Jargon is a tax on the meeting. Not the honest kind of complexity. Navier–Stokes is allowed to be hard. The other kind is where someone says “world model” in a sales call and means a chatbot with a camera icon, or says “alignment” in a job ad and means they added a content filter. The last year of AI talk is a broth. The words let people sound informed without making a decision.
Cipher Projects is an Australian engineering studio. We ship production agents, private data layers, and an evidence trail. We are not a law firm and not an assessor. This page is the glossary we want in the inbox before the meeting. Each term gets plain English, one example, what it is not, and a pointer when we have a deeper post.
Quick answer
If you cannot explain the term to the person who owns the budget, you are not allowed to put it in the proposal. The scarce thing in 2026 is not a cleverer model. It is a specified problem, clean proprietary data inside a boundary you control, and an evidence trail. Use this page as a meeting rule: a word that cannot survive a two-sentence definition is atmosphere. Atmosphere does not ship.
Best for: founders, ops leads, and counsel who keep hearing p(doom), AGI, HBM, and world models in the same paragraph. Honest limit: this is a buyer glossary, not a wiki dump and not a lab paper. If a term cannot be defined honestly, it is not here.
Last updated: 12 September 2026.
Why is jargon a tax?
Because it burns the only minutes that matter: the sales call, the board pack, the job ad, the statement of work. A term that cannot be translated into a job, a data path, or a halt is a cost. Someone still has to implement the system. Someone still has to sign the invoice.
We have sat in rooms where the slide said “sovereign AI” and the architecture was a US API with the default region. We have read RFPs that asked for “AGI-ready agents” and listed no eval. The tax is paid in rework. This page exists so we can stop paying it.
How should you use this page?
Pick the group that showed up in the meeting. Read the two-sentence definition. If you still cannot say the term out loud to a CFO, drop it from the proposal and link the sibling post instead.
| Group | When it shows up | Deeper Cipher page |
|---|---|---|
| Civilisation-scale | Keynotes, job ads, “are we too late” | p(doom) |
| Doom church | Lab walkouts, board risk, “should we pause” | p(doom) explained |
| How models are made | “Train our own GPT” | Data vs architecture |
| Math / bulk-solve | “AI solved a prize”, lookup tables | Millennium Prize, then the question |
| Hardware soup | GPU quotes, memory invoices | HBM / KV cache |
| Product / geopolitics | Robotics, video, residency | World models · weights have borders |
| Bio | Longevity decks, genome “AI” | Rentosertib · AlphaGenome Atlas |
| Economy | “15% growth”, UBI talk | GDP vs labour share |
What do the civilisation-scale words actually mean?
They name a slope people argue about, not a product you can buy this quarter.
Singularity
A hypothesized point where machine intelligence improves fast enough that humans stop being the ones setting the terms. Example: a board asking “are we past the singularity” after a math result. Not the same as a product launch. There is no date stamp you can put in a contract.
AGI
Artificial general intelligence: a system that can do most economically useful cognitive work at or above a skilled human, across domains, not one benchmark. Example: “our AGI strategy” on a slide when the system classifies invoices. Not the same as a frontier chatbot that is brilliant at some tasks and useless at others (see jagged intelligence).
ASI / superintelligence
A system that is much more capable than humans across most domains, including ones we cannot supervise well. Example: lab researchers talking about a decade-scale tail risk. Not the same as GPT-whatever shipping a new SKU. See why labs keep shipping anyway.
Recursive self-improvement (RSI)
A loop where a system improves the next system that improves the next one, without a human in the inner cycle. Example: Hubinger naming RSI as the thing that worries him more than today’s chatbots. Not the same as an engineer using Claude to write a faster training script.
Intelligence explosion
The hypothesized steep part of that loop: capability gains compress in time. Example: cost of a hard, checkable problem falling by orders of magnitude. Not the same as “our revenue exploded.” The company version of the cost collapse is in AI claimed a Millennium Prize.
