction and Reasoning Corpus presented small grids where a solver inferred a transformation from a few worked examples.
Why it mattered. You can try the underlying idea with colored squares and pencil. Solving one new example tests application of a rule. Needing to discover a different rule tests adaptation.
Where the evidence stops. The benchmark encodes design choices about what prior knowledge people share. A score measures success on this family of tasks under a stated protocol.
May 22, 2020 · Retrieval gives a language model an external source to consult
Research preprint · Patrick Lewis and colleagues
The researchers paired a text generator with retrieval from an external collection. The system selected passages relevant to a question and used them while generating responses on knowledge-intensive tasks.
Why it mattered. Imagine supplying the correct revision of a lab manual when asking a wiring question. The content the system can retrieve and the model’s ability to interpret it can be evaluated separately.
Where the evidence stops. Retrieval can return the wrong passage. Even a relevant passage can be misread. A grounded answer requires checking that its actual claim follows from its source.
Research paper · Tom Brown and colleagues at OpenAI and collaborating institutions
GPT-3 was evaluated on tasks presented as text with instructions or a small number of worked examples. In the few-shot setting, the researchers supplied those examples at evaluation time without updating the model’s weights.
Why it mattered. A reader could now teach the format of a task in the request itself. For an Academy guide, the distinction between an example in the prompt and additional model training matters in everyday use.
Where the evidence stops. Performance varied across tasks. A few accurate answers can be consistent with several explanations, including a familiar task format. They do not establish reliable learning on every novel problem.
Nov 2020 · African language communities help decide how translation is built and judged
Community research · Wilhelmina Nekoto and the Masakhane research community, including translators and participants without formal machine-learning training
The project reported machine-translation benchmarks for more than 30 African languages and human evaluations for about a third of them. Participants helped produce data, models and evaluations.
Why it mattered. A system cannot be called broadly useful on the strength of a few widely resourced languages. Speakers’ judgments and access to relevant examples change both what can be built and what can be measured.
Where the evidence stops. The paper reports different levels of evaluation across languages. The existence of a benchmark for a language does not establish that its translations are ready for every setting.
2021 · Researchers challenge the assumption that scale alone delivers understanding
Research position paper · Emily Bender, Timnit Gebru, Angelina McMillan-Major and Shmargaret Shmitchell
The authors argued that expanding language-model training raises questions about data documentation, costs, whose language is represented, and what fluency can lead people to infer. They proposed considering these choices before scaling up.
Why it mattered. The production of training text and the people affected by a system belong in its technical history. A reader assessing an assistant can ask who supplied the text and evidence used to interpret its output.
Where the evidence stops. This is a documented scholarly argument. Treat its recommendations as arguments to examine. Individual empirical claims about a later model still require evidence for that model.
Jul 2021 · An independent scientific challenge tests AlphaFold’s predictions
Research paper and blind assessment · John Jumper and the AlphaFold team, CASP assessors and experimentalists who determined the reference structures
AlphaFold2’s developers reported accurate predictions of the three-dimensional shapes of many proteins in the 2020 CASP14 assessment. The 2021 methods paper and the organizers’ account give two views of the result. The structures used to check the predictions had been withheld from entrants.
Why it mattered. A scientific test can be designed so that the answer already exists in a lab and is hidden until predictions are submitted. That is a powerful alternative to judging the plausibility of an impressive-looking output.
Where the evidence stops. This was evidence for structure prediction on the assessed targets. An accurate prediction on this task does not establish the ability to choose and carry out an unrelated scientific investigation.
2022 onward · Assistants, agents and the tests of generality
Chat interfaces, feedback from people and access to tools changed what users could do. Evaluators now have to specify the model, the tools it could use, the help it received, the budget and the tasks on which it failed.
Question to carry forward. What observation would make you revise your judgment of a modern model’s generality?
2022 · Human demonstrations and rankings teach instruction following
Research paper · Long Ouyang and colleagues, the human labelers and the people who supplied prompts
For InstructGPT, people wrote examples of desired behavior and ranked outputs. The researchers used this feedback to further train language models. On the tested prompt distribution, raters preferred some outputs from a much smaller tuned model to those from a larger untuned model.
