Science and technology — C2

By B2 you owned the everyday vocabulary of scientific topics and gadgets. At C1 you added research-method vocabulary and the working terms of major fields. At C2 you cross into the discourse where the philosophy of science, frontier AI safety research, and the molecular biology of mRNA vaccines are taken apart in detail. You can read a Nature editorial on a retraction, a Quanta Magazine feature on the AdS/CFT correspondence, an Anthropic paper on interpretability, a MIT Technology Review analysis of CRISPR clinical trials, and a New Yorker profile of an AI alignment researcher — without translation drag and without misreading the technical subtext.

The vocabulary in this lesson sits at the seam of philosophy of science, statistical methodology, AI/ML, quantum physics, and biotechnology. It is the working language of Nature, Science, Quanta, MIT Technology Review, Wired, The Atlantic’s science writers, and the long preprints of DeepMind, OpenAI, Anthropic, and academic ML labs. A C2 speaker reads these without help, including the methodological caveats most authors slip past readers who lack the vocabulary to notice.

A pragmatic note: scientific language is supposed to be precise but is often misused even by scientists. Theory in everyday speech means guess; in science it means a well-supported framework. Significant in everyday speech means important; in statistics it has a specific meaning related to p-values that decouples it from importance. At C2 you read these double meanings and avoid the confusions.

Philosophy of science — what the scientific method actually says

• the scientific method — the loose family of practices: observation, hypothesis, prediction, experiment, revision
• inductive reasoning vs deductive reasoning — generalizing from cases vs reasoning from premises
• abductive reasoning / inference to the best explanation — Peirce-Lipton; pick the explanation that best accounts for the evidence
• hypothesis — testable claim
• null hypothesis vs alternative hypothesis — H0 is the default (no effect); H1 is what you’re testing for
• theory (in science) — a well-supported integrative framework (the theory of evolution, general relativity); contrast everyday theory = guess
• law — a regularity expressed mathematically (Newton’s laws); not stronger than a theory, just narrower
• model — a simplified representation; all models are wrong, some are useful (George Box)
• paradigm (Kuhn) — the shared framework of a scientific community
• paradigm shift — fundamental change in framework (often misused for any change)
• normal science vs revolutionary science (Kuhn) — within-paradigm puzzle-solving vs paradigm replacement
• the demarcation problem — distinguishing science from non-science
• falsifiability / refutability (Popper) — a claim is scientific only if it could in principle be falsified
• risky predictions — predictions a theory makes that could fail; Popper’s mark of good science
• ad hoc rescue / ad hoc modification — patching a theory to avoid refutation (pejorative)
• unfalsifiable / unfalsified vs falsified — different statuses
• the underdetermination of theory by data (Duhem-Quine) — multiple theories can fit the same evidence
• the Duhem-Quine thesis — auxiliary assumptions always accompany hypotheses; what gets falsified is the whole bundle
• research programme (Lakatos) — hard core + protective belt of auxiliary hypotheses
• degenerating vs progressive research programme — losing ground vs predicting new phenomena
• methodological anarchism (Feyerabend) — anything goes; deliberately provocative
• scientific realism vs anti-realism / instrumentalism — theories aim at truth vs at useful prediction
• the no-miracles argument (for realism) — success would be a miracle if theories weren’t approximately true
• the pessimistic meta-induction (against realism) — most past theories were false; current ones probably are too
• reductionism — explanation in terms of lower-level components
• emergence / emergent properties — higher-level properties not obvious from lower-level
• strong emergence vs weak emergence — irreducible vs in-principle-reducible
• mechanism — a causal process; modern philosophy of biology and neuroscience prize mechanistic explanation

Methodology — the evidence hierarchy

• evidence hierarchy / evidence pyramid — ranking of study designs by inferential strength
• case report / case series — lowest tier; observational on few patients
• cross-sectional study — snapshot at a single time
• case-control study — retrospective; compares cases to controls
• cohort study — prospective or retrospective; follows groups over time
• observational study — does not assign treatment
• randomized controlled trial (RCT) — random assignment to treatment vs control; the workhorse of clinical research
• double-blind RCT — neither participants nor researchers know assignments
• placebo-controlled — control receives inert intervention
• active-controlled — control receives existing standard of care
• non-inferiority trial vs superiority trial — different hypothesis structures
• intention-to-treat (ITT) vs per-protocol analysis
• pre-registration — locking in hypotheses and analysis before data collection
• registered report — journal accepts based on method before results
• meta-analysis — quantitative synthesis of multiple studies
• systematic review — structured qualitative synthesis (often paired with meta-analysis)
• PRISMA — reporting guideline for systematic reviews
• Cochrane review — gold-standard evidence syntheses from the Cochrane Collaboration
• GRADE — framework for rating evidence quality and recommendation strength
• forest plot / funnel plot — visualization tools
• heterogeneity / I-squared — variation across studies in a meta-analysis
• publication bias / the file-drawer problem — null results going unpublished
• the replication crisis — see also the academia lesson; many published results fail replication
• p-hacking / data dredging / garden of forking paths — analytic flexibility producing false positives
• HARKing — Hypothesizing After the Results are Known
• multiple comparisons / multiple testing problem — running many tests inflates false-positive rate
• Bonferroni correction / false discovery rate (FDR) / Benjamini-Hochberg — corrections

