Skip to content

Update · A runaway answer now stops at 768 MiB in every lesson

Choose a tag to compare

@TarrySingh TarrySingh released this 08 Nov 12:01
· 1 commit to master since this release

An update to Synapsa Commons: free, hands-on AI courses that run anywhere. No new lessons this time. This release makes every one of the 103 lessons safer to run on small machines.

What changed

  • A runaway answer now stops at 768 MiB, down from 2 GiB, in every lesson. If an answer keeps allocating memory, the check it belongs to stops it, marks that row as failed and the notebook carries on. On a 2 GB machine of the kind Binder gives you, an answer that grows without end now stops at about 885 MiB, where it used to reach 2.1 GB and could take the kernel down with it.
  • The memory watch checks more often. It now catches fast-growing answers sooner, and if an answer's own error handling swallows the stop, the watch stops it again. In our measurements this made the lessons no slower.
  • Three fixes in the grader. Memory is now freed after a row that does not pass, so one bad answer can no longer use up the memory of the rows after it. A score now rounds down, never up. The whole-tree check now reads each lesson's real point totals.
  • One source is corrected. In the lesson on living on free tiers, the OpenRouter quote no longer copies a placeholder that had not been filled in on the page. The passage was re-read on 4 November 2026.
  • The lessons themselves are unchanged. Explanations, exercises, data and difficulty are exactly as before. Only the safety net around the checks changed.

Checked before release

  • Today all 103 lessons passed every quality gate on macOS. Every worked solution scores 100%. Every blank notebook scores 0%, except five C and C++ lessons, which score 5 to 9% because their "it compiles" rows pass.
  • On 7 and 8 November the same lesson code was checked in more depth:
    • All 103 lessons, run unanswered, print the same results on Python 3.11 and 3.12.
    • About 1,900 deliberately wrong answers were graded. None scored full marks, and no correct answer lost a point.
    • The largest amount of memory any correct check needed was 990 MiB, in a row that is allowed 1536 MiB.
    • Five lessons were checked on Linux, on an emulated x86 machine, and three heavy ones inside a 2 GB container.
  • Limits still apply. On native Windows there is no time or memory limit, so use WSL, Colab or Codespaces there. One long native call, such as a very large numpy operation, is stopped only when it returns.
  • Worked solutions and full grading come with enrolment on Synapsa, opening in the coming weeks.