Oluwadi
Investigating the Overlooked
Region
What
Who For
Analysis
Oluwadi
Investigating the Overlooked
A computer executes a structured language exactly. A compiler, a database engine and a theorem prover each accept a formal syntax with defined semantics, and each returns the same result for the same input. Natural language has no such definition. The work of getting machines to use human knowledge is the work of translating from one to the other, and the history of it is a history of who did the translating and what it cost.
The arithmetic in the second half of this piece is exact, and every step of it is deterministic. The same input and the same weights, the numbers that training sets, return the same output. That output is a probability distribution over the next word: a list that gives every possible word a probability not below zero, with the probabilities summing to one. [1] [2] The probabilities come from the language, and from compounding.
English lets you say the same thing many ways. "The meeting ended early," "The meeting wrapped up ahead of schedule" and "We got out of the meeting early" mean about the same. So when a model has written "The meeting," there is no single right next word. There is a set of acceptable ones: ended, finished, wrapped, broke up. The model's job is to put a probability on each word in that set, and on every other word in the language, nearly all of them close to zero.
Words also run the other way: one word can mean many things. WordNet, a dictionary built at Princeton that groups words by meaning, lists 78,019 single-word entries in version 3.1, and 23,564 of them, 30 percent, have more than one sense. The common words have the most. "Break" has 75: break a window, break a promise, break for lunch, break a record, break even, a lucky break. The same word in the same spot points at different things depending on what came before it. A model that cannot tell which sense is meant will pick the wrong next word.
Irregular verbs show the same unevenness. Most English verbs follow a rule: walk becomes walked. A few do not: go becomes went. Fewer than 3 percent of modern English verbs are irregular, and yet the ten most common verbs are all irregulara: be, have, do, go, say, can, will, see, take, get. The words people use most are the ones that break the rule. Researchers who followed English verbs over time found that the half-life of an irregular verb scales as the square root of its frequency. [3] That means a verb used a hundred times as often lasts about ten times as long, and the rare irregular verbs are the ones that give way and become regular. What is left as exceptions is the everyday core.
Nouns do the same. The plural of person is people, which WordNet lists as an entry of its own, with four senses against three for person. WordNet also keeps exception lists for the forms its suffix rules cannot derive: 2,054 for nouns and 2,401 for verbs. Each entry is a case where "add an -s" or "add an -ed" gives the wrong answer, and someone had to write the exception down.
This unevenness is part of why English is hard to learn as a second language. A learner can master a rule and then meet the words that break it, and the words that break it are the common ones. In my view the cause is how English was assembled. It has a Germanic base, with French and Latin layered on top and traces of other languages, Celtic among them. English is a mutt, and each parent left its own spellings and its own exceptions. American English is the stepchild of the mutt. It carried the same mix across the Atlantic and took in words from the people it met there. In New Jersey, Absecon, Rahway and Lake Lenape are ordinary names, and all three are Native American in origin. Absecon comes from the Lenape word Absegami, usually glossed as "little water." [55] Tradition says Rahway is named for Rahwack, a local tribal chief. [56] Lake Lenape, in Atlantic County, carries the name of the Lenape people themselves. [57] Three names are an anecdote. The scale is the argument. Wikipedia's list of New Jersey placenames of Native American origin, which is marked as incomplete, had 206 entries when I counted it on October 17, 2026: one county, 84 settlements and 121 bodies of water. [58] Most New Jerseyans say these names every day without hearing another language in them. A high school in Galloway Township carries the name Absegami, and the name is passed along each year to a new class. [59] We keep the lineage even when we do not know the origin. This is my observation. I have not sourced the history of the language, and a linguist may want to qualify it.
Now the compounding. A model writes one word at a time, and each word it chooses becomes part of the context for the next. The probability of a whole sequence is the product of the conditional probability of each word, where the conditional probability of a word is its probability given the words before it. [1] A product behaves differently from a sum. Small shortfalls multiply.