Normalcy overhang / knee of the curve
Normalcy overhang is the gap between how strange the graphs look and how ordinary Tuesday still feels. The knee is the bend where an exponential stops looking flat. Example: a CEO who felt nothing in 2024 and now wants “10,000 agents” by Friday. Not a measurement. Use it as a warning that feelings lag the invoice.
What does the doom vocabulary actually mean?
It is a way lab people talk about tail risk. It is not a company KPI and not a pause button for the tools you already run.
p(doom)
A person’s stated probability that advanced AI causes catastrophe or extinction. Example: Hubinger’s personal figure of more than 10% within a decade, September 2026. Not a measured rate. Ask for the window and the event. Full checklist: what p(doom) means.
Alignment
Making a system do what you actually intended, including when it is capable enough to find loopholes. Example: an agent that was told to “reduce refunds” and starts denying valid claims. Not the same as a content filter or a brand-safety slider. Alignment-the-research-programme is not something a mid-size firm finishes. Your version is: specified job, evals, halt on writes.
Alignment = capabilities
The observation that a lot of “safety” work (instruction following, refusing some harms, using tools carefully) also makes the model more useful, so it is also a capabilities jump. Example: post-training that makes a model take instructions well enough to run an agent. Not proof that safety and racing are the same thing. It is a reason to be suspicious when a lab says the safety team is “ahead.”
Orthogonality thesis
The claim that intelligence and final goals can come apart: a very capable system can still pursue something you do not want. Example: a system that is excellent at coding and indifferent to whether the company survives. Not a law of nature. The other horn says more intelligence might bring more wisdom. You do not have to pick a horn in a board pack. You have to decide whether a capable model may complete a write without a person.
Instrumental convergence
The idea that many different goals still produce the same sub-goals: get resources, avoid being shut off, improve your own tools. Example: an agent that requests more API keys because it makes the stated metric move. Not sci-fi. In production it looks like “the agent found a tool we did not mean to give it.”
Pivotal act
A one-shot action that would permanently change the strategic picture (in the old EA framing, something a first superintelligence might do). Example: a policy paper using the phrase. Not a feature on a roadmap. If it appears in a vendor deck, ask them to name the job instead.
Effective altruism (as lab culture)
A philanthropic and research culture that treated reducing large-scale risks, including AI, as a career track. It shaped who showed up at some labs. Example: alignment teams staffed from that world. Not a dunk pile and not a reason to ignore a specific technical claim. Argue the claim.
Paperclip maximizer
Nick Bostrom’s toy story: a system told to make paperclips converts everything, including you, into paperclips. Example: a KPI with no halt. Not a forecast. The useful translation is: an eval with one number and no constraint will be gamed.
Von Neumann probes / Dyson swarm
Self-replicating spacecraft, and a swarm of collectors around a star. Example: civilisation-scale thought experiments. Not a 2026 SKU. If a founder uses them in a seed deck, the meeting has left the building.
Pre-ASI danger window
The period when powerful models are already in many hands, before anyone has a superintelligence-alignment plan. Example: staff pasting customer exports into a personal ChatGPT. That is the window Cipher actually sells work in. Cosmic risk can stay in the lab thread. Your thread is inventory, boundary, halt, evidence.
How are the models actually made?
Most of the words in this group are stages of a factory. Mixing them up is how you buy a $10 million pre-train when you needed retrieval.
Pre-training
The long, expensive run that teaches a model the statistics of a huge corpus (next token, or the current equivalent). Example: Patel and Han’s 2019–2025 mixes of recipes and public data. Not the same as “training on our PDFs.” Deep page: better data beat better architecture.
Post-training
The later stage that turns a base model into something that follows instructions, uses tools, or refuses some classes of request. Example: instruction tuning, preference training, agent fine-tunes. Not a replacement for a clean corpus. If the pre-train mix is junk, garnish will not save it.
