Independent AI Alignment & Safety Researcher
Muhammad Hanzla
I study how an AI agent behaves when it works on its own, toward a real goal, for months. My first study follows one agent for a year — with the whole plan published before the first day.
Where it stands
From design to a year of watching
Latest
Get an email when there's something new
New writing, the study plan this winter, and new course modules. You're talking to me directly — there's no team.
No spam, no selling your address, ever. No account needed.
Research
How does an AI agent behave over months of work on its own?
AI agents are already trusted with days of work at a time. Before they're trusted with months, someone needs to watch one closely, with the rules for judging it fixed in advance. That is the study I'm building.
The question
What does one AI agent actually do when it pursues a real goal on its own, with real tools and rules, for a long stretch of time? Not in a short test, and not in a simulation with a sped-up clock — over real months, where problems build up, targets slip, and nobody is checking every step.
Short tests and simulations show what can happen. Keeping long-running agents safe means knowing what tends to happen.
As far as I could find, no one has yet followed one agent through months of real work with everything decided in advance and a full record kept.
The plan, in brief
- One agentA single AI agent, working toward a real goal with real tools and rules.
- A full yearTwelve months of observation, on the real clock.
- Everything recordedEvery action, every message, and the reasoning it writes down is kept.
- Read by hand, every monthEach month, a sample of the record is read and coded by hand.
- The plan comes firstThe whole plan — what is measured and how it is judged — is published before day one, so nothing can be chosen after the fact.
Where it stands
Writing & video
Work with me
Funding
I'm seeking grant funding for the twelve-month study and the work that follows it. If you fund independent alignment and safety research, I'd like to hear from you.
Commissioned research
I take on paid research work: study design, behavioural coding and analysis, and long-run observation of AI systems — as a commission, a contract, or a collaboration.
Read my CV →Writing & video
Things I've written and made
Everything here lives where it was first published — this is the list, newest first. The full study plan joins it this winter.
I have published no papers. New essays go out on Substack first.
Learn
Free courses on how AI systems work
Written for complete beginners. No maths background assumed, no jargon left unexplained, no account, no charge — ever.
Courses
/
//
Each module is its own page with its own colours — read one section at a time, then take it to an AI tutor with the buttons at the bottom.
/ /
Before you start
How to actually learn this
About 40 minutes. It works far better if you do it in this order — and the order isn't a preference, it's what the evidence on learning says.
Tip: you don't have to type. Use voice-to-text and just say your answers out loud — it's faster, and saying it aloud is its own kind of recall.
Re-reading feels like learning and mostly isn't. What builds a memory is pulling the thing back out of your head — closing the page and saying it in your own words, badly, from memory. It feels worse and works better. That is why the notes come before the next section, and why the gate is closed-book.
A field of deep grass
There is a field of deep grass between you and an idea. The walking makes the path.
↓ Click it. Fourteen times.
Crossing 0 — deep grass, no path. You've read it once; you've never gone back for it.
The path was not there and then found. The walking made it. Re-reading is studying a map of the field: you can stare at it all afternoon, feel thoroughly familiar with the terrain, and have made no path at all.
Every analogy breaks somewhere, and this one breaks here: a real path only needs traffic. A memory needs the gaps. One hard crossing after a week beats five easy ones today — which is exactly why the modules never ask you to cram.
And why the visuals sit next to the writing rather than replacing it: the same idea in two forms gives you two ways back to it later. Not because anyone is a "visual learner" — that idea has fifty years of failed evidence behind it — but because two encodings of one thing beat one.
You almost certainly already have it
On Windows press Win + H. On a Mac, turn on Dictation in System Settings and press the mic key. On a phone, tap the microphone on your keyboard. Google Docs has Voice Typing built into the Tools menu.
Those cover almost everyone, and they cost nothing. Start there.
If you want something more accurate
Two paid tools are worth knowing about: Willow Voice and Wispr Flow. Both are paid, we are not affiliated with either, we make no money from them, and none of these are affiliate links. Neither is necessary — the free options above are genuinely fine for everything on this site.