Here is an illustration, not a measurement. Suppose that at each step there is a 99 percent chance that the model's choice is acceptable. Over 10 words, the chance that every choice is acceptable is about 90 percent. Over 100 words it is about 37 percent. Over 1,000 words it is about one in twenty-three thousand. The arithmetic treats each choice as independent, which real text is not, so it shows the shape of the effect and not its size. The shape is the point: a per-word reliability that sounds excellent decays quickly over a paragraph.
There is a second effect, and it is worse than the arithmetic. A mistake early in an answer is not diluted by the words after it. It becomes part of what those words are conditioned on, so the later words are chosen to fit it. That is how a fluent, confident paragraph can rest on a wrong first step: everything after the error is consistent with the error. This is my description of the mechanism, not a measured result in the sources.
Both facts point the same way. Because many words are acceptable at each step, the model cannot be checked word by word against a single right answer. Because small differences compound, the answer as a whole can drift even when each word looks fine. The unit that has to be checked is the whole answer.
What a system does with the distribution is a separate choice. Taking the most likely word every time is deterministic. Sampling from the distribution, so that likelier words come up more often but unlikely ones sometimes do, adds randomness on purpose. Neither is a proof.
Randomness also enters in a second place, earlier: when the model is trained. A neural network is a large set of adjustable numbers called weights. Before training, nobody knows the right values, so the weights start as random numbers. AlexNet, the 2012 image network described later in this piece, drew its starting weights from a zero-mean Gaussian distribution. [4]b That is a bell curve. Most draws land near the middle and a few land far out. Zero-mean means the middle sits at zero, so the starting weights are small and just as likely to be positive as negative.
Training then adjusts them, a little at a time. The network is shown an example and gives an answer. The answer is compared with the right one, and the size of the error is measured. Each weight is then nudged slightly in whichever direction would have made the error smaller. That direction is the gradientc. Doing this on the whole data set for every nudge would be slow, so stochastic gradient descent estimates the direction from a small random sample each time. [5] [6] AlexNet used samples of 128 examples.d [7]
Put the two together. The start is random, and every step is a small correction. The result is wherever the corrections arrive. A different random start can arrive at different weights.
The evidence carries the same uncertainty. Error rates, BLEU scores and perplexity are each measured on a sample of language, so each is an estimate and not an exact value. Error rate is the share of test items answered wrongly. BLEU is the overlap in word sequences between a machine translation and human translations. [8] Perplexity is computed from the inverse of the probability a language model gives a test text, and lower is better. [9]
Computer scientists sort formal languages into levels, and each level has a known kind of machine that can recognize it. The simplest level is the regular languages. A finite-state machine recognizes them. It has a fixed set of states, like a turnstile that is either locked or unlocked, and it remembers nothing beyond which state it is in now. The next level is the context-free languages, which underlie the syntax of most programming languages. A pushdown automaton recognizes them. It is a finite-state machine with a stack, which lets it keep track of nesting, such as which parentheses are still open.e [10]
Noam Chomsky, an American linguist at MIT, showed in 1956 that no finite-state process can serve as a grammar for English. [11] A grammar here is a finite set of rules that generates the sentences of a language. In plain terms, a machine that knows only its current state cannot keep track of how English sentences are built. A state machine can follow a structured language and cannot, by itself, follow a natural one.
The oldest account of the job is in Genesis, where the animals are brought to the man "to see what he would call them." [12] Naming is the first half. The second half is deciding where a named thing goes.
Three terms cover that second half. Classification is the act of assigning an item to a category, as a librarian puts a book on a shelf. A taxonomy is the scheme of categories, usually a hierarchy, with broad groups containing narrower ones. An ontology goes further. It declares relations beyond "is a kind of," such as "is part of," under a shared vocabulary. Tom Gruber, a computer scientist, defined it in 1993 as "a specification of a representational vocabulary for a shared domain of discourse," made of "definitions of classes, relations, functions, and other objects." [13]
Librarians did this by hand before there were computers. S. R. Ranganathan, an Indian librarian and mathematician, first published his Colon Classification in 1933. It described each work along independent facets, which he later settled at five: Personality, Matter, Energy, Space and Time. [14] A tree gives an item one position, like a book that sits on exactly one shelf. A faceted scheme gives it a value on each axis, so the same book can be described by its subject, its place and its period at once. That is a different data model, and it is the one that a relational schema resembles.f
The first machines to hold such knowledge were programmed with it by hand. Edward Feigenbaum, a Stanford computer scientist, described the field in 1977 as bringing the tools of AI to problems "requiring experts' knowledge for their solution." In the same paper he named the cost: "the acquisition of domain knowledge was the bottleneck problem in the building of applications-oriented intelligent agents." [15]
A knowledge engineer sat with an expert and wrote down what the expert said as rules in a structured language. The author knew people who did this work, and did the equivalent: sitting with analysts to work out how they worked.