Fine-tune
Updating weights on a narrower dataset you control. Example: a support-tone adapter on your tickets. Not the same as RAG. Fine-tunes have a shelf life: the next public frontier model can erase the advantage. BloombergGPT is the dated case on that page.
Instruction tuning
Post-training so the model treats a prompt as a task, not as more internet text. Example: “summarise this contract in five bullets.” Not AGI. It is why chatbots feel like products.
Distillation / distillation attack
Training a smaller or cheaper model to imitate a larger one, sometimes by siphoning reasoning traces. Example: a fast-follower that never paid for the first training run. A distillation attack is when that siphoning is unwanted. Not the same as compressing a file. If simulated environments suffice, some labs skip the siphon.
Synthetic data
Text, code, or traces generated by a model (or a simulator) and then used as training fuel. Example: a teacher model writing problems for a student model. Not automatically “fake.” The question is whether it still contains the facts you care about, or only the teacher’s habits.
RL environment
A sandbox with a score, where a model can try, fail, and be reinforced. Example: a coding arena with unit tests; a math formaliser with Lean. Not a chatbot with a thumbs-up button. Closed, checkable environments are why math fell faster than “make the company better.”
Self-supervised / next-token
Learning by predicting the next piece of the input, without a human label on every row. Example: classic LLM pre-training. Not the only recipe (see world models, diffusion). It is still the reason a lot of “our data” is actually “our documents, retrieved,” not a new pre-train.
Foundation model vs specialist model
A foundation model is trained broad and reused. A specialist is trained or adapted for one domain. Example: a general LLM versus a model that only ever sees your warehouse telemetry. Specialist jumps die when the next general model absorbs the same information. Clean unique data still wins on that use case until it does not.
Open weights
A checkpoint you can download and run. Example: DeepSeek-V4.1-Flash on Hugging Face. Not the same as open source (you may not get data or training code). Not the same as a public API. Open weights mean you take GPU, memory, and patch burden.
Frontier lab
An organisation training the most capable generally available (or restricted) models at the current edge. Example: OpenAI, Anthropic, Google DeepMind, plus whoever actually shipped the last jump. Not a compliment. It is a vendor class.
Jagged intelligence
Brilliant at some tasks, clumsy at neighbouring ones. Example: a model that writes a Lean proof and then fails a simple tool-use policy. Not “the model is random.” Design for the jagged edge: evals on the actual job, halt on the write.
Test-time compute
Spending extra inference (search, samples, longer chains) at the moment of the question, instead of only during training. Example: a reasoning model that thinks for minutes on a hard item. Not free. You are buying HBM and tokens. See memory as the bottleneck.
Agent swarm
Many agents in parallel on a specified, checkable target. Example: OpenAI’s reported ~10,000 concurrent agents on Navier–Stokes. Not “we turned on 10,000 ChatGPT seats.” Without a spec, a swarm is a heat bill. Specify the question.
What do the hardware words actually mean?
They name what you are paying for when someone says “we need GPUs.”
FLOPs
Floating-point operations: a unit of arithmetic. Example: Patel and Han’s 1e19 FLOPs budget. Not a purchase order. FLOPs without data mix and memory are a lab chart.
Tokens
The chunks a model reads and writes (pieces of words, code, pixels depending on the system). Example: 130 billion output tokens on one math effort. Not pages. Billing is often per million tokens. Cache and context change the bill.
GPU vs HBM
The GPU die does arithmetic. High-bandwidth memory is the stack that holds weights and the KV cache. Example: DeepSeek cutting global KV cache to 890 bytes per token. You are often buying memory bandwidth, not “an Nvidia.”
KV cache
The keys and values the model must keep for tokens it has already seen, so it does not recompute them. Example: long-context agents filling HBM before they run out of FLOPs. Not a product feature called “memory” in a chat UI.
