Start here
Never built with AI before? Come to a weekly talk, watch a past recording, and start with prompting well. You'll be building a simple agent sooner than you think.
Can code a little? Take a wedge of a real problem and ship something small for 20 real users. Pair with a domain expert from the community.
Bring a research question. Especially in biology and health — our life-sciences members and diaspora mentors are the whole point. See the products & services for what that looks like.
Curated resources
5 resources · curated for members
Say What You See
Learn image prompting by describing what you see — Google AI regenerates the image, and you only pass if your words hit the visual threshold. Tips as you go; three attempts per level.
Google Arts & Culture Lab · Jack Wild open →Skills for Design Engineers
Agent-ready skills for better UI taste — animation easing, shadows, Apple-style motion, and how to review motion so interfaces don’t look like slop. Install with npx skills or read the SKILL.md files.
Emil Kowalski (Vercel / Linear) open →The Component Gallery — Components
A living catalog of UI components (buttons, inputs, nav, modals…) with real examples from design systems. Use it to name patterns correctly and steal structure, not just aesthetics.
component.gallery open →The Component Gallery — Design Systems
Browse real-world design systems side by side. Learn how mature product teams document tokens, components, and principles — useful when you build anything members will reuse.
component.gallery open →AGX.fm — an infinite radio station with no recordings in it
A lo-fi station that never repeats and never plays a recording. Every kick, chord, vinyl crackle and reverb tail is synthesized from raw oscillators and noise the moment you hear it, so there is no master, no sample and no rights holder anywhere in it. And it is one broadcast, synced to the clock instead of a server: every listener hears the same track at the same moment, joining mid-track like real radio.
AgenticGHX · built with coding agents open →Shelf open
How to get real work out of an AI collaborator — the skill underneath every other track. Nothing curated here yet — suggest a link in community chat and we’ll review it.
Shelf open
Expand what you think is possible: art, play, world-building, creative AI. Nothing curated here yet — suggest a link in community chat and we’ll review it.
Shelf open
Biology, health, genomics — where many of our members already live. Nothing curated here yet — suggest a link in community chat and we’ll review it.
Shelf open
Devices, sensors, GPUs, edge — when the model has to touch the physical world. Nothing curated here yet — suggest a link in community chat and we’ll review it.
Shelf open
Interfaces, product craft, and making tools people actually use. Nothing curated here yet — suggest a link in community chat and we’ll review it.
Shelf open
Foundations: learning, agents, language, and the math behind the magic. Nothing curated here yet — suggest a link in community chat and we’ll review it.
No resources in this shelf yet.
Interactive explainers
Hands-on visual walkthroughs — poke at the ideas, don't just read them. More to come.
Foundation models: one model, five ways to use it
The whole field orders itself by one question — do the model's numbers change? Explore 21 models across 7 fields, and see why a fine-tuned 39M model beats a zero-shot 1.5B one on Akan, Ewe, Dagbani and Dagaare.
explore →The compute gap: what an AI GPU costs from Accra
Only one of the big three clouds will rent you a modern AI GPU in Africa — and the same 8×H100 machine costs 1.72× more here than in Virginia. Measured from the providers' own price lists.
explore →The token tax: why AI costs more in Twi and Ga
The same sentence costs 2.55× more in Ga than in English — and Ga is the shorter text. Measured with a real tokenizer, with the actual splits shown.
explore →Temperature 0 was never deterministic
Send the same request 30 times at temperature 0 and you get up to 26 different answers — seed pinned, on both major APIs. Where they fork, why the model is genuinely tied there, and why the newest models reject the temperature knob outright.
explore →The judge has a favourite seat
Almost everything in AI is now graded by another AI. We built pairs of answers where the right verdict is known in advance — including pairs where neither is better — and judged each one twice with the order swapped. Six judges, both vendors, and the result is not the one the textbooks predict.
explore →One flyer, two kinds of image
Our own flyer, as a 1.4 MB PNG and an 843 KB PDF: the smaller file keeps its words perfect at any zoom — and hides 72 DPI pixels inside. Raster vs vector, formats, editing and the one-way door of conversion, measured on the real files.
explore →When vector stops winning
Vector graphics are supposed to be small and to scale forever. Both halves stop being true at a measurable point — 30,312 points on a scatter plot. One rule, tested on a flag, a chart, and Ghana's 260 districts.
explore →The citations that don't exist
We asked four models for real, published references across twelve research topics and checked every DOI against the registry DOIs are actually issued in. The worst failure isn't the dead link — it's the DOI that works and belongs to a different paper.
explore →The AI glossary: 76 terms in plain English
Every term the other explainers lean on, in plain English: search 76 entries, filter by topic, then switch to test yourself and flip the page into flashcards.
explore →Agents don't write, they read
"An agent is just a prompt in a loop" — priced, on 36 real agent sessions from this repository. Where the money actually goes, what caching removes, why the growth stops, and the turn count at which a cheaper model stops being cheaper.
explore →What maps decide: Ghana, the UK, and the lying projection
Measured on the same geometry, Ghana and the UK differ in true area by 0.02% — and the web map draws the UK 2.87× bigger. What geospatial data is, six Ghanaian applications that end in a real decision, and a worksheet to design your own.
explore →What is an embedding?
Ten hands-on steps from "word → vector" to how attention reshapes meaning — spin a 3-D word cloud, watch a mini-model train, and see "bank" change sense in context.
explore →Embedding lab: poke a real 200-dimensional space
The companion lab to the embeddings explainer, on real vectors in Google's Embedding Projector. Nine predict-then-check exercises: why antonyms sit on top of each other, why a UMAP plot can't be trusted, and how to audit an embedding for bias.
explore →One geometry, six jobs
Semantic search and recommendation are the same function call. Six downstream uses of embeddings running on one set of 17 points — plus the metric switch that silently breaks deduplication, and the top_k bug that pads every RAG prompt with junk.
Also on this site
Recommended books
Why each title matters for AgenticGHX builders.
browse →Past talks
Every weekly talk, archived. Start with "How to Build an AI Agent."
watch →Blog
Field notes and essays from members — thinking in the open.
read →$ To add a resource: drop the URL + which shelf in chat
(or edit website/src/data/resources.json).
Human approve → site updates. Start with things that teach a skill, not a product pitch.