What the author took from both is that the expert cannot always say what they know. Expertise is closer to innateness than to recall. The expert no more thinks about it than a person thinks about breathing, and by the time someone is an expert it is just normal. The author sees relationships between things and cannot say why. Two psychologists, Nisbett and Wilson, reviewed evidence in 1977 that people may have little or no direct access to their own higher-order thinking, and that their reports of it rest on implicit theories of cause. [16] Later reviewers call the legacy mixed: widely cited, with no research program continuing it. [17]
The consequence was cost. The rules did not stay finished, and there were never enough experts. In the work the author knew of, that cost stopped the project: the translation from natural language to formal rules needed a scarce trained person for every fact.
Conversational programs were built the same way, as state machines and frames. ELIZA matched patternsg in the user's text and returned scripted responses. I built an ELIZA clone as a teenager, so I know how little there is inside one. Task systems such as GUS, a travel-planning system from 1977, treated a request as a form with named blanks, called slots, such as destination and date. The system's job was to fill the blanks and then act. [18] Both put a structured language between the person and the machine, and the person had to stay inside it.
Cyc is the cost at its largest. Douglas Lenat began it in 1984 as an attempt to write down common sense by hand. By his own 1995 account, about a person-century of effort had produced roughly 100,000 concepts and about a million commonsense axioms, entered by hand.h [19]
Two standards grew up around this work. RDF, the W3C's data model, writes every assertion as a triple of subject, predicate and objecti, such as a person, "wrote," and a book. A set of triples is a graph, because each triple links two things. [20] OWL, standardized in 2004, adds vocabulary such as disjointness and cardinalityj to RDF and RDF Schema. [21]
WordNet, built at Princeton, is a smaller and more useful artifact. It groups English words into synsets, sets of synonyms that share one meaningk, and links them with relations such as hypernymy (is-a) and meronymy (part-of). The current release holds more than 117,000 synsets, about 82,000 of them nouns.
In 2009 a Princeton team used the noun hierarchy as the label set for an image collection, ImageNet. The paper says it was "built upon the backbone of the WordNet structure." [22] Each category of picture was a WordNet noun. The collection had 3.2 million images in 5,247 synsets, labeled by crowd workers through Amazon Mechanical Turk.l [22] By 2014 it held 14,197,122 images in 21,841 synsets, and a classification contest on a 1,000-category subset had run since 2010. [23]
In 2012 a deep convolutional networkm, AlexNet, trained on that subset by supervised learning, which means on labeled examples, reached a top-5 errorn of 15.3 percent against 26.2 percent for the next entry. [7] In other words, the right label was among its five best guesses far more often than for any rival. Labeled data and a taxonomy had carried the field across the line.
The taxonomy carried its flaws with it. A 2020 audit of ImageNet's person subtree found 2,832 synsets under it. Of those, 1,593 were flagged as potentially offensive labels that should not be used for image recognition, and they were attached to 600,040 images. Of the remaining 1,239, only 158 were visual categories. The rest, the authors write, reflect annotators' bias. [24] The authors trace part of the problem to "the stagnant concept vocabulary of WordNet." [24] A classifier can be accurate and still be wrong if the scheme it classifies into was wrong first.
