Reading an RSI news item with criteria
Separate forecast from fact, understand the METR curve and the task horizon, and know when measurement breaks down.
The problem
Every month someone announces the AI that improves itself.
Without a criterion, it is hard to separate what was measured from what is a promise. Each item below has a course that deals with it.
- A forecast for 2028 arrives as if it were fact, and you don't know how to check it.
- Revising the answer, changing the agent, and training the model become the same thing in the conversation.
- A higher score on a test hides a worse task outside the sample.
- The change that looked good goes into use and nobody knows how to roll it back.
- The system proposes, measures, and approves its own change, with no outsider checking it.
After the training
By the end of the 3 courses, you can:
Telling the three levels of improvement apart
Revising the answer, improving the system, and investigating recursion, each with what changes and what needs evidence.
Designing a cycle before automating
Separate who proposes, who measures, and who approves, and turn approval into reusable knowledge.
Building an evaluation that doesn't fool itself
Held-out cases the candidate hasn't seen, checking of generated data, and the rule of one change at a time.
Applying LOOP-R in your work
Run, measure, critique, propose, test, validate, and promote, with version tracking, a cost ceiling, and a command to revert.
3 courses in order · 39 lessons + 6 modules
Three courses, from the alert to the applied method.
Alerta IA 2028 gives the context. RSI v6.2 teaches how to design and evaluate a cycle. The LOOP-R course takes the method to your company, for owners and managers without a technical background.
- 1
Context
Alerta IA 2028
Three tracks and six modules: the AI loop that builds AI, the METR curve, MirrorCode, when measurement breaks down, and the 2028 scenarios as hypotheses.
- 2
Fundamentals and practice
RSI v6.2
Six modules from what changed to decision with evidence: a designed cycle, the AlphaEvolve and Darwin Gödel Machine cases, held-out evaluation, applications, and a pilot with a budget and a stop.
18 lessonsOpen course 2 - 3
Method
LOOP-R Course
Five tracks: why your company doesn't improve, the method, the loop assistants, measuring for real, and from task to company.
21 lessonsOpen course 3
Who it's for
Made for you if
- You read about AI that improves itself and want to separate what was measured from what is a promise.
- You are an owner or manager and want a process that improves through testing, with no technical background.
- You already use agents and want to compare versions with held-out tasks before adopting a change.
- You teach or write about AI and need original sources and clear limits.
Not for you yet if
- You want to build in the terminal the technical loop that works for hours and checks itself: see the LOOP-R Engenharia area.
- You are looking for a ready-to-run autonomous RSI system: the project is research and educational content.
Included in the training
Research with sources, projects you can open today.
The program starts from 15 sources read in the original, such as Self-Refine, Reflexion, Darwin Gödel Machine and AlphaEvolve. The RSI Guide, the RSI Copilot, Dream-RSI and the LOOP-R framework come along, with the videos on the topic.
Projects and kits
Videos
Why this training
What changes compared to following headlines.
A forecast doesn't become fact
The 2028 scenarios appear as hypotheses to examine, with no guaranteed timeline.
Whoever does doesn't approve
Proposal, measurement, and approval have separate roles. No worse version replaces the current one by the system's own decision.
Small, real tasks
Customer service without inventing policy, questions the text supports, and notes that become actions without filling in gaps.
Three languages
The three courses and the projects are in Portuguese, English and Spanish.
Questions
Before you start.
Do I need to follow the order?
The suggested order is Alerta IA 2028 → RSI v6.2 → LOOP-R. Those who already follow the topic can start with RSI v6.2.
Do I need a technical background?
No. The LOOP-R course is for owners and managers without a technical background, and RSI v6.2 works with tasks like answering based on a policy or extracting actions from notes.
Which tools will I use?
To understand it, all you need is a browser. To apply it, the LOOP-R framework runs in Claude Code with nine assistants, each with a separate role.
How big is the training?
RSI v6.2 has 18 lessons in six modules, the LOOP-R course has 21 lessons in five tracks, and AI Alert 2028 has six modules in three tracks.
What is RSI?
Recursive self-improvement: when an improvement starts helping to produce the next improvements. More attempts or a higher score do not prove unlimited growth.
Understand the cycle before you trust it.
Start with AI Alert 2028 and move on to RSI v6.2. In 18 lessons you design, evaluate, and decide on an improvement cycle backed by evidence.