Subjects › AGI Agents (6-week course)
W6 · Honest scope & what AGI means
Placing a system on the Levels of AGI, measuring the alignment tax, and saying precisely what it can and cannot do.
The idea
Nobody has built AGI; this course builds the architecture around general-purpose models. 'Levels of AGI' (Morris et al., read in Week 1) separates performance (emerging → superhuman) from generality (narrow vs general), and the Week 6 reflection asks you to place your system on that matrix using eval evidence only. The alignment tax is what a safety layer costs in success rate and LLM calls; Experiment C measures it with the constitutional review on vs off, and whether the review catches anything real is itself an empirical question. A good capstone report backs every capability claim with an eval number (the rubric deducts 10 points otherwise) and states its limitations and residual risks plainly. In your repo: Week06_Integration_Capstone/final_report_template.md, capstone_rubric.md, project_tracks.md; TUTOR_NOTES.md §2 principle 10.
Lesson
No lesson built yet. The worked solutions below are the material it would be assembled from.
Questions (5)
| Question | Type | Difficulty | Drift | Accuracy |
|---|---|---|---|---|
| Which sentence describes the finished AGI_Prototype honestly? | mcq | Easy -1.30 | — | — (0) |
| Week 6 Experiment C (alignment tax). On a 30-task suite your agent succeeds on 24 tasks with the constitutiona… | numeric | Easy -0.60 | — | — (0) |
| Week 6 Experiment C (alignment tax), 30-task suite. Switching the constitutional review on lowers success by a… | mcq | Hard 0.50 | — | — (0) |
| The Levels of AGI framework rates a system on two separate axes: performance (emerging → superhuman) and gener… | mcq | Moderate 0.20 | — | — (0) |
| A classmate's report says: 'Our agent has memory, planning, tools and self-critique, so it is an early AGI.' I… | short | Very hard 1.50 | — | — (0) |