A second line of work started from the other end. It did not write language down in a structured form. It treated language as a probability distribution to be estimated from examples. The standard tool was the n-gram model, which predicts the next word from the few words before it: after "peanut butter and," it guesses "jelly." Yoshua Bengio and colleagues named its limit in 2003. A test sequence will usually not match any sequence seen in training, so n-grams generalize only by stitching together short overlapping chunks. [1]
Their answer was to learn a vector for each wordo. A vector is a list of numbers, which places the word at a point in space, and words used in similar ways land near each other. A sentence never seen in training could then still receive high probability if its words sat near words from sentences the model had seen. [1]
The method for fitting such models was already there. Rumelhart, Hinton and Williams had published backpropagation in 1986, a procedure that repeatedly adjusts a network's connection weights to reduce the difference between its output and the desired output. [25] It is how the nudge described above is worked out for every weight in every layer.
Irregular verbs gave the method its first public test. In 1986 Rumelhart and McClelland built a neural model that learned to map verb stems to past tense forms, and in 1988 two other researchers, Pinker and Prince, published a critique of it. [26] In 2018 Kirov and Cotterell reran the task with encoder-decoder networks and argued that these address most of Pinker and Prince's objections and warrant re-examination as models in linguistics. [26] The dispute is the one this piece follows: rules written down, or structure learned from examples.
What was missing was hardware. A CPU executes a few operations at a time, very fast. A GPU, a graphics processor, performs the same operation on many numbers at once. In 2009 a Stanford team reported that a GPU implementation of deep belief network trainingp ran up to 55 times faster than a dual-core CPU, and took a four-layer model with 100 million parameters from several weeks to about a day. [27]
In 2012 AlexNet took five to six days to train on two GTX 580 GPUs with 3 GB of memory each, and its authors wrote that the results "can be improved simply by waiting for faster GPUs and bigger datasets to become available." [7] The mathematics preceded the hardware by more than two decades. Data was the other constraint, and ImageNet arrived in 2009. Which of the two limited the field more is not settled.
Language needs a model that carries information across a sequence. The recurrent neural network does thatq. At each step it feeds its hidden layer back in as input at the next stepr, so the network's state holds a trace of what it has already read. Jordan described the idea in 1986, and Elman developed it in 1990, applying it to tasks that ran from a temporal version of XOR to discovering syntactic and semantic features of words. [28] At scale it arrived in 2010: a mixture of recurrent language models roughly halved perplexity against a strong n-gram baseline, and cut word error rate by about 18 percent on a Wall Street Journal speech recognition task.s [29]
Recurrence has a weakness, and the 1997 long short-term memory (LSTM) paper names it. During training, the error signal flows backward through the steps, and across many steps it tends either to blow up or to vanisht. A signal that vanishes teaches the early steps nothing, so long-range dependencies are not learned. LSTM adds gated unitsu that keep the error flowing across more than 1,000 time steps. [30]
The results followed. A deep LSTM network reached 17.7 percent error on the TIMIT phoneme recognition benchmarkv in 2013. [31] By 2014 an LSTM encoder-decoderw reached a BLEU score of 34.8 on English-to-French translation, against 33.3 for a phrase-based system. [32] It compressed each source sentence into one fixed-length vector, which Bahdanau, Cho and Bengio called "a bottleneck." They replaced it with a soft search over the source sentence, which is attention.x [33]
After 2012 the direction reversed. The hardware began to choose the mathematics. A recurrent network computes each step from the state of the one before it, so the steps of a single sequence cannot run at the same time. A GPU is built to run many operations at once, and recurrence cannot use that. The Transformer dropped recurrence for attention alone, and its authors reported a model that was "more parallelizable and requiring significantly less time to train," trained in 3.5 days on eight GPUs. [34]
In the same year Google reported a chip built around one operation, matrix multiplication, which combines two grids of numbers by multiplying and adding. The chip had a matrix multiply unit of 65,536 8-bit multiply-accumulate cellsy and a peak of 92 trillion operations per second. It was 15 to 30 times faster than its contemporary CPU or GPU on neural network inference, which means running a trained model on new input. [35]
The techniques map to tasks as follows. Read each row from left to right: the technique, what it was built for, and one finding reported in its source.
