The AI Garden newsletter

One concept explained properly, and nothing else.

Each email walks through one AI idea in plain words,
in 3 minutes.

19 Sept 20265 min

ModelsNote

Jev: the AI model that answers with numbers, not words

TypeSafe AI’s first model never writes a sentence. It reads your text, answers fixed questions with probabilities, and leaves the decision to your code.

On a weathered greenhouse potting bench, a wooden seed-sorting tray with three compartments holds a handful of seeds, most in the first compartment and one in the third. A small brass dial gauge with a blank face stands beside it, with seed packets, a trowel and potted plants around.

17 Sept 20262 min

ModelsNote

“Take a deep breath” helped an AI with maths

Another AI found the odd instruction. What happened in the test is more interesting than the idea of a robot calming its nerves.

On a greenhouse potting bench, one brass toy robot rests a hand on a fan of blank cards. A single card lies between it and a second robot beside a closed book, with potted plants behind them.

16 Sept 20265 min

ModelsNote

Hugging Face: from a teen chat app to a $12.9 billion emoji

In 2017 it was an app for bored teenagers, named after the 🤗 emoji. Then its makers gave away the code behind it, and the whole AI world moved in.

A sunlit greenhouse potting bench. A wooden cabinet of small drawers stands open on rows of plain cream seed packets. Leaning against it is a round yellow enamel sign painted with a smiling face holding up two open hands. Spilled seeds lie beside a seedling in a terracotta pot.

16 Sept 20266 min

ModelsNote

Six ideas behind AI language models

A short history of how neural networks carry context, turn one sequence into another, and handle more work.

Illustration of a greenhouse potting bench with a winding paper ribbon, notebooks, a wooden gate, paper trays, seed compartments and a lamp.

12 Sept 20265 min

SafetyNote

Zero-days: the software flaws nobody knows about yet

A zero-day is a weak spot in software that its makers haven’t found, so they have had zero days to fix it. AI models are now finding them too.

A walled kitchen garden in soft morning light. On the left a weathered wooden gate is shut with a brass padlock. Further along the same fence, low to the ground, one board is missing, and a small brown rabbit peeks in through the gap beside a few nibbled lettuce leaves.

12 Sept 20264 min

ModelsNote

Nano Banana: the AI image model named at 2:30 a.m.

Google’s hit picture-making AI got its silly name when a product manager mashed two nicknames together in the middle of the night. Here’s that story, and how AI turns words into a picture.

A moonlit greenhouse potting bench. On a small wooden easel, a framed canvas is grey static on its left half and a painted yellow banana on its right, the static breaking up into the painting. Beside it a tiny banana sits on a terracotta saucer, and an open brass pocket watch shows half past two.

12 Sept 20265 min

ModelsNote

Six of seven chatbots picked the same “random” number

Ask for a number from 1 to 50 and 27 keeps coming back. What that says about the way a model chooses every word.

Seven wooden plant labels stand in a seed tray of dark soil on a greenhouse bench. Six are hand-written in pencil with the number 27, and the seventh, turned slightly away, reads 42. Terracotta pots, young seedlings and a walled garden sit behind the glass.

11 Sept 20265 min

ApplicationsNote

Airline Ticket Prices: How AI Picks Prices in Between

Since the 1980s, airlines have sold seats at a few fixed prices and switched between them as a plane filled up. Now AI can pick a price in between and change it during the day.

On a sunlit greenhouse potting bench, a wooden stair-step stand holds plain cream paper tickets on its steps. Beside it, a smooth rising brass rail carries a folded paper aeroplane partway along. Terracotta pots and a walled garden arch sit behind.

9 Sept 202610 min

ScienceNote

Navier-Stokes: The million-dollar puzzle about flowing water

Start with a swirl of water, draw its motion with arrows, and use simple arithmetic to see what it would mean for the equations to break down.

Teal dye curls through water in a clear glass bowl on a greenhouse bench. A wooden spoon lies beside the bowl, with ferns and terracotta pots behind it.

9 Sept 20265 min

SafetyNote

SynthID: How a Watermark Hides Inside AI Word Choices

Google DeepMind published the method in Nature in October 2024 and ran it on about 20 million Gemini replies without anyone noticing. Claude is adopting it. This is how a watermark hides in the choice of words.

A sheet of thick cream paper is propped against a terracotta pot on a weathered greenhouse bench. Low sun lights it from behind so that a faint fern-leaf watermark shows through the paper. A fountain pen lies in front, with ferns, seed trays and pots behind.

8 Sept 20265 min

SystemsNote

Quantisation: How Large Models Fit on Your Laptop

A language model is really just a very long list of numbers. Round those numbers off, and the file gets dramatically smaller and runs faster, for a small trade in quality.

On a weathered wooden potting bench in a sunlit greenhouse, a thick, closed, green leather-bound book lies beside a slim closed book in the same green leather, with an identical pressed fern frond resting on each cover, and terracotta pots of ferns soft in the background.

8 Sept 20264 min

ModelsNote

Mixture of Experts: The Experts Aren’t Separate AIs

Inside one model, a router sends each token through a few internal blocks. Follow one token through the machinery and see what “expert” really means.

Brass irrigation valves fan out toward terracotta herb pots on a greenhouse bench, with two outlets spraying water.

8 Sept 20264 min

ModelsNote

Long Context: Why AI Can Miss the Page That Matters

A document can fit inside an AI model’s context window and still be used unevenly. The position of the useful passage can change the answer.

A long cream accordion-fold document stretches across a weathered greenhouse potting bench, with one panel in the middle marked by a green leaf and partly hidden between the folds.

8 Sept 20265 min

SystemsNote

Embeddings: How AI Search Works Without Exact Keywords

Type "money back" and a keyword search can miss the refund page. AI search turns every sentence into a point on a map and picks the nearest ones, so it finds the page. It can still pick the wrong passage first.

A sheet of cream card on a wooden potting bench in a sunlit greenhouse holds dozens of pressed specimens in groups: fern fronds on the left, round leaves on the right, small flower petals below, a few lone specimens between them, and one fresh green leaf beside the round leaves.

8 Sept 20264 min

SystemsNote

AI Memory: Why “Remember This” Is Not Model Training

An assistant can carry your preferences into a later conversation without rewriting what its language model learned. Here is where that memory actually lives.

One concept explained properly, and nothing else.

Jev: the AI model that answers with numbers, not words

“Take a deep breath” helped an AI with maths

Hugging Face: from a teen chat app to a $12.9 billion emoji

Six ideas behind AI language models

Zero-days: the software flaws nobody knows about yet

Nano Banana: the AI image model named at 2:30 a.m.

Six of seven chatbots picked the same “random” number

Airline Ticket Prices: How AI Picks Prices in Between

Navier-Stokes: The million-dollar puzzle about flowing water

SynthID: How a Watermark Hides Inside AI Word Choices

Quantisation: How Large Models Fit on Your Laptop

Mixture of Experts: The Experts Aren’t Separate AIs

Long Context: Why AI Can Miss the Page That Matters

Embeddings: How AI Search Works Without Exact Keywords

AI Memory: Why “Remember This” Is Not Model Training

One thoughtful walk through AI each week.