Skip to content

progress.md

Latest commit

 

History

History
339 lines (250 loc) · 48.3 KB

progress.md

File metadata and controls

339 lines (250 loc) · 48.3 KB
Raw

Progress Log

2026-05-04

  • Started update batch after user approval.
  • Loaded planning, YouTube PDF, Bilibili PDF, and verification skills.
  • Verified local tooling availability.
  • Counted current notes: 290 PDFs and 289 TeX files before this batch.
  • Confirmed Bilibili new videos have public metadata but no platform subtitles.
  • Marked Bilibili note generation blocked due to B站 412 plus missing Whisper.
  • Updated README and TRACKING with current inventory and active-course backlog.
  • Downloaded CS336 Spring 2026 official materials for lecture05-10 into cs336-2026/.
  • Verified downloaded CS336 files by file and wc -c.
  • Final verification: 290 *-notes.pdf files, no stale README/TRACKING strings found, git diff --check clean, and CS336 downloaded files have valid file types.
  • Created NOTE_GENERATION_TODO.md to track pending note generation.
  • Confirmed 4 A100 80GB GPUs are visible via /proc/driver/nvidia/gpus.
  • Confirmed faster-whisper and cached large-v3 model exist, but YouTube and Bilibili downloads currently require cookies.
  • Downloaded Modern Agent 17 via public Bilibili playurl API, extracted audio, transcribed on A100 with faster-whisper large-v3, generated key frames, wrote TeX, and compiled lecture17-notes.pdf.
  • Attempted Agentic RL 16 via public Bilibili playurl API; only preview-length media was accessible, so no incomplete PDF was generated.
  • Verification after Modern Agent 17: note count is 291, generated PDF/TeX/SRT exist, no stale README/TODO strings found, and git diff --check is clean.
  • Switched focus to CS336 at user request.
  • Generated cs336-2026/lecture01/lecture01-notes.tex from official Spring 2026 lecture01-slides.py.
  • Compiled initial cs336-2026/lecture01/lecture01-notes.pdf draft with XeLaTeX; output was 9 pages and was later superseded by the quality rewrite.
  • Updated README/TRACKING/TODO after CS336 lecture01; note count temporarily became 292 before lecture02 was added.
  • User clarified that CS336 notes must meet the repository depth standard rather than a thin summary.
  • Rewrote cs336-2026/lecture01/lecture01-notes.tex into a deeper 964-line course note, removed manual TikZ figures, and added official image assets under cs336-2026/lecture01/images/.
  • Recompiled cs336-2026/lecture01/lecture01-notes.pdf twice with XeLaTeX; current PDF is 27 pages by mutool, with 11 images including cover, 22 teaching boxes, 13 section-level summaries, and 9 code listings.
  • Patched tools/scripts/check_quality.sh to fall back to mutool when pdfinfo is unavailable and to count actual box environments.
  • Final verification: tools/scripts/check_quality.sh cs336-2026/lecture01/lecture01-notes.tex later reports 28p ⭐⭐⭐, and git diff --check passed.
  • Continued CS336 Spring 2026 with lecture02 at the same depth standard.
  • Downloaded lecture02 figure assets, wrote cs336-2026/lecture02/lecture02-notes.tex, and compiled lecture02-notes.pdf twice with XeLaTeX.
  • Lecture02 verification: tools/scripts/check_quality.sh cs336-2026/lecture02/lecture02-notes.tex later reports 26p ⭐⭐⭐; note count became 293 PDFs.
  • Resumed CS336 Spring 2026 at user request with stricter quality bar: deep exposition, careful technical detail, and one verified lecture at a time.
  • Started cs336-2026/lecture03/, using official lecture03 slides plus the existing CS336 lecture03 note as a cross-check, not as a replacement for fresh writing.
  • Wrote cs336-2026/lecture03/lecture03-notes.tex as a fresh Spring 2026 architecture/hyperparameters note with conservative recipe tables, formulas, code snippets, official slide figures, and explicit engineering judgment sections.
  • Compiled cs336-2026/lecture03/lecture03-notes.pdf with XeLaTeX; quality verification reports 35p 11s 24b 33f ⭐⭐⭐, and repository note count is now 294 PDFs.
  • Continued to cs336-2026/lecture04/ at user request.
