Large language models

What this subject is

A large language model is a neural network trained to predict the next token in a sequence, fitted by gradient descent over a very large corpus of text, and then adapted by supervised instruction tuning and preference optimisation into something that answers rather than merely continues. Every property people find remarkable and every property people find alarming descends from that one objective operating at scale: fluency, breadth, in-context learning, confident error, sensitivity to phrasing, absorbed bias, and memorisation of the corpus. Shannon posed next-symbol prediction as a measurable task in 1951; Bengio and colleagues gave it distributed representations in 2003; Vaswani and colleagues gave it a parallelisable architecture in 2017; Kaplan and colleagues showed in 2020 that loss falls predictably with scale, and Hoffmann and colleagues corrected in 2022 how that budget should be spent. Christiano and colleagues supplied the method in 2017 by which human preferences steer a policy, and Ouyang and colleagues applied it in 2022 to make base models usable. Everything else in this map is either a consequence of that chain, a measurement of it, or a dispute about what the measurements mean.

What the package holds

Curated scaffolding your agent loads before it researches, so it starts from vetted ground rather than a cold search.

• Tier 1: 56
• Tier 2: 6
• Tier 3: 8
• Tier 4: 2

The questions and claims below are quoted from the package files.

Where the field disagrees

Each one carries real proponents on more than one side, so your agent cannot quietly pick a winner.

• Do large language models acquire genuinely new abilities at scale thresholds, or does metric choice manufacture the appearance of discontinuity?

  4 named positions · unresolved and worth teaching as unresolved. There is broad agreement that pretraining loss falls smoothly, that some downstream metrics jump, and that metric choice affects the shape. There is no agreement on whether anything qualitative happens inside the model, and the dispute is partly definitional. Any teaching that declares a winner here is ahead of the evidence.

• Does a system trained only on the form of language have anything that deserves to be called understanding?

  5 named positions · philosophically unresolved and probably not resolvable in its current framing. The productive move for a learner is to refuse the umbrella term: ask which specific competence, measured by which instrument, and how performance degrades off distribution.

• Will continued scaling of the current recipe keep delivering, or is a different architecture required?

  4 named positions · live, commercially loaded, and heavily distorted by incentives on both sides. Predictions in this controversy have a poor track record in both directions; date every claim and check it against measurements rather than announcements.

• Is hallucination an engineering defect that better models and methods will remove, or a consequence of the objective that can only be bounded?

  4 named positions · the theoretical results establish a floor above zero under stated assumptions; the engineering results establish that the current rate is well above that floor. Both can be true, and honest teaching gives the mitigation methods with their measured effect sizes rather than a verdict.

• Does learning from human feedback align a model with human intent, or teach it to produce outputs humans approve of?

  4 named positions · no serious party claims preference optimisation delivers truthfulness; the dispute is over how far the proxy gap can be narrowed and by what. Teach it as an optimisation of predicted approval with known, measurable side effects.

Myths the package corrects

Widely held claims with the evidence that settles or bounds them.

• “Model quality is set by parameter count — a bigger model is a better model, and the parameter number is the headline figure worth comparing.”

  debunked-as-stated

  hoffmann-2022 trained over 400 models to fit the compute-optimal frontier and found that parameters and training tokens should scale roughly in proportion — meaning the flagship models of that era were substantially undertrained. A 70-billion-parameter model trained on about 1.4 trillion tokens outperformed models several times its size. Practice then moved past the compute-optimal point in the other direction, training smaller models on far more tokens, because training compute is paid once while inference compute is paid on every request. Distillation (hinton-2015) compounds the break: a small model trained on a large one's output distribution inherits much of its behaviour, so parameter count does not even rank models within one family. ouyang-2022 removes what remains of the ranking — human raters preferred a 1.3-billion-parameter instruction-tuned model to a 175-billion-parameter base model. What survives is narrower and still true: holding data, method and post-training fixed, more parameters helps. That condition is essentially never satisfied in a comparison anyone publishes. As a coda on how the field self-corrects, besiroglu-2024 found errors in one of Chinchilla's three estimation approaches while supporting its headline conclusion.

