Wink Pings

Andrew Ng compresses AI engineering into one hour—five timestamps worth pausing for

Free one-hour course, from vibe coding to coding agents, evals, and judgment. Timestamps, key numbers, plus a breakdown of a long article on Fable 5's self-improving system.

Andrew Ng posted a free one-hour course on X, and the timestamps are already cut. The course breaks the 2026 AI engineering path from vibe coding to coding agents, evals, and judgment into five segments. Original post here: https://x.com/0xMovez/status/2122694058845839769 .

Five timestamps

  1. 10%, 8:30: Why engineers who only write specs are hard to hire.
  2. 30%, 13:30: Four AI skills distilled from 20,000 job posts.
  3. 50%, 14:10: What trade-offs a coding agent hides.
  4. 75%, 19:40: The 80% plateau, where evals start taking over.
  5. 100%, 29:40: Five skills AI cannot automate.

Under the tweet, someone worries that junior engineer hiring is shrinking, and Movez replies, “90% of juniors were never needed in the first place.” That judgment is blunt, and the 8:30 segment is exactly about why engineers who only write specs are hard to hire. Another reply mentions that no one will cut “shaping the build”: engineers no longer just receive specs; they have to write the specs themselves. There is also a concrete experience about evals: ask for failing tests first. A bill-splitting function produced a rounding bug once in three tries with a normal prompt; with tests first, it passed all three times, at about 2 extra cents each time. That detail is more useful than empty talk about evals.

The harder part in the referenced article

In the same quote chain, Codez posted a 14-step long article about Claude Fable 5, on building a self-improving agent system that accumulates. Article link: http://x.com/i/article/2065077530571264000 . Core facts: Fable 5 is the first publicly released Mythos-class model from Anthropic, released on June 9, 2026, supporting multi-day autonomous sessions, sub-agent delegation, and vision-based self-checking of outputs; pricing is $10 per million input tokens and $50 per million output tokens. The article repeatedly emphasizes one thing: self-improving and self-learning are two different things. The model weights do not change; what changes is the system built around the model.

The four-layer stack is divided like this:

  1. Primitives: model, sub-agents, worktree, tools.
  2. Orchestration: /goal and Outcomes run a self-correction loop, Dynamic Workflows orchestrate multi-step tasks, and Routines put runs in the cloud.
  3. Memory: STATE.md, Skills, and a knowledge base, writing down the lessons from every session.
  4. Self-improvement: vision self-checks, eval loops, and rule distillation, writing results back into memory and Skills.

There are three key numbers. In the Parameter Golf experiment, Fable 5 plus an independent verifier gained about 6x more improvement than Opus 4.7. On Continual Learning Bench, Fable 5 reached 73% verification coverage, versus 17% for Opus 4.7. The article breaks memory into five stages: Fail, Investigate, Verify, Distill, Consult. STATE.md corresponds to project memory, and Skills corresponds to cross-project process memory. At the end of each session, write STATE.md first; at the start of the next, read it first. Without this step, Fable 5’s memory advantage falls back to Sonnet level.

The safety boundary is also stated plainly. Fable 5 ships with a classifier that refuses requests in high-risk domains such as cybersecurity, biology, chemistry, and model distillation, and automatically falls back to Opus 4.8. The article reminds readers to treat the fallback as a known path, rather than waiting for a loop to silently fail before checking. In another discussion thread, someone posted a purported Mythos system prompt; the Gist shows Claude product information, tool definitions, storage APIs, and search instructions. Its authenticity cannot be independently verified, so do not treat it as official documentation.

There is also a community-curated awesome-agent-loops: https://github.com/serenakeyitan/awesome-agent-loops , which lists usable prompts for /loop, /goal, and /schedule by scenario. If you want to try it, start with “write the failing test first” or give the agent an independent verifier; that is cheaper than switching to a stronger model.

发布时间: 2026-11-18 04:31