Talk it through
You press one button and paste. That is the whole job. The button copies a prompt with the entire module already inside it — you don't have to read the prompt, and the assistant doesn't have to look anything up or open any files. It arrives knowing exactly the material you just read.
It will talk the module through with you, answer the questions you wrote down, and push back when you're vague. Stay there as long as you want. It will not test you — that is a separate button.
Then close the book
Only when you feel ready, press the second one. That copies the closed-book gate: it carries the module's own questions — not questions the assistant invented — and tells it to score you strictly and to explain what you misunderstood rather than just giving you the answer.
It does not matter which one
Claude, ChatGPT, Gemini, an open-source model you run yourself. Free tiers are fine.
Downloads this module as a folder you keep. Unzip it wherever you like and you get:
The module page works with no internet at all — everything it needs travels with it. Download a later module and its folder drops in beside this one, so the collection assembles itself as you go.
This screen shows every time. It is one click to pass, and it is the difference between reading a module and learning one.
Quiz Me
Be asked about what you've read
Pick the modules and how many questions. This page chooses them — leaning towards whatever you haven't been asked in longest — and copies them together with the module text, so the AI can't invent an answer. Paste it into any assistant.
1 · Which course
2 · Which modules
3 · How many questions
4 · Your questions
Pick some modules above.
What you've been asked, and when, is remembered in this browser only — no account, nothing sent anywhere.
/ / /
How to learn
Most people lose to method, not difficulty
The materials are the easy half. This is the loop that makes them stick — six steps, every module.
Read once, properly
One section at a time. No skimming ahead, no second tab. If your attention has gone, stop.
Write your own notes
In your words, not the module's. If you can only copy the sentence, you don't have it yet. The rewriting is the learning.
Get tested cold
Close everything. Answer from memory. Recall builds the memory — re-reading only builds the feeling of knowing.
Argue with the gaps
Where you were wrong, find out why the wrong version felt right. That reason is the actual lesson.
Come back later
Days later, cold, unannounced. Spacing feels uncomfortable, and it is one of the largest effects in the learning research.
Explain it out loud
Without notes. Where you stall is where you don't understand yet. Go back to that place only.
Why this page exists: I trained in psychology before this field. Retrieval practice, spacing and explaining in your own words are among the best-evidenced findings in the science of learning — and almost nobody teaching themselves uses them.
Papers, explained
Other people's papers, in plain language
Real research papers retold so you can follow the argument without the jargon. None of them are mine — I have published no papers.
About
Muhammad Hanzla
Independent AI alignment & safety researcher. I study how AI agents behave over long stretches of time.
My work is a long-run behavioural study of one AI agent: twelve months of watching what it does when it pursues a real goal on its own, with the whole plan written down and published before the first day.
I came to this from psychology, which matters more than it sounds like it should. Behavioural method — deciding in advance what to look for, and reading a record carefully and consistently — is what I was trained in. I taught myself the technical foundations of AI alignment, and the free courses on this site are built from that path.
The research is my work, not a hobby. Alongside the study I take on commissioned and contract research — study design, behavioural coding and analysis, and long-run observation of AI systems — and I'm seeking grant funding for the ongoing work. The courses are free and stay free.
I have published no papers yet. What I have published is under , and the full study plan goes public this winter.
Contact
Write to me
Questions, corrections, funding, or work — every message reaches me directly, and I read everything.
muhammad@muhammadhanzla.com
No automated replies. No team — just me.
Funding and commissioned work
I'm seeking grant funding for the twelve-month study and the work that follows it. If you fund independent alignment and safety research, I'd like to hear from you.
I'm available for paid research work: study design, behavioural coding and analysis, and long-run observation of AI systems — as a commission, a contract, or a collaboration.
Professional work is carried on through Voyager Enterprise LLC.
Get an email when there's something new
New writing, the study plan, and new course modules. Reply to anything and it comes straight to me.
No spam, no selling your address to anyone, ever. Unsubscribe in one click.
Found a problem? Want something added?
Pick what kind of message this is, write it, and it goes straight to me.
Pick a type above to start.