  • Confirmed Spring 2026 lecture04 is broader than the old CS336 lecture04: it covers attention alternatives (linear attention, Mamba-2, Gated Delta Net, sparse adaptation/DSA) before mixture-of-experts.
  • Wrote cs336-2026/lecture04/lecture04-notes.tex as a fresh Spring 2026 long-form note tying attention alternatives and MoE together through sparse/conditional computation.
  • Compiled cs336-2026/lecture04/lecture04-notes.pdf with XeLaTeX; quality verification reports 38p 14s 21b 41f ⭐⭐⭐, and repository note count is now 295 PDFs.
  • Continued to cs336-2026/lecture05/ at user request.
  • Confirmed Spring 2026 lecture05 is a 55-slide GPU performance lecture: GPU/TPU architecture, memory hierarchy, low precision, fusion, recomputation, memory coalescing, tiling, matrix-performance anomalies, and FlashAttention.
  • Generated 55 slide images under cs336-2026/lecture05/slides-images/.
  • Wrote cs336-2026/lecture05/lecture05-notes.tex as a fresh Spring 2026 long-form GPU performance note with matmul arithmetic-intensity derivation, low-precision tradeoffs, attention IO accounting, and FlashAttention design checklist.
  • Compiled cs336-2026/lecture05/lecture05-notes.pdf with XeLaTeX; quality verification reports 31p 9s 21b 36f ⭐⭐⭐, and repository note count is now 296 PDFs.
  • Continued to cs336-2026/lecture06/ at user request.
  • Confirmed Spring 2026 lecture06 is an executable-source lecture on benchmarking, profiling, kernel fusion, torch.compile, and Triton kernels for GeLU, softmax, row sum, and matmul+ReLU.
  • Downloaded/copied lecture06 visual assets into cs336-2026/lecture06/images/, including official referenced images and selected L05 GPU/roofline/tiling recap figures.
  • Wrote cs336-2026/lecture06/lecture06-notes.tex as a fresh Spring 2026 long-form Kernels/Triton note with detailed GPU programming-model review, benchmarking/profiling methodology, fusion memory accounting, and Triton kernel walkthroughs for GeLU, softmax, row sum, and matmul+ReLU.
  • Compiled cs336-2026/lecture06/lecture06-notes.pdf twice with XeLaTeX; single-note quality verification reports 31p 13s 31b 12f ⭐⭐⭐, and repository note count is now 297 PDFs.
  • Final L06 verification: tools/scripts/check_quality.sh cs336-2026 reports lecture01-06 all ⭐⭐⭐, find . -name '*-notes.pdf' | wc -l reports 297, git diff --check passed, and the L06 LaTeX log scan found no LaTeX errors, undefined control sequences, rerun requests, or overfull boxes.
  • Continued to cs336-2026/lecture07/ at user request.
  • Confirmed Spring 2026 lecture07 is an executable-source lecture on multi-GPU parallelism: collectives, interconnect hardware, NCCL/PyTorch distributed, communication benchmark methodology, and bare-bones data/tensor/pipeline parallel MLPs.
  • Downloaded lecture07 visual assets into cs336-2026/lecture07/images/, including official CS336 diagrams for ranks/interconnect/data/tensor/pipeline parallelism and PyTorch/NCCL diagrams for broadcast, scatter, gather, reduce, all-gather, all-reduce, and reduce-scatter.
  • Wrote cs336-2026/lecture07/lecture07-notes.tex as a fresh Spring 2026 long-form Parallelism note with detailed sections on collectives, hardware topology, NCCL/PyTorch distributed, communication benchmarking, and data/tensor/pipeline parallel MLP implementations.
  • Compiled cs336-2026/lecture07/lecture07-notes.pdf twice with XeLaTeX; single-note quality verification reports 25p 12s 30b 14f ⭐⭐⭐, and canonical repository source-note count excluding .web-build is now 298 PDFs.
  • Final L07 verification: tools/scripts/check_quality.sh cs336-2026 reports lecture01-07 all ⭐⭐⭐, canonical source-note count excluding .web-build reports 298, git diff --check passed, and the L07 LaTeX log scan found no LaTeX errors, undefined control sequences, rerun requests, or overfull boxes.
  • Continued to cs336-2026/lecture08/ at user request.
  • Confirmed Spring 2026 lecture08 is a 73-page official PDF deck titled "Parallelism Basics".
  • Attempted to install missing pdftotext support via sudo apt-get install poppler-utils; host requires an interactive sudo password, so the install was blocked.