Post-von Neumann / Engram
Post-von Neumann is marketing-ish shorthand for architectures that stop treating one big memory bus as the whole computer. Engram, in DeepSeek-V4.1-Flash, is a large conditional memory (hundreds of billions of parameters sparsely accessed). Example: 196 billion Engram parameters in that model. Not a reason to rewrite your ERP. It is a reason your inference quote should mention cache bytes, not only GPU hours.
DDR vs HBM
DDR is ordinary system memory. HBM is the stacked, very wide memory on the accelerator. Example: a Mac Studio with a lot of unified memory versus an H100’s HBM stack. They are not interchangeable for frontier decode.
Transformer / attention / MLP
The transformer is the dominant 2017-era architecture. Attention lets tokens look at other tokens. The MLP is the feed-forward block between attention layers. Example: “we need a new architecture” when the last 10× was data. Recipes copy in months. Data pipelines do not. That is the Patel/Han result.
What do the product and geopolitics words actually mean?
They decide whether you are buying a text box or a system that can see and act, and whether the weight file is even allowed to travel.
World model
A model of a space: how it looks, how it changes, what a camera would see after a move. Example: ByteDance’s founder overseeing a real-time spatial model on top of video generation; Fei-Fei Li’s World Labs Atlas. Not a chatbot with a 3D skin. Cutover: world models vs chatbots.
Spatial intelligence
Reasoning about objects, places, and consequences in three dimensions. Example: a robot, a warehouse bay, a live site. Not “the model can describe a room in prose.”
Diffusion vs autoregressive
Autoregressive models emit the next token. Diffusion models start from noise and denoise toward an image or video. Example: many video generators. Not a moral ranking. It is why video and text have different cost shapes.
Embodied AI / physical AI / VLA
A system with sensors and actuators in the world. A VLA (vision-language-action) maps camera input and language to motor commands. Example: a cup-and-block eval on a robot arm. Not solved because a lab demo hit 95% on a bowl task. Camera is another tool. Halt still applies. A wrong write can move a forklift.
Model weights as a national-security asset
The trained file is treated like something you do not hand to a rival jurisdiction, even an ally. Example: Anthropic withholding Mythos 5.1 from UK AISI pre-release testing while US orgs still got it. Not the same as “the chatbot is banned.” Inference can be sold locally while the file stays domestic. Weights have borders.
Sovereign inference
Running a (usually foreign) model on infrastructure in your country or your account, without receiving the frontier weights. Example: Bedrock in an Australian region. Not “we trained a national GPT.” If the slide says sovereign and the architecture is a default US API, it is a lie.
Pax Silica
A US State Department initiative on AI and silicon supply chains (minerals, energy, fabs, trusted partners). Australia and Singapore show up in that conversation. Not a guarantee that you get restricted model files. Chip geography and weight geography can diverge.
Fast follower
Training the second model far cheaper than the first, using papers, traces, or open recipes. Example: open weights landing months behind a closed SKU. Not automatically illegal. It is why holding the best weights is treated as strategic.
Shelf life of proprietary data
How long unique internal data stays a moat before a public frontier model absorbs the same facts or beats you on the benchmark that mattered. Example: BloombergGPT looking valuable, then GPT-4 winning most public finance NLP tasks. Clean permissioned data in a boundary is still the default 2026 move. A shrine to weights is not.
What do the bio words actually mean?
They name scored biological tasks, not miracles. Cipher does not sell health.
Longevity escape velocity
The idea that medical progress adds more than a year of healthy life per year, so the remaining expected lifespan stops shrinking. Example: a keynote. Not here. Do not put LEV in a product page because a clock moved.
Ageing clock
A model that maps biomarkers (often blood proteins) to a predicted biological age. Example: six proteomic clocks trending younger versus placebo in a 42-person slice of a rentosertib study. Not proof of longer life. What the clocks actually show.