Statistical literacy at C2

• statistical significance — narrowly: p-value below a threshold; broadly often misused
• p-value — probability of data this extreme or more, assuming the null hypothesis
• effect size — magnitude of the difference (often more important than significance)
• confidence interval (CI) — range of plausible values for an estimate
• credible interval — Bayesian analogue
• prior vs posterior vs likelihood — Bayesian vocabulary
• frequentist vs Bayesian statistics — long-run frequency vs degree of belief
• type I error vs type II error — false positive vs false negative
• power / statistical power — probability of detecting a true effect
• underpowered study — too few subjects to reliably detect effect of expected size
• selection bias / confounding / measurement error
• causal inference — identifying causation from observational data
• DAG (directed acyclic graph) — modern causal-inference tool (Pearl)
• instrumental variable / regression discontinuity / difference-in-differences / propensity-score matching — quasi-experimental methods
• counterfactual — what would have happened under different conditions
• Simpson’s paradox — aggregate trend reverses when stratified

AI and machine learning — the C2 vocabulary

The pace of AI vocabulary change is unusual. A C2 speaker should track terms current as of 2025-2026 and recognize what was hot in earlier waves.

Core concepts

• artificial intelligence (AI) — broad umbrella
• machine learning (ML) — systems that learn from data
• deep learning — ML using deep neural networks
• neural network / NN / deep neural network (DNN) — layered computational model
• artificial general intelligence (AGI) — broad, human-level intelligence; contested concept
• artificial superintelligence (ASI) — beyond human level
• narrow AI / task-specific AI — single-purpose
• frontier model — leading-edge large model
• foundation model — large model trained broadly, adapted for tasks
• large language model (LLM) — text-trained foundation models
• multimodal model — handles text, images, audio, video
• transformer — the architecture (Vaswani et al., 2017) behind modern LLMs
• attention mechanism / self-attention — how transformers route information
• pretraining vs fine-tuning — large general training vs task-specific adjustment
• RLHF (reinforcement learning from human feedback) — alignment via human preference data
• DPO (direct preference optimization) — newer RLHF alternative
• constitutional AI — Anthropic’s approach using AI-generated feedback against a principles list
• prompt engineering / prompt design — crafting model inputs
• system prompt vs user prompt
• context window — how much input the model can attend to
• chain-of-thought (CoT) — eliciting intermediate reasoning steps
• scaling laws — predictable relationships between model size, data, and performance (Kaplan, Chinchilla)
• emergent capabilities — abilities appearing at scale; the existence and measurement are contested
• the bitter lesson (Sutton) — generic compute-scaled methods beat clever handcrafted ones over time

Safety and alignment

• AI alignment / the alignment problem — making AI pursue intended goals
• outer alignment vs inner alignment — specifying the right objective vs the model internalizing it
• mesa-optimization / mesa-optimizer — a model that contains its own optimizer with possibly different goals
• deceptive alignment — model behaves aligned during training but defects later
• reward hacking / specification gaming — exploiting the objective rather than fulfilling it
• Goodhart’s law — when a measure becomes a target, it ceases to be a good measure
• scalable oversight — overseeing models more capable than the overseers (debate, market-style aggregation)
• interpretability / mechanistic interpretability — understanding what models are doing internally
• circuits / features / superposition — interpretability primitives
• steering vectors / activation steering — manipulating model behavior via internal interventions
• red-teaming — adversarial testing
• jailbreak / prompt injection — bypassing model safeguards
• hallucination — generating false content presented as factual
• sycophancy — model telling users what they want to hear
• catastrophic risk vs existential risk (x-risk) — large-scale vs civilization-ending
• doomer vs accelerationist (e/acc) — slowdown advocates vs speed advocates (informal)
• AI pause / the open letter — 2023 call for a six-month pause on frontier training
• responsible scaling policy (RSP) — commitments to capability-evaluated pauses