  • Used existing mutool to extract lecture08 per-page text into /tmp/lecture08-text-*.txt and render 73 slide images under cs336-2026/lecture08/slides-images/.
  • Wrote cs336-2026/lecture08/lecture08-notes.tex as a fresh Spring 2026 long-form Parallelism Basics note with detailed sections on networking, collective communication, ZeRO/FSDP, pipeline/tensor/sequence/expert/context parallelism, 3D/4D recipes, and recent model configurations.
  • Compiled cs336-2026/lecture08/lecture08-notes.pdf twice with XeLaTeX; single-note quality verification reports 41p 14s 23b 51f ⭐⭐⭐, and canonical repository source-note count excluding .web-build is now 299 PDFs.
  • Final L08 verification: tools/scripts/check_quality.sh cs336-2026 reports lecture01-08 all ⭐⭐⭐, canonical source-note count excluding .web-build reports 299, git diff --check passed, and the L08 LaTeX log scan found no LaTeX errors, undefined control sequences, rerun requests, or overfull boxes.
  • Installed poppler-utils after the user provided the sudo password for this session; pdftotext, pdfinfo, and pdfimages are now available.
  • Continued to cs336-2026/lecture09/ at user request.
  • Confirmed Spring 2026 lecture09 is a 57-page official PDF deck titled "Scaling Laws - Basics".
  • Used pdfinfo/pdftotext to inspect/extract L09 text and rendered 57 slide images under cs336-2026/lecture09/slides-images/.
  • Wrote cs336-2026/lecture09/lecture09-notes.tex as a fresh Spring 2026 long-form Scaling Laws Basics note with detailed sections on data scaling, data mixture/repetition, model-engineering scaling laws, critical batch size, muP, joint data-model scaling, Kaplan vs Chinchilla, IsoFLOPS, and deployment-aware overtraining.
  • Compiled cs336-2026/lecture09/lecture09-notes.pdf with XeLaTeX until cross-reference warnings cleared; single-note quality verification reports 36p 12s 23b 44f ⭐⭐⭐, and canonical repository source-note count excluding .web-build is now 300 PDFs.
  • Final L09 verification: tools/scripts/check_quality.sh cs336-2026 reports lecture01-09 all ⭐⭐⭐, canonical source-note count excluding .web-build reports 300, git diff --check passed, and the L09 LaTeX log scan found no LaTeX errors, undefined control sequences, rerun requests, or overfull boxes.
  • Continued to cs336-2026/lecture10/ at user request.
  • Confirmed Spring 2026 lecture10 is an executable-source lecture titled "inference", covering inference metrics, arithmetic intensity, KV cache, prefill/generation, latency-throughput tradeoffs, KV compression, quantization, pruning/distillation, speculative sampling, continuous batching, and PagedAttention.
  • Downloaded lecture10 visual assets into cs336-2026/lecture10/images/, including official CS336 figures, Scaling Book Transformer/inference/GQA diagrams, a continuous batching diagram, and a quantization diagram; converted WebP inference diagrams to PNG for local TeX compilation.
  • Wrote cs336-2026/lecture10/lecture10-notes.tex as a fresh Spring 2026 long-form Inference note with detailed derivations, 29 figures, 41 teaching boxes, and code snippets for KV-cache decoding, latency-throughput modeling, speculative sampling, continuous batching, and PagedAttention block tables.
  • Compiled cs336-2026/lecture10/lecture10-notes.pdf with XeLaTeX until cross-reference warnings cleared; single-note quality verification reports 30p 13s 41b 29f ⭐⭐⭐, and canonical repository source-note count excluding .web-build is now 301 PDFs.
  • Final L10 verification: tools/scripts/check_quality.sh cs336-2026 reports lecture01-10 all ⭐⭐⭐, canonical source-note count excluding .web-build reports 301, git diff --check passed, and the L10 LaTeX log scan found no LaTeX errors, undefined control sequences, rerun requests, or overfull boxes.
  • User reviewed lecture01 and tightened the writing standard: important figures need explicit read-the-figure interpretation, concept-dense passages need concentrated explanations, foundational background concepts should use diagrams/tables/formulas, and available slides should be covered comprehensively.