Why it mattered. What a user finds helpful depends on more than scale or next-word prediction. An evaluation should ask who rated the answers, which prompts they saw and how the alternatives were presented.
Where the evidence stops. A preference result reflects the selected prompts and raters. An answer can win a preference comparison and still contain a factual error.
Mar 2022 · A smaller model trained on more text can win at the same training budget
Scaling experiment · Jordan Hoffmann and colleagues at DeepMind
The team compared ways of allocating a fixed amount of computation between model size and training text. Their 70-billion-parameter Chinchilla model was trained on about four times as much data as the 280-billion-parameter Gopher and outperformed it on many reported tasks at a similar training-compute budget.
Why it mattered. A parameter count alone leaves out how much data the model saw and how resources were spent. Those variables can change the ranking even before considering the tools available to a user.
Where the evidence stops. The comparison supports a conclusion about the tested budgets, models and tasks. Extending it indefinitely or declaring a general-intelligence threshold would require different evidence.
Oct 6, 2022 · ReAct interleaves model output, actions and observations
Research preprint · Shunyu Yao and colleagues
ReAct prompted language models to alternate between working through a task, taking permitted actions and using the returned observations. The authors tested settings including question answering with an external knowledge source and interactive simulated tasks.
Why it mattered. A tool can change what a system knows during an interaction. To compare assistants fairly, record which tools were available, what those tools returned and how many attempts each system had.
Where the evidence stops. The experiments use specified actions and environments. A displayed reasoning trace is an observable output and cannot by itself establish what caused an internal decision.
Nov 9, 2022 · BigScience documents an international approach to building BLOOM
Project research paper · BigScience Workshop participants, multilingual data contributors and the French public computing infrastructure
The BigScience collaboration documented a multilingual model and shared model artifacts and a detailed model card. It trained BLOOM on France’s Jean Zay public supercomputer. The date marks the project paper’s first preprint.
Why it mattered. Readers can inspect model documentation to ask whose languages were included, how computing time was supplied and what they are permitted to do with the weights. Building a frontier-scale model also required public infrastructure and coordinated labor.
Where the evidence stops. Access to model weights, training material and the hardware needed to repeat training are distinct. A language appearing in a list is no guarantee of high performance for a particular speaker or task.
Nov 16, 2022 · HELM makes the choice of metrics part of the result
Evaluation study · Percy Liang and the Stanford-led HELM collaboration
HELM organized language-model evaluations by scenarios and measured criteria including accuracy, robustness, fairness and efficiency. It reported gaps in what had been evaluated and compared systems under a common framework.
Why it mattered. A traveler choosing a translation tool might value correctness, speed and cost differently. A single rank silently chooses how much each criterion matters. Several measurements make that choice visible.
Where the evidence stops. A common framework can improve comparison while still leaving languages, tasks and deployment conditions uncovered. The benchmark’s coverage is part of the finding.
Nov 30, 2022 · ChatGPT gives people a conversational way to try instruction following
Public research preview · OpenAI’s ChatGPT team, the trainers and users who provided feedback
OpenAI released ChatGPT as a research preview with dialogue and follow-up turns. Its launch post included examples and acknowledged plausible-sounding errors and sensitivity to wording.
Why it mattered. A non-programmer could now try the same question with a different instruction and compare the replies. What changed here was public access to an interface as well as the behavior of a trained model.
Where the evidence stops. This date refers to the ChatGPT preview. A fluent exchange is an observation about the interface and its responses. Checking a particular factual answer requires evidence outside the conversation.
Mar 15, 2023 · GPT-4’s report describes image and text inputs
Technical report · OpenAI’s GPT-4 research team
The technical report described a model accepting images and text and producing text, with results across several exams and benchmarks. It also described limitations and withheld substantial training and architecture details.
Why it mattered. An assistant could in principle answer about a diagram as well as a written paragraph. That creates a second question for a user comparing platforms. Which input types did the tested interface actually expose?
Where the evidence stops. The paper documents a reported capability. Its publication date cannot establish universal product access, the provenance of every answer, or independent reproduction of its training run.
