AI is already helping build the next AI.
From a frightening headline to a critical reading of the warning for 2028—in plain English.
Frontier leaders — Anthropic, DeepMind and OpenAI — have put 2028 on the table as the horizon when AI could accelerate its own research. This course distills that warning and separates solid evidence from hype —so you can follow developments with informed attention, not panic.
No sign-up · opens directly in your browser · your progress, questions, and notes stay on your device.
From headline-watcher to critical reader of the warning
Most people encounter this subject through a headline — "lab confirms AI self-improving by 2028" —and feel only alarm. This course does the opposite: it gives you the tools to assess the warning for yourself.
You stop asking "Is this the end?" and start asking "What here is solid, what is speculation, and which signals should I watch?"
You do not need to code or be a researcher. You need to understand the loop, know the evidence on the table, and recognize where AI measurement itself starts to fail. In the end, panic becomes informed attention.
Without critical reading, the warning becomes noise
- ✗Reacting only to a scary headline leads to anxiety.
- ✗"It is all hype" becomes denial of what is already real.
- ✗Repeating isolated numbers — “60%,” “56%,” “19 days” — becomes rumor.
- ✗Confusing RSI with the singularity, consciousness, or robots leads to confusion.
With critical reading, the warning becomes informed attention.
Six ideas unlock the whole warning
You do not need to code or be a researcher—you need to understand what is happening and separate signal from noise. These six ideas are the course’s backbone.
The RSI loop
AI designs a better AI, which designs the next one—faster with each cycle. The warning’s central idea, without jargon.
Why code comes first
In software, the write → run → fix cycle closes in seconds. That is why the leap appears in code before anywhere else.
The METR curve
The “task horizon”: the longest task, measured in human time, that a model can complete alone with about 50% success.
MirrorCode
Rebuild a real program in the dark, using only its executable—the benchmark that tests how far AI can go on its own.
When measurement breaks
Cheating, sandbagging, and why detecting a failure can be a good sign: monitoring is working.
2028 scenarios
Not a date for the singularity: a range of possible futures and a checklist of concrete signals to watch.
The complete course
🔁 RSI fundamentals
The warning’s central idea in plain English: what recursive self-improvement is, why the limit shifts, and why the leap appears first in code.
- ·The loop: AI building AI
- ·Why it starts with code
📊 The evidence
The concrete evidence: METR’s task-horizon curve and the MirrorCode benchmark—what they measure and, honestly, what they do NOT prove.
- ·The METR curve: the task horizon
- ·MirrorCode: rebuilding in the dark
🛡️ Safety and 2028
Where measurement breaks (cheating, sandbagging), what each lab is saying, and how to read 2028 scenarios without panicking.
- ·When measurement breaks
- ·The race and 2028 scenarios
Solid evidence vs. speculation
The 2028 warning mixes real trends with numbers from a single video or article. This course teaches you to distinguish them—and marks every figure or model version repeated from the video as (according to the video—unverified).
✓ More solid
Categories and trends supported by broad evidence, not dependent on a single source.
- ✓The task-horizon trend from METR (the curve rises over time).
- ✓The MirrorCode concept: rebuilding software in the dark as a measure of what AI can do alone.
- ✓The measurement instability when a model cheats (cheating / reward hacking).
- ✓Anthropic’s public statement that most of its code is written by Claude.
- ✓RSI as a serious topic raised by Clark (Anthropic), Hassabis (DeepMind), and Hinton.
⚠ Treat as claims to verify
Exact numbers and names from one video or article—useful as illustrations, not established facts.
- ⚠Exact model names and versions ("Opus 4.7", "GPT-5.6 Sol", "Claude 3 Opus 52×").
- ⚠The "~60% by 2028" —one person’s estimate, not a field-wide consensus.
- ⚠Exact figures repeated from the video: 56%, 19 days, US$2,600, US$200M.
- ⚠Any claim that the curve "proves" a predetermined outcome—extrapolating a straight line is no guarantee.
Course rule: whenever we cite a model version or figure from the video, we mark it (according to the video—unverified).
Informed attention, not panic.
The 2028 warning is not a prophecy—it is an invitation to follow developments critically.
Understand the loop, examine the evidence, learn where measurement fails—and you will be able to separate signal from noise on your own, without waiting for the next headline.