  • Updated AGENTS.md, QUALITY.md, CLAUDE.md, and tools/skills/video-render-common/writing-and-figures.md to capture these new standards.
  • Started regenerating CS336 Spring 2026 lecture01 and lecture02 from official executable sources under the new standards, rather than continuing to patch the existing notes.
  • Downloaded additional local image assets for regenerated lecture01/02, including industrialization, DGX B200 topology, Marin scaling forecast/result, The Pile chart, FP8, and roofline images, so the PDFs do not depend on remote image URLs.
  • Replaced cs336-2026/lecture01/lecture01-notes.tex and cs336-2026/lecture02/lecture02-notes.tex with fresh new-standard drafts, then compiled both PDFs with XeLaTeX until rerun warnings cleared.
  • Final lecture01/02 regeneration verification: lecture01 reports 20p 7s 32b 15f ⭐⭐⭐, lecture02 reports 20p 9s 38b 11f ⭐⭐⭐, tools/scripts/check_quality.sh cs336-2026 reports all ten CS336 2026 notes as ⭐⭐⭐, git diff --check passed, and both lecture logs have no LaTeX errors, undefined control sequences, rerun requests, or overfull boxes.
  • Added mandatory PDF visual QA workflow to shared note-generation skills and project standards, including page rendering, contact sheet inspection, full-size suspicious-page review, and required TeX fixes before completion.
  • Installed missing project tooling/dependencies: latexmk, system ripgrep, fd-find, current-venv Python packages yt-dlp, openai-whisper, torch CPU, pypdf, pdfplumber, pymupdf, pandas, matplotlib, beautifulsoup4, and lxml; verified imports and tests/test_generate_site.py pass.
  • Added workflow tooling: build_lecture_manifest.py, check_note_coverage.py, and render_pdf_qa.py, plus docs/NOTE_GENERATION_WORKFLOW.md; smoke-tested manifest, coverage, visual QA, and source-only check_quality.sh . behavior on CS336 lecture01/02.
  • Extended workflow standards and check_note_coverage.py to require first-use explanations for dense systems terms including ZeRO, sharding, fused kernels, collectives, optimizer state, activation checkpointing, DRAM/SRAM/HBM, and perplexity; smoke test flags old CS336 lecture01 for these exact gaps.
  • Completed CS336 2026 lecture01 and lecture02 new-workflow artifacts: manifests, coverage matrices, blueprints, PDF visual QA contact sheets, and checked QA reports.
  • Regenerated CS336 2026 lecture03 under the new workflow as a 44-page slide-complete note with all 67 source slides included and 68 figures total; quality script reports 44p 11s 27b 68f ⭐⭐⭐.
  • Improved render_pdf_qa.py after lecture03 visual QA found near-blank pages from pdftoppm; the script now detects near-blank rendered pages and falls back/checks rendered output.
  • Regenerated CS336 2026 lecture04 under the new source-first workflow as a 41-page slide-complete note with all 60 source slides included, attention/MoE glossary coverage, and read-the-figure explanations; quality script reports 41p 12s 42b 61f ⭐⭐⭐, LaTeX log scan is clean, and visual QA contact sheet was reviewed.
  • Regenerated CS336 2026 lecture05 under the new source-first workflow as a 38-page slide-complete GPU note with all 55 source slides included, 56 figures, 44 teaching boxes, and read-the-figure explanations; quality script reports 38p 8s 44b 56f ⭐⭐⭐, LaTeX log scan is clean, and visual QA contact sheet was reviewed.
  • Regenerated CS336 2026 lecture06 under the new source-first workflow as a 20-page source-node-complete kernels/Triton note with 8 figures, 36 teaching boxes, and detailed code walkthroughs for benchmarking, profiling, GeLU fusion, Triton GeLU, softmax, row-sum, and tiled matmul+ReLU; quality script reports 20p 15s 36b 8f ⭐⭐⭐, LaTeX log scan is clean, and visual QA contact sheet was reviewed.
  • Re-verified CS336 2026 lecture04 at the user's request before continuing serially: two XeLaTeX passes produced a 41-page PDF, check_quality.sh reports 41p ⭐⭐⭐, check_note_coverage.py reports figs=61 readfig=105 boxes=42 term_digest=3 formulas=2 code=0 summaries=10 with no hard errors, all 60 slide images are referenced, log scan is clean for hard LaTeX issues, and the refreshed PDF QA contact sheet/report were reviewed and marked complete.