| Technique | Task it was built for | What the source reports |
|---|---|---|
| Hand-written rules | Problems "requiring experts' knowledge for their solution" [15] | Knowledge acquisition was the bottleneck [15] |
| Taxonomy, ontology, RDF, OWL | Fixing the vocabulary and relations of a domain [13] [21] | ImageNet's label set was built on WordNet's [22] |
| Hidden Markov modelz | Speech recognition [36] | Three standard problems: sequence likelihood, best hidden path, parameter fitting [36] |
| n-gram model | Next word from the few before it [1] | Generalizes only by stitching short chunks seen in training [1] |
| IBM alignment modelsaa | Statistical machine translation [37] | Five models, parameters estimated from sentence pairs [37] |
| Recurrent network | Sequences, as a language model for speech recognition [28] [29] | About half the perplexity; word error down 18 percent on Wall Street Journal [29] |
| LSTM | Long-range dependencies; phoneme recognition [30] [31] | 17.7 percent error on TIMIT [31] |
| LSTM encoder-decoder | Translation [32] | BLEU 34.8 against 33.3 for a phrase-based system [32] |
| Convolutional network | Image classification [7] | 15.3 percent top-5 error against 26.2 [7] |
| Attention, Transformer | Translation and alignment [33] [34] | 28.4 BLEU English-to-German; 3.5 days on eight GPUs [34] |
| Distillation | Compressing a large model into a small one [38] | Within 0.3 points of a ten-model ensemble [38] |
| RLAIF | Replacing human preference rankings [39] [40] | Harmless rate 88 percent against 76 for human feedback [40] |
| Synthetic generation | Supplying training data [41] [42] | Contested: gains in two studies; collapse when it replaces real data, not when it accumulates [43] [44] |
Some of these competed for the same task. N-gram and recurrent language models both served speech recognition, and the recurrent model won. Others belong to different tasks. What changed between them was the hardware and the data available to train them.
Language models removed the scheme. Next-token prediction is self-supervised learning: the training target is the next token, a word or a piece of a word, of the text itself, so the data supplies its own labels. Nobody writes a grammar, a taxonomy or an ontology for it. The structure is learned from co-occurrence in a very large record of what people wrote. What caused the results that followed is contested. Kaplan and colleagues found that loss, the measure of how wrong a model's predictions are, falls as a power law in model size, dataset size and compute, the total arithmetic spent on training, so that each tenfold increase in one changes the loss by a fixed factor, and that architectural details such as depth and width have minimal effect within a wide range. [45] Architecture is not irrelevant, since the Transformer paper reports higher translation quality as well as speed. [34] The author's position follows Kaplan: the supply of text and the arithmetic to use it were the conditions, and the architectures that mattered were the ones that used the arithmetic well.
That record is finite. Epoch AI estimates the effective stock of public human textab at about 400 trillion tokens, and projects that models will be trained on datasets roughly equal to it between 2026 and 2032, with a median of 2028. [46] That is a forecast, and it rests on assumptions about data quality and repeated passes.
The response is a belief, not a result: that models will have to be trained on data other models produced. It is worth separating three things that get called by that name.
Knowledge distillation trains a smaller student model to match a larger teacher's output distribution. Hinton, Vinyals and Dean reported on a speech recognition task that a distilled model came within 0.3 points of an ensemble of ten, ten models whose outputs are combined. [38]
In reinforcement learning from human feedback, RLHF, people rank a model's outputs and the model is trained to score well against a reward learned from those rankings. [47] Reinforcement learning from AI feedback, RLAIF, replaces the human rankings with rankings produced by another model. In Anthropic's Constitutional AI,ac "the only human oversight is provided through a list of rules or principles." [39] A Google study found RLAIF comparable to RLHF, with a harmless-response rate of 88 percent against 76 percent. [40]
Synthetic data generation, the third, has contested evidence. Self-Instruct started from 175 hand-written tasks, had a model produce about 52,000 instructionsad with 82,000 instances, and improved the base model by 33 percent, reaching within about 5 percent of a model trained with private human annotation. [41] The phi-1 code model, trained on 6 billion tokens of filtered web text and 1 billion tokens of GPT-3.5-generated textbooks, scored 50.6 percent on HumanEval pass@1.ae [42] A Nature paper found that training on model output "indiscriminate[ly]" caused model collapse: irreversible defects, with the tails of the original distribution disappearing. [43] A later study found that collapse follows when each generation's synthetic data replaces the real data, and does not follow when synthetic data accumulates alongside it. [44] Whether frontier training can run on synthetic data is unresolved. The claim here is only that the belief exists, and the author has seen at least six companies selling coherent synthetic data on it.
