ABSLI · AI enablement
Internal learning experience
Distribution technology teams

Fig. 00 — Introduction
19.0760° N · 72.8777° E
Mumbai

You don't need to become an AI engineer. You need to learn how to make AI useful.

UserBuilderArchitect

From your first good prompt to designing systems that can reason, retrieve information, use tools and take action.

Fig. 01 — The route

Five stages. One climb.

Each stage adds a capability you did not have before. Select a station to see what you learn there, what you can build once you have it, and where the line goes next.

Stage 01 of 05Use

What you learn
What you can now build
Keywords unlocked
Next step

Reel 01–05 — Watch first

Watch before you build.

Five widely watched films, in the order that makes sense. Select a reel below, then press play. The first one is the only prerequisite for everything on this page.

Start here · Reel 01 But what is a GPT?3Blue1Brown · 27 min Plays on YouTube ↗

Why this one first

It shows, visually, what a language model does with the words you type. Once you have seen that, every other term on this page becomes a variation on one idea.

Open on YouTube

Fig. 03 — Live pipeline

How a modern AI system works.

One pipeline, four modes. Select a mode and watch the signal travel. Click any node to open it in the inspector. The agent mode will stop and wait for you.

Signal 00/00 — idle

Fig. 04 — Technology index

The tech, decoded.

Twenty-one parts of the machine. Each one answers two questions: what is it, and why would I care. Related parts are wired together, so you can follow a thread.

FoundationsPart 001 / 021

Prompt

What is it?

Why would I care?

Mental model

Wired to

Fig. 05 — Escalation

Where AI capability escalates.

Three levels. Same model, more connections. Watch the system grow: what changes is not the intelligence of the model but what it is allowed to reach and do.

Level 1 of 3

Basic

AI works primarily on what you give it.

Summarise
Rewrite
Extract
Compare
Classify
Analyse images
Analyse files
Generate structured output

Components03

AutonomyNone

NeedsA good prompt

Level 2 of 3

Intermediate

AI gets access to knowledge and systems.

Search private information
Retrieve knowledge
Call APIs
Use tools
Process documents
Trigger workflows

Components06

AutonomyLow

NeedsData access

Level 3 of 3

Advanced

AI can pursue goals across multiple steps.

Plan
Choose tools
Execute steps
Observe results
Retry
Delegate
Request human approval
Evaluate its own outputs

Components11

AutonomyBounded

NeedsEvals + guardrails

Exhibition — eight rooms · one screening room

People are building things that didn't look possible a year ago.

Not another chatbot. Not another to-do app. Real systems that see, move, reason, build and interact with the physical world.

Exhibit 01 / 08Accessibility · Computer vision

Live camera

Keep rightDrop-off on the left

Path confidence: 0.91
Slope: −11.5°
Clearance: 1.6 m
Obstacle: 2.9 m
Frame: 01100 · 30 fps

Camera
Perception
Safe path
Direction

A phone that finds the walkable path.

Harvard researchers built Mobilio, a navigation app that runs entirely on a smartphone. A custom computer-vision model reads the live camera feed from a pedestrian's point of view and finds walkable paths, blends in GPS, motion sensors and the phone's LiDAR where present, and guides blind and low-vision users with spatial audio cues.

Rendered scene. The trail above is drawn to show how a walkable-path model reads terrain. It is not source footage and not the app's real interface.

Why it's wild

In a 14-person study it cut indoor obstacle contacts by about 41% compared with a cane and Google Maps. No new hardware. The phone in your pocket.

Tech underneath

Read the project

Exhibit 02 / 08AI + hardware · Self-verification

It looked at the screen it had programmed. Then it fixed the code.

A developer who had never wired a board gave Claude Code a goal, a Wi-Fi network and a spare phone pointed at an ESP32's small display. The agent designed the setup, wrote and flashed the firmware, pushed updates over the air and wrote its own tests. Then it noticed, through the camera, that the clock font was silently dropping characters — a bug every automated test had passed — and fixed it.

Why it's wild

The tests said green. The camera said otherwise. The agent trusted the camera.

Tech underneath

Explore on GitHub

Conceptual loop. A summary of the workflow the project describes; the demo video is at the link.

Exhibit 03 / 08Robotics · LiDAR · Navigation

Eight non-roboticists. One robot dog. A beach ball.

Anthropic split eight researchers with no robotics background into two teams and asked both to make a Unitree Go2 fetch a beach ball. The team using Claude connected to the dog's camera and LiDAR, wrote its own controller — one member even got it taking plain-English commands — and by the end of the day the robot could find the ball and push it around on its own. They finished shared tasks in roughly half the time of the team without AI.

Why it's wild

A language model became the interface between people and a machine they had never touched.

Tech underneath

Read the experiment Watch the day

Exhibit 04 / 08Multi-agent · Autonomous software

Sixteen agents. One compiler. Nobody watching.

An Anthropic researcher put 16 Claude instances to work in parallel on a single shared codebase — each in a simple loop with a test harness — and then mostly walked away. Two weeks and roughly 2,000 sessions later, they had written a 100,000-line C compiler in Rust that can build the Linux kernel.

Why it's wild

The hard part was not the code. It was designing the harness so sixteen agents could work side by side without undoing each other's work.

Tech underneath

Read the write-up

Shared repository one codebase · task locks so two agents never take the same job

Test harness the only judge of "done" · thousands of test programs

Loop: pick task → work → test → commit repeat until nothing fails · ~2,000 sessions · 2 weeks

Exhibit 05 / 08AI controlling software

An AI that uses a computer the way you do.

01 ScreenshotLook

02 Find fieldLocate

03 Cursor · x 412 y 233Move

04 ClickAct

05 Type · re-checkVerify

Anthropic's computer use lets Claude look at a screenshot, work out where to move the cursor by counting pixels, click, type and look again — filling a form from a spreadsheet, moving between applications — with no integration built specially for it.