  • Regenerated CS336 2026 lecture07 under the new source-first workflow from the official executable lecture source: the note is now 26 pages with 15 figures, 18 read-the-figure explanations, 42 teaching boxes, 8 terminology-digestion hits, 7 formulas, and 7 code listings. Final verification reports 26p ⭐⭐⭐; coverage has no hard errors; all local images are referenced; LaTeX log scan is clean; visual PDF QA contact sheet was reviewed and checked.
  • Regenerated CS336 2026 lecture08 under the new source-first workflow from the 73-page official slide deck: the note is now slide-complete with all 73 slide images included, 53 pages, 74 figure inclusions, 59 read-the-figure/table/formula explanations, 76 teaching boxes, and explicit terminology digestion for collectives, sharding, ZeRO/FSDP, TP/SP/EP/CP, and pipeline concepts. Final verification reports 53p ⭐⭐⭐, coverage has no hard errors, LaTeX log scan is clean, and visual PDF QA contact sheet was reviewed and checked.
  • Regenerated CS336 2026 lecture09 under the new source-first workflow from the 57-page official slide deck: the note is now slide-complete with all 57 slide images included, 42 pages, 58 figure inclusions, 53 read-the-figure/formula/table explanations, 59 teaching boxes, 5 formulas, and 3 code listings. Final verification reports 42p ⭐⭐⭐, coverage has no hard errors, LaTeX log scan is clean, and visual PDF QA contact sheet was reviewed and checked.
  • Regenerated CS336 2026 lecture10 under the new source-first workflow from the official executable lecture source: the note is now source-node-complete with all local teaching PNGs included, 21 pages, 29 figure inclusions, 23 read-the-figure explanations, 34 teaching boxes, 4 terminology-digestion hits, and formulas for arithmetic intensity/KV cache/speculative sampling. Final verification reports 21p ⭐⭐⭐, coverage has no hard errors or warnings, LaTeX log scan is clean, and visual PDF QA contact sheet was reviewed and checked.
  • Downloaded newly available official CS336 2026 materials for lecture11-13 from the Stanford lectures repository: lecture11 PDF deck (58 pages), lecture12 executable source, and lecture13 executable source.
  • Generated CS336 2026 lecture11 under the new source-first workflow from the 58-page official slide deck: the note is slide-complete with all 58 slide images included, 43 pages, 59 figure inclusions, 53 read-the-figure/formula explanations, 60 teaching boxes, and terminology digestion for WSD, muP, optimizer scaling, Muon, Chinchilla methods, and recent scaling recipes. Final verification reports 43p ⭐⭐⭐, coverage has no hard errors, LaTeX log scan is clean, and visual PDF QA contact sheet was reviewed and checked.
  • Generated CS336 2026 lecture12 under the new source-first workflow from the official executable lecture source: localized all 42 referenced image assets, wrote a 29-page evaluation note with 44 figure inclusions, 25 read-the-figure explanations, 40 teaching boxes, and terminology digestion for evaluation design, perplexity, zero-shot, methods/models/agents, ELO/LLM judges, ecological validity, contamination, and benchmark quality. Final verification reports 29p ⭐⭐⭐, coverage has no hard errors or warnings, LaTeX log scan is clean, and visual PDF QA contact sheet was reviewed and checked.
  • Generated CS336 2026 lecture13 under the new source-first workflow from the official executable lecture source: localized all 18 referenced image assets, wrote a 20-page Data I note with 19 figure inclusions, 13 read-the-figure explanations, 46 teaching boxes, and terminology digestion for crawlers, WARC/WET, robots.txt/ToS, copyright/fair use/licensing, Common Crawl, filtering, deduplication, model-based filtering, provenance, and data audit loops. Final verification reports 20p ⭐⭐⭐, coverage has no hard errors or warnings, LaTeX log scan is clean, and visual PDF QA contact sheet was reviewed and checked.