2024 · An AGI framework separates breadth, depth and deployment autonomy
Conference position paper · Meredith Ringel Morris and colleagues
A November 2023 preprint appeared as an ICML position paper in 2024. The authors proposed levels of performance and generality and discussed how these interact with the autonomy granted to deployed systems.
Why it mattered. Try describing a chess champion and a helpful but inconsistent all-purpose assistant on two axes. The chess system can have extraordinary depth and narrow breadth. Giving an assistant more permissions introduces a different consideration.
Where the evidence stops. The levels are a proposed framework. Placing a system at any level requires a definition of the tasks, comparison group and evidence. Other definitions of AGI remain in use.
Jan 2025 · DeepSeek distinguishes two recipes for training reasoning models
Technical report · DeepSeek-AI researchers and the people who prepared data and evaluation tasks
The team reported that R1-Zero applied reinforcement learning to a pretrained base without a preliminary supervised fine-tuning stage. The separate R1 system used initial demonstration data and a multistage training process. The report described strengths and problems such as language mixing in R1-Zero.
Why it mattered. Small differences in training stages can change what a headline means. Ask which model, which initial training and which task produced the reported result.
Where the evidence stops. These are the developers’ training descriptions and benchmark results. Omitting one supervised stage does not erase the base model’s training data or the people who constructed the evaluation.
Mar 18, 2025 · METR measures task difficulty in human time
Independent evaluation · Thomas Kwa and colleagues at METR, and the experts timed on reference tasks
The researchers related agent success on a particular collection of tasks to the time human experts needed for those tasks. Their 50% horizon estimates the human task duration at which the fitted model predicts a one-in-two chance of agent success.
Why it mattered. If an evaluator says a task lasts an hour, ask whose hour. Human experience, task selection and the required success probability can all change the interpretation.
Where the evidence stops. The horizon is not the wall-clock runtime of the AI. It is an estimate for the study’s task distribution. Extrapolating a trend to an AGI arrival date adds assumptions the experiment does not test.
Mar 25, 2026 · ARC-AGI-3 makes exploration an action-by-action test
Benchmark launch · ARC Prize’s benchmark team, environment creators and human test participants
ARC Prize launched interactive tasks where a solver must explore unfamiliar environments, infer a goal and act under learned rules. Its subsequent technical report describes how scoring uses human action baselines and accounts for efficiency.
Why it mattered. A static answer and a sequence of exploratory actions reveal different failures. A solver can now waste moves gathering information, pursue the wrong goal or adapt efficiently after an unexpected observation.
Where the evidence stops. The launch and the later technical report are separate records. The evidence comes from the benchmark’s creators and describes performance under their tasks and scoring. It supplies no universal threshold for AGI.
Nov 17, 2026 · Two indices tell different stories about the same release
Contemporary comparison · OpenAI’s GPT-6.2 team and the independent evaluators at Artificial Analysis
After the release of GPT-6.2 Astra and Sol, Artificial Analysis reported that Astra gained five points over GPT-6 Astra on its Intelligence Index. On its Coding Agent Index, Astra was flat against the same predecessor in the Codex harness at maximum effort. The date is the evaluator’s report.
Why it mattered. This is a worked example for reading a weekly review. Ask which tasks changed, which stayed flat, what tools were allowed and whether the comparison resembles the job you care about.
Where the evidence stops. These are two named indices at one date with particular evaluation settings. Neither score determines the outcome for all programming tasks or answers whether a model meets an AGI definition.
On paper, replace “I am …” with “Why are you …?” Try it on “I am curious” and “I am fixing a radio.” What would a listener infer? Compare your two lines with the ELIZA transcript.
Check your reasoning. The same substitution works on both strings. The second reply can sound odd. Neither exchange shows whether the rule knows what a radio is.
Show a friend two drawings of shapes moved one square to the right. Ask them to transform a new drawing. Then give examples that imply a different rule, such as making a mirror image. Record how many examples or hints each of you needed.
Check your reasoning. Report the number of examples, the hidden rule and any hints. Success with a new drawing under the old rule and success after the rule changes are different pieces of evidence.