Compute and infrastructure

• GPU / TPU / accelerator — computing hardware for ML
• H100 / A100 / B200 — successive NVIDIA flagship GPUs
• training run / training compute / FLOPs (floating-point operations)
• training cluster / supercomputer
• inference — running a trained model
• inference cost / token cost
• inference compute — using more compute at inference for harder problems (the o1/o3 paradigm)
• test-time compute / inference scaling
• distillation — training a smaller model to mimic a larger one
• quantization — reducing numerical precision to shrink model size
• MoE (mixture of experts) — sparse architecture activating only some weights per input
• agentic / AI agent — model that takes multi-step actions, uses tools, plans
• tool use / function calling — model invoking external systems

Quantum physics and computing — vocabulary at C2

• quantum mechanics — the framework
• wave function / state vector / the ket
• superposition — being in multiple states simultaneously
• entanglement — correlations stronger than classical
• decoherence — loss of quantum behavior through environmental interaction
• measurement / observation / collapse — the contested interpretive vocabulary
• the measurement problem — what happens at observation
• the Copenhagen interpretation / many-worlds (Everett) / pilot-wave (Bohm) / objective collapse (GRW)
• Bell’s theorem / Bell inequality / Bell test — distinguishing quantum from local-hidden-variable theories
• EPR pair / Bell state — entangled-pair vocabulary
• the no-cloning theorem — quantum states can’t be perfectly copied

Quantum computing

• qubit (quantum bit) — quantum information unit
• classical bit vs qubit
• logical qubit vs physical qubit — error-corrected vs raw
• quantum gate / quantum circuit
• NISQ (Noisy Intermediate-Scale Quantum) — current-era quantum hardware
• fault-tolerant quantum computing — with error correction
• quantum error correction (QEC) / surface code / topological code
• quantum supremacy / quantum advantage — quantum beats classical on some task (Google 2019, controversial)
• quantum algorithm — Shor’s, Grover’s, HHL
• post-quantum cryptography (PQC) — classical cryptography secure against quantum attacks
• Shor’s algorithm — quantum factoring; threatens RSA
• Grover’s algorithm — quantum search; quadratic speedup
• decoherence time / T1 / T2 — coherence lifetimes

Biotechnology — CRISPR, mRNA, and the modern biotech vocabulary

• DNA / RNA / mRNA / tRNA / rRNA / microRNA / miRNA / siRNA
• the central dogma — DNA → RNA → protein (with revisions)
• transcription vs translation — DNA to RNA vs RNA to protein
• gene expression / regulation of gene expression
• the genome / the transcriptome / the proteome / the metabolome
• GWAS (genome-wide association study)
• the human genome project — completed 2003
• the pangenome — diverse human genomic reference

CRISPR

• CRISPR / CRISPR/Cas9 / CRISPR/Cas12 / CRISPR/Cas13 — gene-editing systems
• the guide RNA (gRNA) — directs Cas to target
• PAM (protospacer adjacent motif) — sequence required adjacent to target
• knockout vs knock-in — disabling vs inserting
• prime editing / base editing — newer CRISPR variants with finer precision
• HDR (homology-directed repair) vs NHEJ (non-homologous end joining) — repair pathways
• off-target effects — unintended edits
• somatic editing vs germline editing — body cells vs heritable
• embryo editing — He Jiankui’s 2018 controversial work
• Cas9 specificity / high-fidelity Cas9
• CRISPR therapeutics — Casgevy (sickle-cell), exa-cel, others

mRNA and modern vaccines

• mRNA vaccine — encodes antigen for the body to produce
• lipid nanoparticle (LNP) — delivery vehicle for mRNA
• modified nucleoside (pseudouridine) — Karikó-Weissman innovation that reduced immunogenicity
• spike protein / antigen presentation
• neutralizing antibodies vs non-neutralizing antibodies
• cellular immunity (T-cell) vs humoral immunity (B-cell, antibody)
• immunogenicity / reactogenicity — immune response vs side effects
• antibody-dependent enhancement (ADE) — feared but largely unrealized SARS-CoV-2 concern
• vaccine efficacy vs vaccine effectiveness — trial vs real-world
• breakthrough infection — infection despite vaccination
• variants of concern (VOC) / variants of interest (VOI)