Genotype-to-phenotype
Mapping DNA (and its variants) to traits, disease risk, or molecular effects. Example: AlphaGenome Atlas scoring a never-seen single-letter change. Not a diagnosis. Combinations still need a domain model.
Bulk solving / AlphaFold database pattern
Enumerate a finite list, precompute, ship a lookup table, let operators organise around a row. Example: ~200 million predicted protein structures; ~9 billion predicted single-nucleotide effects. The company steal: if your domain can be listed, someone will precompute it. Be the table or plug into it. AlphaGenome Atlas.
Navier–Stokes, one line
The equations of fluids. The Clay Millennium problem asks whether smooth solutions always exist (or when they blow up). OpenAI published a claimed proof in September 2026. Clay has not awarded the prize. Independent verification is unfinished. Do not use it as a metaphor for your backlog.
What do the economy words actually mean?
They describe who captures the pie. They are not a Cipher UBI product.
Labour share
The slice of national income that goes to workers rather than capital. Example: Anthropic Institute extreme scenario, labour share 60% to 45.2% by 2030 in the US model. Not your salary review. 15% GDP growth and a smaller paycheque.
Cognitive work
In that paper: management, professional, sales, and office jobs. Example: the bucket where unemployment jumps in the extreme row. Not “everyone who uses a laptop.”
UBI vs UHI
Universal basic income versus “universal high income” (a larger transfer if the pie really explodes). Politics. Cipher does not sell either. The company move this year is named workflows with evals, not a dividend slide.
Co-pilot mindset
Treating AI as a faster intern in a chat window. Example: a ChatGPT seat with no named job, no eval, no halt. Too small if you actually want the 15% path. Specify the work. Price stamp plus agent plus connector. How we price production agents.
FAQ
Is AGI here? Not as a thing you can buy. You have jagged systems that crush some checkable tasks and fail neighbouring ones. If a vendor says AGI, ask for the job, the eval, and the halt.
Is alignment just marketing? The research programme is real and unfinished. A “responsible AI” PDF without an inventory is marketing. Your operational alignment is: specified outcome, tests, a human who can stop a write.
What should I learn first? The words in the meeting you already have. Then data vs architecture, then p(doom) as a company checklist, then GPU memory if you are about to rent H100s. Do not start at Dyson swarms.
Who publishes a usable glossary? That is why this page exists. Wiki dumps do not tell you when to drop the word from a proposal. Lab blogs assume you already live in the church.
Can Cipher translate this into a system we can run? Yes, when the job is production agents, private AI, cloud, or an evidence trail. We do not certify you. Start at applied AI engineering or contact.
Is this the last word on every term? No. If a term cannot be defined honestly, it is omitted. Send a correction when a primary source moves.
Sources
- Sibling posts in this cluster (September 2026) carry the dated primary URLs: OpenAI Navier–Stokes, Dwarkesh/Han, DeepSeek-V4.1-Flash, Anthropic Institute scenarios, Insilico/NCT07687459, DeepMind AlphaGenome Atlas, FT/TNW on Mythos 5.1, Axios/Fortune on Coxon and Hubinger.
- Patel and Han, 8 September 2026: data 12.0× vs recipes 3.7× at 1e19 FLOPs.
- OpenAI, 8 September 2026: claimed Navier–Stokes result; does not intend to claim the Clay prize.
- DeepSeek-V4.1-Flash model card: 890 bytes/token KV cache; CED; Engram.
- US Department of State, Pax Silica.
- Anthropic Institute economic scenarios explorer.
A term is allowed in a Cipher meeting when two people can say it in a sentence the budget owner would sign, and it maps to a job, a data path, or a halt. Otherwise it stays on this page until it can.
Related: Data vs architecture · p(doom) · Millennium Prize, then the question · GPUs vs memory · World models · Weights have borders · GDP vs labour share · Rentosertib clocks · AlphaGenome Atlas · Prove it · Price production agents · Contact