• “Large models undergo qualitative phase changes — abilities that simply are not there below a scale threshold appear above it — so something genuinely new comes into existence as models grow.”

  contested-correction

  Present this as a live dispute, because it is one. schaeffer-2023 showed that the sharp curves concentrate under discontinuous scoring rules — exact-string match, multiple-choice accuracy — and largely smooth out when the same model outputs are rescored with continuous metrics such as token edit distance or Brier score; the authors then induced apparent emergence in vision models by choosing the metric, and confirmed the pattern across BIG-bench tasks. The paper received a NeurIPS 2023 outstanding-paper award. A third position (Lu, Bigoulaeva, Sachdeva, Tayyar Madabushi and Gurevych, ACL 2024) attributes much of the apparently emergent performance to in-context learning and instruction tuning rather than to any scale-triggered latent ability. Against all of this stands olsson-2022, which found induction-head formation to be a genuine phase change visible in the training loss — so discrete internal transitions demonstrably occur, whatever explains any particular benchmark jump. What everyone agrees on: pretraining loss falls smoothly, some downstream metrics jump, and metric choice affects the shape. What nobody has established: that a benchmark discontinuity indicates a qualitative change inside the model. Teaching either side as settled is ahead of the evidence.

• “Hallucination is a defect of current systems; the next larger model, or the next training trick, will fix it.”

  debunked-as-stated

  Hallucination follows from how the objective works. Training rewards likely continuations of a finite, imbalanced corpus; nothing anywhere in the objective rewards abstention, and the model has no internal marker separating a fact recalled from a plausible completion. kalai-vempala-2024 proves the consequence: a model calibrated on its training distribution generates false statements at a rate tied to the fraction of facts appearing exactly once in training, so reducing hallucination requires either abstaining or becoming less calibrated, each with a cost. xu-2024 argues a stronger inevitability result and is contested precisely on its scope — it concerns worst-case agreement with arbitrary computable functions rather than practical accuracy on realistic queries — and the honest reading treats it as a boundary argument, not as a licence to stop improving. Scale alone is not even monotonic: lin-2022 found larger models less truthful than smaller ones on questions engineered to invite common human misconceptions, because scale amplifies whatever the corpus contains. Part of the observed rate is also an artefact of how the field scores itself — benchmarks that award credit for a confident wrong answer and nothing for "I do not know" train guessing in. What works is measurable and partial: retrieval grounding, abstention training, self-consistency, verification against tools, and citation checking, with effect sizes catalogued in ji-2023. Mitigation is real; elimination is not on offer.

• “The model has a database of facts inside it and looks up the answer when asked, or searches the internet in the background.”

  debunked

  A base model performs one forward pass over its weights; there is no index, no record of where anything came from, no retrieval step, and no internet access unless a tool has explicitly been wired in as a separate component (lewis-2020 describes what that component looks like when it exists). What the weights hold is real but is not a database. petroni-2019 measured it directly: pretrained models recall a non-trivial share of relational facts with no retrieval component at all, and that recall is uneven, sensitive to how the question is phrased, and unreliable for rare entities. kandpal-2023 quantified the shape — accuracy tracks how often the relevant entity relation appeared in the pretraining corpus, strongly enough that closing the long-tail gap by scale alone would require implausibly large models. carlini-2021 fixes the other boundary: verbatim training sequences, including personal data, can be extracted by querying, so the correct picture is neither "just patterns, no content" nor "a stored encyclopaedia" but a lossy, frequency-weighted compression with no index and no provenance. The practical consequences follow immediately: recall degrades on rare topics, the model cannot tell you its source, and it cannot distinguish internally between remembering and confabulating.

• “The model scores in the top few percent on professional and academic exams, so it will perform at that level on real tasks of that kind.”

  debunked-as-stated

  Three independent problems break the inference and they compound. First, contamination: test items appear in web-scale corpora whose contents are undisclosed, so for closed models the question is unfalsifiable in principle (sainz-2023); magar-schwartz-2022 shows memorisation and exploitation are separable and that whether contamination inflates a score depends on model size, duplication and task, which makes this an empirical matter rather than an accusation. Second, construct validity: raji-2021 documents how benchmarks built for narrow situated tasks get promoted into general capability claims the instruments were never designed to support, and chollet-2019 argues that measuring skill rather than skill-acquisition efficiency guarantees the instrument can be bought with data. Third, aggregation: liang-2023 evaluated dozens of models across dozens of scenarios and found rankings differ by desideratum — accuracy, calibration, robustness, fairness, bias, toxicity, efficiency — so any single headline number is a decision about what to ignore. A fourth, mundane problem sits underneath: the evaluated artefact is usually not the deployed one, which arrives quantised or distilled behind a system prompt and a serving stack. Note the symmetric error, which this package treats as equally serious: none of this establishes that benchmarks measure nothing. bowman-2022 argues that systematic underclaiming misleads policy and safety work as reliably as hype does. The correct posture is neither trust nor dismissal but interrogation — what was measured, with what prompting protocol and compute budget, was contamination checked, and is the construct the thing you care about.

Related topics

• Cryptography
• Systems thinking
• Economic psychology