Why it's wild

No API required. If a person can click it, the model can attempt it — which is exactly why it needs a human close by.

Tech underneath

Read the announcement

Conceptual frames. A simplified picture of the see-locate-act-verify cycle, not a screen recording.

Exhibit 06 / 08Generative design · 3D printing

Parts described in plain English. Printed in plastic.

Describe a mount, a case or a stand. With this open-source skill, Claude writes parametric CAD code, renders the part from several angles, reviews its own render, and exports the file a 3D printer accepts. Parts designed this way — phone stands, drawer bins — are published for anyone to print.

Why it's wild

A physical object became a text file — which is exactly the kind of thing a language model can write, look at and fix.

Tech underneath

Explore on GitHub

Illustrative drawing. The bracket and dimensions are an example of the output type, not a specific published part.

Day 2 · 07:001 of 14 know about the party

Memory stream

Exhibit 07 / 08Multi-agent · Memory · Simulation

Twenty-five agents moved into a town.

Stanford researchers gave 25 language-model agents names, jobs, memories and a small pixel town. They woke up, cooked, went to work and chatted. When one agent decided to host a party, the invitation spread from conversation to conversation and guests showed up at the right place and time.

Why it's wild

Nobody scripted the party. It emerged from memory, reflection and planning.

Tech underneath

Explore on GitHub Read the paper

Conceptual simulation. Fourteen of the twenty-five residents are drawn here, enough to follow one invitation as it spreads. Not a recording of the original run.

Exhibit 08 / 08Robotics · Imitation learning

A robot arm you teach by hand.

Hugging Face's open-source robotics library and the SO-101 arm — around €114 in parts — let anyone move a leader arm by hand, record a handful of demonstrations and train a policy on a laptop until the follower arm repeats the task on its own.

Why it's wild

Robot learning used to need a lab. Now it needs a desk, a webcam and an afternoon.

Tech underneath

Explore on GitHub

Screening room / 04 clipsShared into the team chat

Four more, caught on video.

Short clips of working systems, kept here so they are easy to find. Nothing plays until you press it. Every one is someone else's project — the account on each card is the one to credit.

**Grading a rock face for safe footing.** NewZ Info Tech 0:55 · computer vision · safety

**Where shoppers actually look.** @gogreen_ai 0:39 · point cloud · retail

**Scoring how a cow walks, from five points.** @digitallivestock 0:54 · pose estimation · livestock

**A town of agents running a business.** @agenthusiastllm 1:19 · agents · concept demo

Not our work. Four clips shared into the team chat and parked here. Credit stays with the accounts named on each card. Watch clip 04 the way this page asks you to watch anything — the claim is large and nothing in it is verified.

Fig. 07 — Problem identification

Don't start with AI.Start with friction.

Then check whether the friction is the kind AI is good at

Instrument A — AI fit scanner

Describe the problem. Read the verdict.

Switch on everything that is true about one real task in your week. The readout updates as you go. It is a decision aid, not a score.

Signals 00Risks 00

Select what's true about the problem.

Decision aid, not a score. The scanner points. People decide.

Good signals

Repetitive
High volume
Unstructured information
Search, summarise, classify
Multiple manual handoffs
Clear inputs
Output can be verified
Human currently spends time finding information
People copy information between systems
Many similar cases

Red flags

No reliable data
No way to verify output
Extremely rare cases
Very high consequence of error
Full autonomy required from day one
Problem changes every time
Process itself is broken
AI is being added only because "we need AI"

Instrument B — Value × feasibility

Where does your idea land?

Show the examples to see how typical ideas spread out. Then switch to blank mode and place your own.

Four quadrants, one question

Value is what the problem costs today in hours, errors or delay. Feasibility is whether the data, rules and verification exist to build something in weeks, not quarters.

Quick wins go first. Strategic bets need a sponsor and a pilot. Experiments are cheap to try and cheap to stop. The last quadrant is where projects go to be forgotten.

Your ideaNot placed yetBlank mode, then click the map.

Fig. 08 — The build loop

How a good AI project actually runs.

Ten steps, then around again. Select a step on the ring or move through them in order. The gold signal never stops — that is the point.

Step 01 / 10Spot friction

Find the task people complain about, avoid, or do at 6 pm. Friction is the signal. AI is not the starting point.

A prototype proves the idea works once. Production proves it survives Monday.

Fig. 09 — Mission launcher

Build one yourself.

Concepts learned

Open resource

Fig. 10 — Route map

Two destinations. One line.

Every station is a real, open course or resource. Choose where you want to end up, then select a station to see what you learn there and what you can do afterwards.

Foundation Zone 01

Better AI use Zone 02

Build Zone 03

Connect Zone 04

Agents & AI coding Zone 05

Terminus: you can design and prototype AI systems.

Optional credentials

Anthropic partner certifications. Not on the main line, not required for anything on this page. For partner organisations that want a formal credential.

Partner onlyPaidOptional

Partner certifications

Associate $99
Developer $125
Architect Foundations $125
Architect Professional $175

Fig. 11 — Final destination

The same person, a few months apart.

Nothing about the model changed. What changed is what they can see in a problem.

Start

"I can prompt ChatGPT."

Finish

"I can look at a problem and sketch:"

01Data
02Model
03Knowledge
04Tools
05Workflow
06Guardrails
07Human
08Measurement

That's the shift.

From using AI to designing with AI.

Fig. 12 — One action

Do one thing today.

Watch "But what is a GPT?" by 3Blue1Brown. It is the only prerequisite.

Then use AI on one real problem from your week. Not a demo. A real one.

Start with the first video