2026-05-12 YouTube ttkd0t5qTD4

  • Started user-requested YouTube note task for ttkd0t5qTD4.
  • Loaded youtube-render-pdf and planning-with-files workflows.
  • Verified local yt-dlp, ffmpeg, xelatex, and pdfinfo availability.
  • Anonymous YouTube metadata/format extraction failed because YouTube requires sign-in/bot confirmation. Will try local browser-cookie based extraction next.

2026-05-12 YouTube auth work

  • User asked to solve the repeated YouTube login challenge fundamentally before continuing note generation.

  • Confirmed yt-dlp is current (2026.03.17) and configured only with --js-runtimes node.

  • Confirmed no ordinary browser profiles exist for --cookies-from-browser; found cached Playwright Chromium, which can be used to create a reusable authenticated profile.

  • Installed bgutil-ytdlp-pot-provider==1.3.1.

  • Installed then corrected curl_cffi to 0.14.0; yt-dlp --list-impersonate-targets now shows Chrome/Safari/Firefox targets via curl_cffi.

  • Verified plugin loading: yt-dlp reports PO Token Providers: bgutil:http-1.3.1.

  • Remaining auth issue at this point: local bgutil HTTP server is not yet running, so tokens cannot be generated.

  • User completed Google device-code OAuth authorization. Token cache was saved successfully.

  • Continuing diagnosis because extraction now fails at a post-auth YouTube API HTTP 400, not at the earlier sign-in challenge.

  • Added tools/scripts/youtube_auth_check.sh and expanded .gitignore for YouTube cookie files.

  • OAuth was authorized but proved insufficient for actual media formats on this host; cookie-based YouTube account auth is now the required next input.

  • Authenticated YouTube format listing succeeded after adding --remote-components ejs:github; real audio/video formats became available.

  • Downloaded and merged the original 4K video to youtube/ttkd0t5qTD4/original.mkv; downloaded subtitle SRT files, thumbnail, and info JSON.

  • Moving to transcript cleaning, frame extraction, TeX writing, PDF compilation, and visual QA.

  • Completed ttkd0t5qTD4 note generation: wrote TeX, compiled PDF, ran quality, coverage, log scan, diff whitespace check, and visual PDF QA.

2026-05-12 Zhang Xiaojun YouTube batch

  • Started batch request for all Zhang Xiaojun YouTube interviews.