AmE-specific vs international vocabulary

Collocations

• frame / formulate / pose / posit / advance a hypothesis
• test / probe / interrogate a claim
• rule out / rule in an explanation
• confound / disentangle / tease apart causes
• control for confounders
• adjust for / stratify by
• replicate / fail to replicate / partially replicate
• publish / appear in / land in Nature, Science, Cell
• scale up / scale down / scale across
• train / pretrain / fine-tune / distill / quantize a model
• deploy / serve / host / containerize a model
• prompt / jailbreak / red-team / steer a model
• align / misalign / dealign an objective
• edit / knock out / knock in / base-edit / prime-edit a gene
• vector / deliver / target a therapy

Phrases and locutions

• the evidence is mixed / the evidence is robust
• a smoking gun — decisive evidence
• the dose makes the poison (Paracelsus)
• correlation is not causation
• the burden of proof lies with…
• a parsimonious explanation — Occam’s razor
• extraordinary claims require extraordinary evidence (Sagan)
• the null result
• a robust finding / a fragile finding
• statistically significant but not clinically meaningful
• effect size of practical importance
• the field is converging on
• the consensus is shifting
• the field is divided
• moving the field forward
• on the frontier of
• the cutting edge / the bleeding edge
• state of the art (SOTA)
• proof of concept (PoC)
• scaling pays off / scaling hits a wall
• first principles thinking

Common Russian-speaker mistakes

1. Theory meaning guess or hypothesis. Calque of Russian теория used loosely. In English scientific register theory = well-supported framework; hypothesis = testable proposed explanation; guess / speculation / conjecture = informal proposals. I have a theory that… in scientific context will be parsed as a strong claim; for tentative ideas use I hypothesize or I suspect or my hunch is.
2. Significant meaning important. In statistics statistical significance has a narrow technical meaning related to p-values. The difference was significant in a scientific context implies p < 0.05 or similar, not important. For important use important, meaningful, substantial, sizable, material, of practical importance. A clinically meaningful difference contrasts with a statistically significant but clinically negligible difference.
3. To prove applied to scientific theories*. Science doesn’t prove in the deductive sense; it accumulates evidence, supports, corroborates, fails to falsify. Mathematics proves. Newton proved gravity is wrong; Newton formulated the law of gravity or Newton developed the theory is right. The experiment proves the theory should be the experiment supports / corroborates / is consistent with the theory.
4. Researches as plural noun*. Research is uncountable in English: do research, conduct research, a body of research, the research suggests. The plural is studies, research projects, research efforts. Never researches as a noun; researches exists only as the third-person-singular verb (she researches X).
5. Quantum used loosely as large jump. In physics quantum means the smallest discrete unit (quantum of energy, quantum of action). The popular-press quantum leap meaning big jump is somewhere between a metaphor and a mistake; in scientific writing avoid using quantum to mean large. A quantum leap forward in a serious science article sounds amateur.
6. Experiment for any test or trial. AmE experiment has a specific meaning: a controlled comparison with intervention. Trial (clinical context), test, study, investigation are alternatives. We did an experiment to see what users prefer in a product context is usually called an A/B test or a user study, not an experiment.
7. Mathematic / mathematics as plural*. AmE mathematics is grammatically singular (mathematics is) and abbreviates to math (singular). BrE uses maths. I love mathematics; mathematics is hard — never mathematics are. Mathematic without -s exists only as an attributive adjective in rare combinations (mathematic precision); the standard adjective is mathematical.

Summary

• Philosophy of science vocabulary covers Popperian falsifiability, Kuhnian paradigms, Lakatos’s research programmes, the Duhem-Quine thesis, and the realism-vs-instrumentalism axis.
• Evidence hierarchy (case report → cohort → RCT → meta-analysis) and methodology vocabulary (preregistration, registered reports, GRADE, PRISMA) define credible empirical work.
• Statistical literacy at C2 distinguishes p-values from effect sizes, frequentist from Bayesian, type I/II errors, power, and causal inference (DAGs, IV, RD, DiD).
• AI vocabulary spans foundation models, transformers, RLHF/DPO, scaling laws, alignment (outer/inner/mesa), interpretability, red-teaming, and the AGI/x-risk debate.
• Quantum vocabulary covers superposition, entanglement, decoherence, the measurement problem, interpretations (Copenhagen/many-worlds/pilot-wave), and quantum computing (qubits, NISQ, supremacy).
• Biotech vocabulary spans CRISPR/Cas9 (guide RNA, PAM, prime/base editing, germline vs somatic) and mRNA vaccines (LNPs, modified nucleosides, immunogenicity).
• Russian false friends: theory for guess, significant for important, prove for support, researches as plural, experiment for any test.

Next theme: Environment and climate — C2 — anthropogenic forcing, radiative forcing, climate sensitivity, tipping points, geoengineering, carbon sequestration, ESG.