Read the METR abstract. Write down whose time was measured, what the stated success probability represents and which kinds of task were sampled. Name one task from your own life that the study did not test.
Check your reasoning. Human experts supplied the task durations. The 50% horizon is a modeled success threshold on the sampled tasks. It gives no completion probability for your own untested task.
Algorithm. A sequence of rules for solving a problem. A chess search can apply rules for legal moves and for choosing which move to examine next.
Heuristic. A shortcut that directs effort toward a promising answer. It can save a search through millions of possibilities and can also miss the best answer.
Model and system. A model is a learned or specified way to make predictions or decisions. A deployed system may add a search engine, documents, instructions, memory or human approval around it.
Training and inference. Training changes a model using examples or feedback. Inference uses the resulting model to produce an answer. Supplying an example in a chat can change the answer without updating the model’s weights.
Weights and parameters. Stored numbers a model uses to process its input. Learning can change the strength of a connection, much as adjusting a mixer changes how strongly a sound contributes to the final mix.
Generalization and transfer. Generalization means doing well on new examples. Transfer asks whether something learned in one setting helps on a different task. Reading a new handwritten digit and learning a new game test different kinds of novelty.
Benchmark and baseline. A benchmark specifies tasks and scoring. A baseline is a reference result, such as the score of a person, an earlier system or a simple rule, measured under stated conditions.
Attention. A calculation that weights relationships between pieces of an input. In translation, the words most relevant to the next output word can change from one step to the next.
Reinforcement learning. Learning to choose actions using feedback called reward. Designers choose what earns a reward, so a high score can reflect their choices as well as the system’s ability.
Agent and autonomy. An agent acts and observes the results. Autonomy describes the freedom people give it to act. Access to more tools can improve an outcome without changing the model.
Training exposure. A test item, or something very close to it, may have appeared during development. An unseen puzzle, a held-out task or an independent replication can make a generality claim easier to evaluate.
Graphics processing unit (GPU). A processor that can do many similar calculations in parallel. The repeated numerical operations used to train neural networks are a good fit for this kind of hardware.
Open questions for the next claim
What would count as an unfamiliar task?
Look for a protocol that withholds the rules, documents relevant training exposure and records how many examples or hints the solver receives. Changes to the protocol can alter the conclusion.
Compare a model with and without access to tools or reference material. Preserve the inputs, budget, model version and scoring method. A gain can come from the surrounding system.
A forecast needs a stated definition, assumptions about future progress and a way it could turn out wrong. The timeline records forecasts as attributed proposals only when historically consequential. A milestone requires evidence of something that actually happened.
An accessible independent overview. Compare its compact historical narrative with the original documents linked under the early milestones. A Short History of AI
An edited collection for readers who want to explore approaches that differ from today’s chat assistants. Artificial General Intelligence
Each note identifies the author or institution, the kind of record and where to check the claim. The source year can differ from the event year. Several later histories preserve evidence about earlier events.