  • First action: enumerate channel videos and podcast playlists, then build a deduplicated generation queue.

  • Created Zhang Xiaojun YouTube batch queue files under youtube/zhangxiaojun/. Next canonical interview target is episode 138 (vG1RBqn1sG4).

  • Downloaded episode 138 work media (vG1RBqn1sG4) as 720p source.mp4; YouTube reported no subtitles for requested languages, so local transcription is required.

  • Stopped CPU transcription path and began GPU root-cause validation for faster-whisper/CTranslate2.

  • Fixed transcription workflow to use CTranslate2 CUDA via tools/scripts/transcribe_faster_whisper.py. Episode 138 large-v3 GPU transcription completed in about 756 seconds with 6416 segments.

  • GPU transcription root fix verified: episode 138 transcript now exists as transcript.zh.srt/txt/json from faster-whisper large-v3 CUDA.

2026-05-13 Zhang Xiaojun serial generation

  • User requested serial high-standard generation for the broad AI/internet queue, newest to oldest.

  • Treat EP140 as completed via youtube/ttkd0t5qTD4. Starting EP139 Xxz5uh0L1mE next. EP138 source/transcript work is preserved but paused to respect newest-to-oldest ordering.

  • Episode 139 media and GPU transcript completed: source.mp4, transcript.zh.srt/txt/json, and chapter transcript split are available under youtube/zhangxiaojun/ep139-Xxz5uh0L1mE/.

  • Completed EP139 Agent survey note: youtube/zhangxiaojun/ep139-Xxz5uh0L1mE/ep139-notes.pdf (20 pages, quality ⭐⭐⭐, coverage passes, visual QA checked).

  • Applied feedback: fixed no-slides podcast visual policy. EP139 now uses cover + concept diagrams only; repeated speaker frames were removed. Updated podcast workflow and quality script so interview notes are not incentivized to add repeated speaker frames for figure quota.

  • Continuing serial Zhang Xiaojun generation with EP138 Luo Fuli (vG1RBqn1sG4), using existing source media and faster-whisper large-v3 CUDA transcript.

  • Completed EP138 Luo Fuli note: youtube/zhangxiaojun/ep138-vG1RBqn1sG4/ep138-notes.pdf (20 pages, quality ⭐⭐⭐, coverage passes, visual QA checked; cover + concept diagrams only, no repeated speaker frames).

  • Continuing serial Zhang Xiaojun generation with EP137 Hong Letong (bv8ghyTFF9w).

2026-05-13 CS336 workflow audit after user quality feedback

  • User reported that CS336 notes have many images but still lack enough detailed prose explanation and transitions between small sections.

  • Audited the workflow and sample CS336 notes. New static metrics confirmed the issue: several figure-heavy notes have low prose-per-figure and weak subsection openers despite previously scoring ⭐⭐⭐.

  • Updated check_note_coverage.py to detect figure-heavy-prose-thin, thin-local-figure-explanations, and weak-section-openers.

  • Updated check_quality.sh to display prose-per-figure and demote figure-heavy notes below the prose-density threshold.

  • Updated AGENTS.md, QUALITY.md, docs/NOTE_GENERATION_WORKFLOW.md, and shared video-writing rules to require prose-led flow, section bridge paragraphs, pre-figure setup, post-figure synthesis, and narrative blueprint fields.

  • Added docs/CS336_WORKFLOW_AUDIT_2026-05-13.md with root cause, evidence, and follow-up repair plan.

  • Rewrote CS336 2026 lecture12 as the first prose-led/teacher-voice sample: expanded from 29p to 32p, improved prose-per-figure from 149 to 260, added 13 teacher-voice markers, strengthened transitions between evaluation families, and passed the stricter coverage checker with no warnings.

  • EP137 Hong Letong source media and faster-whisper large-v3 CUDA transcript completed; chapter transcript split is ready. Status set to paused-ready-transcript for next serial writing step.

