Recently discovered terms, newest first.
A Lightricks image-to-video model name that appeared on Hugging Face’s trending list.
A Qwen text-to-image model version that entered Hugging Face trending as a new name.
A named autonomous penetration-testing project that broke out as a GitHub repository.
An open-source spreadsheet written in Rust; the repo name itself became the search term.
ByteDance’s video-generation model family, used as a named search term after each version drop.
A dated Anthropic Claude Haiku release used as a search name the week it shipped.
An image-editing product whose name spiked as people searched for a Photoshop alternative.
Google’s tool for spotting SynthID watermarks; the durable search is “ai image detector.”
Writing software by telling a model what you want and keeping what runs, instead of typing every line.
Shopping or checkout where an agent finds, compares, and buys on a person’s behalf.
Oxford’s 2024 word for the feeling that too much low-quality online content is dulling your mind.
An open standard for connecting models to tools and data sources through a shared interface.
Spending more inference-time thinking on a hard question instead of only training a bigger model.
The hidden cost of prompts that nobody owns, versions, or tests as the model changes.
Charli XCX’s 2024 album-as-aesthetic: lime green, messy confidence, mid-party energy.
Low-effort, high-volume AI text or media that fills feeds without adding much.
A job posting the company does not actually intend to fill right now.
AI systems that take a goal and then plan, call tools, and act across steps without a prompt each time.
A model that routes each token to a few specialist sub-networks instead of using every parameter.
Oxford’s 2023 word for romantic charm, especially the kind you perform in chat.
Showing up at the office just long enough to be seen, then leaving to work elsewhere.
An online identity that performs 1950s-style homemaking, often as a political or commercial aesthetic.
Model weights you can download, even if the license is not fully open source.
Spending a whole day in bed on purpose, as rest rather than as illness.
Interfaces that treat the room as the canvas: headsets, hands, and mixed overlays.
Using already-staked tokens again to secure another protocol, and to earn another yield.
A product or company designed around models from the start, not a feature bolted onto old software.
The quality drop when models are trained on too much of their own previous output.
Expensive taste that hides logos: good cloth, quiet color, no obvious brand.
A program that uses a model to decide actions toward a goal, often with tools and memory.
Search that answers with a model instead of, or on top of, a list of blue links.
Speeding up generation by letting a cheap draft model guess tokens a larger model then checks.
Creators telling audiences not to buy a hyped product, or to buy a cheaper stand-in.
A famous person whose career is easier to explain once you know their parents.
A prompt or trick that gets a model to ignore the safety rules its maker set.
Oxford’s 2022 word for dropping the performance of being put-together and living feral on purpose.
When a model states something false with the same confidence as a true answer.
Models that produce text, images, audio, or video instead of only classifying an input.
Quietly looking for the next job while you still have this one.
Cory Doctorow’s word for platforms that get worse as they squeeze users, then business customers.
Companies doing climate work but saying less about it, to avoid greenwashing attacks.
Hiding instructions in text or pages so a model follows the attacker instead of the developer.
Making a job so empty or hostile that the person leaves without a layoff.
Doing the job as written and refusing the unpaid extra that used to look like loyalty.
A phishing-resistant login that uses a device-bound public key instead of a password.
Moving supply chains to politically friendly countries instead of the cheapest ones.
Countries and firms reducing their use of the U.S. dollar for trade or reserves.
Asking a model to write intermediate steps before the final answer.
Training a model to prefer answers humans rank as better, using a reward model.
A cheap way to adapt a large model by training small rank-decomposition matrices.