1. Notes on the Analytical Engine
Ada Lovelace, in her translation of Luigi Menabrea · Contemporary publication · 1843 · Checked Nov 19, 2026
Read Note A on representing relations and Note G on Bernoulli numbers. Read the source
2. Notes
University of Oklahoma Libraries, Galileo’s World · Museum collection · Checked Nov 19, 2026
See the attribution and the description of the original printed work. Read the source
3. An Investigation of the Laws of Thought, chapter II
George Boole, Cambridge reproduction · Original publication · 1854 · Checked Nov 19, 2026
Read the chapter summary on classes, signs and the rules for combining them. The book metadata gives the original year. Read the source
4. On Computable Numbers, with an Application to the Entscheidungsproblem
Alan Turing · Original paper · 1936 · Checked Nov 19, 2026
See the received and read dates on page 230 and section 6 on the universal machine. Read the source
5. A Logical Calculus of the Ideas Immanent in Nervous Activity
Warren McCulloch and Walter Pitts · Original paper · 1943 · Checked Nov 19, 2026
Read the introductory assumptions about all-or-none activity and the relation to logic. Read the source
6. Oral History of Jean Bartik
Jean Jennings Bartik interviewed by Gardner Hendrie, Computer History Museum · Participant oral history · 2008 · Checked Nov 19, 2026
Find the discussion of six women developing ENIAC programming and the account of the February public announcement. Read the source
7. For the Record: Totally Wired
University of Pennsylvania · Institutional history · 2006 · Checked Nov 19, 2026
Read the description of the six women who connected units with cables and set switches. Read the source
8. Cybernetics, or, Control and Communication in the Animal and the Machine
MIT Libraries on Norbert Wiener’s 1948 book · Library collection · 2011 · Checked Nov 19, 2026
Read the account of feedback and control alongside the original book’s publication details. Read the source
9. Computing Machinery and Intelligence
Alan Turing · Original paper · 1950 · Checked Nov 19, 2026
Compare section 1 on the imitation game with section 7 on learning machines. Read the source
10. A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence
John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon · Original proposal · 1955 · Checked Nov 19, 2026
Read the dateline and first two paragraphs, then the list of research problems. Read the source
11. The Founding and Future of AI
Dartmouth College · Institutional history · Checked Nov 19, 2026
See the account of the summer 1956 gathering. Read the source
12. Empirical Exploration of the Logic Theory Machine
Allen Newell, J. C. Shaw and Herbert A. Simon, RAND · Original technical report · 1956 · Checked Nov 19, 2026
The cover of P-951-A is dated October 10, 1956. Read the introduction and the proof-search tests. Read the source
13. The Perceptron
Frank Rosenblatt · Original paper · 1958 · Checked Nov 19, 2026
Read the learning procedure and the discussion of the model’s limitations. Read the source
14. Some Studies in Machine Learning Using the Game of Checkers
Arthur Samuel, IBM · Original paper · 1959 · Checked Nov 19, 2026
Compare the sections on rote learning and learning by generalization. Read the source
15. How DENDRAL Was Conceived and Born
Joshua Lederberg, Computer History Museum archive · Participant retrospective · Checked Nov 19, 2026
Find “early 1965” in Lederberg’s chapter on the beginning of the collaboration. Read the source
16. Heuristic DENDRAL
NASA Technical Reports Server, 1968 conference record · Contemporary technical record · 1968 · Checked Nov 19, 2026
See the program description, authors and conference date. Read the source
17. ELIZA
Joseph Weizenbaum, Communications of the ACM · Original paper · 1966 · Checked Nov 19, 2026
Read the opening dialogue, the description of scripts and the discussion of the user’s interpretation. Read the source
18. Shakey the Robot
Nils Nilsson, editor, SRI · Project retrospective · 1984 · Checked Nov 19, 2026
Read the opening synopsis for the project dates and the robot’s planning and manipulation tasks. Read the source
19. Shakey the Robot
DARPA · Institutional history · Checked Nov 19, 2026
Find Charles Rosen’s proposal and ARPA’s support in 1966. Read the source
20. Language and Machines
Automatic Language Processing Advisory Committee, National Research Council · Original public report · 1966 · Checked Nov 19, 2026
Open the 1966 volume and read its conclusions and recommendations on translation and linguistics. Read the source
21. ALPAC: The (In)famous Report
John Hutchins, MT News International, 1996 · Historical analysis · 1996 · Checked Nov 19, 2026
Read the opening discussion of the US setting and the section revisiting the recommendations. Read the source
22. Artificial Intelligence: A General Survey
James Lighthill for the UK Science Research Council · Original commissioned report · 1973 · Checked Nov 19, 2026
Find the discussion of combinatorial explosion and the report’s treatment of different branches of AI. Read the source