The human work did not disappear. It moved. In 2023 TIME reported that workers in Kenya labeled violent and abusive text so a safety filter could be trained, for take-home pay of about $1.32 to $2 an hour against a contract rate of $12.50. [48] In October 2025 a company supplying expert reviewers to AI labs raised $350 million at a $10 billion valuation, listing more than 30,000 doctors, lawyers and scientists paid over $85 an hour. [49]
Follow where the structure sits at each stage. The knowledge engineer wrote it in a formal language. The ontologists and the ImageNet team wrote it as a taxonomy of categories. The language model learned it from text. Constitutional AI writes it as a list of plain-English principles and lets a model interpret them.
That last step is a reversal. The specification went back from a structured language to natural language, and the interpreter is a model whose behavior has not been proven. A formal specification can be checked by a machine. A model's answers come from a distribution over acceptable sentences: they can be sampled and measured and cannot be proven, and a list of principles can only be checked by whoever reads the output. That person is the one who checks the answer. Whether a model's adherence to written principles can be verified at all is unresolved.
I know formal languages in the mathematical sense, and I work to make the machine I use deterministic, even though I am not. My hormones could be off. The person who checks the answer is a body too, and some days that body is not at its best. That is one more reason the answer has to be checked against the record and not against the checker's mood.
Counts of WordNet synsets, senses and exception-list entries are the author's count from the Princeton WordNet 3.1 database files. The account of knowledge engineering is the author's, from people who did the work and from the author's own time sitting with analysts, and not from watching it done. The account of companies selling synthetic data is the author's observation, and no company is named. The closing paragraph about the author's own non-determinism is his account. The ELIZA clone is the author's own history, and the author's training includes formal languages in the mathematical sense. The paragraph on why English is hard to learn is the author's observation, not a sourced finding. The 99 percent arithmetic is an illustration, not a measurement. The position that the supply of text and arithmetic were the conditions of scale is the author's, and the text marks where the record is contested. How I work with the AI assistant that helps me write, and why the read stays mine, is set out in The Harness Is the Mechanism. The Read Is Still Ours..
About the cartoon. In September 1947, operators of the Mark II computer at Harvard found a moth in a relay, taped it into their logbook and labeled it "first actual case of bug being found." The logbook is held by the Smithsonian's National Museum of American History. [54] Engineers had used "bug" for a fault in a machine before that, including Thomas Edison in the 1870s, so the moth is the literal case and not the origin of the word. [54] The drawing is a general scene from that lab, not a portrait of anyone.
a. Irregular verb: a verb whose past tense is not formed with the suffix -ed, as with go and went. A regular verb is one that is, as with walk and walked. [3]
b. Gaussian (normal) distribution: a bell-shaped distribution defined by a mean, its center, and a standard deviation, its spread. Zero-mean means centered on zero. [4]
c. Gradient: the direction and rate at which the error changes as each weight changes. Stochastic gradient descent repeatedly moves the weights against the gradient, estimating it from a small random sample of the training data at each step. [5] [6]
d. Batch: the group of training examples used for one weight update. [5]
e. Pushdown automaton: a finite-state machine with a stack, which lets it track nested structure such as matching parentheses. Regular language: a language a finite-state machine can recognize. Context-free language: a language generated by a grammar whose rules each rewrite one symbol regardless of its surroundings. A grammar is a finite set of rules that generates the strings of a language. [10]
f. Facet: an independent axis along which an item is described, such as place or time. A tree gives each item one parent. A faceted scheme gives each item a value on every axis. Relational schema: the definition of the tables and columns of a relational database, in which each row holds a value for each column.
g. ELIZA: an early conversational program that matched patterns in the user's text and returned scripted responses, built to imitate a Rogerian psychologist. [18]
h. Axiom: a statement accepted as true and used to derive other statements. Cyc's are commonsense statements. [19]
i. IRI (Internationalized Resource Identifier): a globally unique identifier, written like a web address, that names a thing in RDF. A triple's subject and predicate are IRIs, and its object may be an IRI or a literal value. [20]
j. Disjointness: two classes share no members, as in nothing being both a cat and a dog. Cardinality: a limit on how many values a property may have, as in a person having exactly one birth date. RDF Schema: a vocabulary for declaring classes, properties and subclass relations on top of RDF. [21]
k. Synset: a set of words that are synonyms in one sense, treated as a single concept. The ImageNet paper calls it a "synonym set." [22] Hypernymy: the relation from a specific term to a more general one, as from "dog" to "animal." Meronymy: the relation from a part to its whole, as from "wheel" to "car."
l. Amazon Mechanical Turk: a marketplace where people are paid per small task, here the labeling of images. [22]
m. Convolutional network: a neural network whose layers apply the same small filter across the whole input, so that a pattern is detected wherever it appears. Deep: having many layers between input and output. [7]
n. Top-5 error: the fraction of test images for which the correct category is not among the model's five highest-ranked guesses. [23]
o. Vector: an ordered list of numbers. An embedding is the vector a model produces for an item, such as a word or an image, so that items that are alike land near each other. [1] [50] Nearness is measured with a distance or similarity function, and finding the stored vectors nearest to a query is nearest-neighbor search. [51]
p. Deep belief network: a neural network built by stacking layers that are trained one at a time. Parameters: the learned numbers of a model, the same thing as its weights. [27]
q. Neural network: a function built from layers of simple units, each of which computes a weighted sum of its inputs and passes the result on.
r. Hidden layer: a layer between input and output whose values are not observed directly. In a recurrent network its values at one time step are fed back as input at the next. Time step: one position in a sequence. [28]
s. Word error rate: the share of words in a transcription that are wrong, counting substitutions, deletions and insertions against a reference transcript.
t. Error signal: the quantity passed backward during training that tells each weight how to change. When it shrinks toward zero, early steps learn nothing. When it grows without bound, training breaks. [30]
u. Gate: a unit whose output, between zero and one, multiplies another signal, so that the network can decide how much to pass on or hold. [30]
v. Phoneme: the smallest unit of sound that distinguishes words in a language. TIMIT: a corpus of recorded read speech with hand-verified phonetic transcriptions, used to test speech recognizers. [52]
w. Encoder-decoder: a model in two parts. The encoder reads the input sequence into an internal representation, and the decoder generates the output sequence from it. [32]
x. Attention: lets the decoder weight different parts of the source sentence as it produces each output word. [33] Transformer: a network built from attention alone, without recurrence. [34]
y. Multiply-accumulate (MAC): multiply two numbers and add the result to a running total. Matrix multiplication is made of MACs. TOPS: trillions of operations per second. [35]
z. Hidden Markov model: a statistical model of a sequence in which each observed item depends on a hidden state that changes step by step, and the next state depends only on the current one. [36]
aa. Alignment: a mapping from words in a source sentence to words in its translation. The IBM models assign probabilities to alignments and estimate their parameters from sentence pairs. [37] Phrase-based system: a statistical translation system that translates sequences of words and not single words.
ab. Effective stock: Epoch's estimate of the amount of text usable for training after quality filtering. [46]
ac. Constitutional AI: Anthropic's method of training with AI feedback guided by a written list of principles. [39]
ad. Instruction: a task described to a model in natural language. Instance: one input-output example of a task. [41]
ae. HumanEval: a set of programming problems scored by whether the code a model writes passes tests. pass@1: the fraction of problems solved with one attempt each. [53]