23. AI in the UK: Ready, Willing and Able?
House of Lords Select Committee on Artificial Intelligence, 2018 · Parliamentary retrospective · 2018 · Checked Nov 19, 2026
Read the historical discussion in paragraphs 23–24, especially the caveat about research funding. Read the source
24. Start of the Fifth Generation Computer Project
Information Processing Society of Japan Computer Museum · Institutional history · Checked Nov 19, 2026
Read the opening account of the 1982 launch and the project’s knowledge-processing aims. Read the source
25. Inference Machines in FGCS Project
Shunichi Uchida, ICOT, 1987 · Original technical report · 1987 · Checked Nov 19, 2026
Read the abstract on sequential and parallel inference machines. Read the source
26. Learning Representations by Back-Propagating Errors
David Rumelhart, Geoffrey Hinton and Ronald Williams · Original paper · 1986 · Checked Nov 19, 2026
Read the opening explanation of error signals and the examples of learned internal representations. Read the source
27. Probabilistic Reasoning in Intelligent Systems
Judea Pearl, Morgan Kaufmann · Research monograph · Checked Nov 19, 2026
See the book description and chapters on belief networks and the propagation of evidence. Read the source
28. Probabilistic Reasoning in Intelligent Systems, 1988 edition
University of Bologna research bibliography · Bibliographic record · 1988 · Checked Nov 19, 2026
Read the original Morgan Kaufmann edition’s publication year. Read the source
29. Learning to Predict by the Methods of Temporal Differences
Richard Sutton · Original paper · 1988 · Checked Nov 19, 2026
Read the abstract and introduction on differences between successive predictions and the link to Samuel’s checkers. Read the source
30. Deep Blue
IBM · Participant institution’s history · Checked Nov 19, 2026
Read the separate accounts of the 1996 and 1997 matches and the design team. Read the source
31. Deep Blue team after the 1997 rematch
Computer History Museum · Museum collection · Checked Nov 19, 2026
Read the match result and the caption identifying the human team. Read the source
32. Gradient-Based Learning Applied to Document Recognition
Yann LeCun, Léon Bottou, Yoshua Bengio and Patrick Haffner, Proceedings of the IEEE · Original paper · 1998 · Checked Nov 19, 2026
Read the description of convolutional networks, handwriting and document systems. Read the source
33. Grand Challenge
DARPA · Organizer’s event record · Checked Nov 19, 2026
Compare the March 2004 result with the October 2005 finishers and course. Read the source
34. Stanley: The Robot That Won the DARPA Grand Challenge
Sebastian Thrun and the Stanford Racing Team, Journal of Field Robotics · Participant technical paper · 2006 · Checked Nov 19, 2026
Read the opening description of the two races and the account of the vehicle’s sensors and driving software. Read the source
35. Artificial General Intelligence
Ben Goertzel and Cassio Pennachin, editors, Springer · Research volume · 2007 · Checked Nov 19, 2026
Read the book overview and compare the different approaches in its contents. Read the source
36. Universal Intelligence: A Definition of Machine Intelligence
Shane Legg and Marcus Hutter · Original research paper · 2007 · Checked Nov 19, 2026
Read the definition and the discussion of weighting environments and practical computability. Read the source
37. ImageNet: A Large-Scale Hierarchical Image Database
Jia Deng and colleagues, CVPR · Original dataset paper · 2009 · Checked Nov 19, 2026
Read the abstract and data-collection section on WordNet categories and crowd verification. Read the source
38. Towards Fairer Datasets
Kaiyu Yang and colleagues · Authors’ dataset audit · 2019 · Checked Nov 19, 2026
Read the abstract on the people subtree, unsuitable concepts and representation. Read the source
39. ImageNet Classification with Deep Convolutional Neural Networks
Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, NeurIPS · Original conference paper · 2012 · Checked Nov 19, 2026
Read the abstract and the sections on training across GPUs and ImageNet classification. Read the source
40. Neural Machine Translation by Jointly Learning to Align and Translate
Dzmitry Bahdanau, Kyunghyun Cho and Yoshua Bengio · Original preprint · 2014 · Checked Nov 19, 2026
Read the abstract and the illustrations of alignment between input words and translated words. Read the source
41. Human-Level Control through Deep Reinforcement Learning
Volodymyr Mnih and colleagues, Nature · Original paper · 2015 · Checked Nov 19, 2026
Read the setup, figure 3 and the per-game results in the extended data. Read the source
42. Mastering the Game of Go with Deep Neural Networks and Tree Search
David Silver and colleagues, Nature · Original paper · 2016 · Checked Nov 19, 2026
The abstract describes expert games, self-play, search and the match with Fan Hui. Read the source
43. Google DeepMind Challenge Match: Lee Sedol v AlphaGo
British Go Association · Contemporary match record · Checked Nov 19, 2026
Read the five dated game reports, including Lee Sedol’s game-four win. Read the source
44. Attention Is All You Need
Ashish Vaswani and colleagues, NeurIPS · Original conference paper · 2017 · Checked Nov 19, 2026
Read the abstract and the sections on self-attention and translation trials. Read the source
45. Gender Shades
Joy Buolamwini and Timnit Gebru, PMLR · Independent evaluation · 2018 · Checked Nov 19, 2026
Read the abstract and the results by subgroup for the three commercial systems. Read the source
46. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova · Original preprint · 2018 · Checked Nov 19, 2026
See the October 2018 submission and the description of bidirectional pretraining. Read the source
47. BERT
Devlin and colleagues, NAACL 2019 · Conference publication · 2019 · Checked Nov 19, 2026
Use the proceedings date to distinguish the 2019 conference paper from the 2018 preprint. Read the source
48. On the Measure of Intelligence
François Chollet · Research proposal and benchmark · 2019 · Checked Nov 19, 2026
Read the definitions of skill acquisition and priors, then the introduction of the Abstraction and Reasoning Corpus. Read the source
49. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Patrick Lewis and colleagues · Original paper · 2020 · Checked Nov 19, 2026
Read the abstract and model description, including the difference between learned parameters and retrieved documents. Read the source
50. Language Models Are Few-Shot Learners
Tom Brown and colleagues, NeurIPS · Original conference paper · 2020 · Checked Nov 19, 2026
Read the abstract and the definition of few-shot demonstrations without gradient updates at test time. Read the source
51. Participatory Research for Low-Resourced Machine Translation
Wilhelmina Nekoto and colleagues, Findings of EMNLP · Original community research · 2020 · Checked Nov 19, 2026
Read the abstract, participant roles and which languages received human evaluation. Read the source
52. On the Dangers of Stochastic Parrots
Emily Bender, Timnit Gebru, Angelina McMillan-Major and Shmargaret Shmitchell · Research position paper · 2021 · Checked Nov 19, 2026
Read the abstract and the recommendations about data documentation, costs and affected communities. Read the source
53. Highly Accurate Protein Structure Prediction with AlphaFold
John Jumper and colleagues, Nature · Original research paper · 2021 · Checked Nov 19, 2026
Read the abstract and the account of the blind CASP14 assessment. Read the source
54. Critical Assessment of Methods of Protein Structure Prediction, Round XIV
Andriy Kryshtafovych and colleagues, CASP organizers · Independent assessment · 2021 · Checked Nov 19, 2026
Read the abstract for the assessors’ description of CASP14 and the comparison with experimental structures. Read the source
55. How Have AlphaFold2’s Predictions of Protein Structure Been Validated?
EMBL’s European Bioinformatics Institute · Accessible scientific overview · Checked Nov 19, 2026
See the explanation of the CASP14 experiment, its comparison with other entrants and the limits of the targets. Read the source
56. Training Language Models to Follow Instructions with Human Feedback
Long Ouyang and colleagues, NeurIPS · Original conference paper · 2022 · Checked Nov 19, 2026
Read the labeler demonstration and preference-ranking procedure and the caveats on whose preferences were learned. Read the source
57. Training Compute-Optimal Large Language Models
Jordan Hoffmann and colleagues · Original scaling study · 2022 · Checked Nov 19, 2026
Compare Chinchilla and Gopher under the authors’ matched training-compute budget. Read the source
58. ReAct: Synergizing Reasoning and Acting in Language Models
Shunyu Yao and colleagues · Original preprint · 2022 · Checked Nov 19, 2026
Read the action-and-observation examples and the environments used in the experiments. Read the source
59. BLOOM: A 176B-Parameter Open-Access Multilingual Language Model
BigScience Workshop · Project research paper · 2022 · Checked Nov 19, 2026
Read the abstract and the sections on multilingual data, collaboration and released artifacts. Read the source