  • Completed EP137 Hong Letong note: youtube/zhangxiaojun/ep137-bv8ghyTFF9w/ep137-notes.pdf (21 pages, quality ⭐⭐⭐, coverage hard checks pass, visual QA checked; cover + concept diagrams only, no repeated speaker frames).

  • Continuing serial Zhang Xiaojun generation with EP136 Guangmi LLM quarterly report (u1Lzp-7Ybn8).

  • EP136 Guangmi source media and faster-whisper large-v3 CUDA transcript completed; chapter transcript split is ready for survey note writing.

  • Completed EP136 Guangmi quarterly report note: youtube/zhangxiaojun/ep136-u1Lzp-7Ybn8/ep136-notes.pdf (20 pages, quality ⭐⭐⭐, coverage hard checks pass, visual QA checked; cover + concept diagrams only, no repeated speaker frames).

  • Continuing serial Zhang Xiaojun generation with EP135 Tristan/Natural Selection (x8qdqWIVVTA).

  • EP135 Tristan/Natural Selection source media and faster-whisper large-v3 CUDA transcript completed; chapter transcript split is ready.

  • Completed EP135 Tristan/Natural Selection note: youtube/zhangxiaojun/ep135-x8qdqWIVVTA/ep135-notes.pdf (21 pages, quality ⭐⭐⭐, coverage hard checks pass, visual QA checked; cover + concept diagrams only, no repeated speaker frames).

  • Continuing serial Zhang Xiaojun generation with EP134 Xie Chen data survey (owjTOT14bG0).

  • EP134 Xie Chen data survey source media and faster-whisper large-v3 CUDA transcript completed; chapter transcript split is ready.

  • Completed EP134 Xie Chen data survey note: youtube/zhangxiaojun/ep134-owjTOT14bG0/ep134-notes.pdf (20 pages, quality ⭐⭐⭐, coverage hard checks pass, visual QA checked; cover + concept diagrams only, no repeated speaker frames).

  • Continuing serial Zhang Xiaojun generation with EP133 Saining Xie marathon interview (iiBY0fqpThI).

  • EP133 Saining Xie source media and faster-whisper large-v3 CUDA transcript completed; chapter transcript split is ready. Status set to paused-ready-transcript for next serial writing step.

  • Completed EP133 Saining Xie marathon interview note: youtube/zhangxiaojun/ep133-iiBY0fqpThI/ep133-notes.pdf (23 pages, quality ⭐⭐⭐, coverage check passes with figs=9 readfig=10 boxes=30 term_digest=4 summaries=15, visual QA checked; cover + generated concept diagrams only, no repeated body speaker frames).

  • Continuing serial Zhang Xiaojun generation with EP132 Gao Jiyang / Xinghaitu (n4_c_HsodPg): downloaded 720p work video, cover, and metadata. YouTube has no subtitles; faster-whisper large-v3 CUDA transcription is running.

  • Completed EP132 Gao Jiyang / Xinghaitu note: youtube/zhangxiaojun/ep132-n4_c_HsodPg/ep132-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check passes with figs=9 readfig=8 boxes=28 term_digest=5 summaries=14, visual QA checked; cover + generated concept diagrams only, no repeated body speaker frames).

  • Continuing serial Zhang Xiaojun generation with EP130 Zhang Yueguang / Muyan Zhiyu (ruVJ_5dObxs): downloaded 720p work video, cover, and metadata. YouTube has no subtitles; faster-whisper large-v3 CUDA transcription is running.

  • Completed EP130 Zhang Yueguang / Muyan Zhiyu note: youtube/zhangxiaojun/ep130-ruVJ_5dObxs/ep130-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check passes with figs=9 readfig=9 boxes=26 term_digest=3 summaries=15, visual QA checked; cover + generated concept diagrams only, no repeated body speaker frames).

  • Fixed Zhang Xiaojun concept-figure rendering style after user reported tofu-box glyphs in EP133: added tools/scripts/render_zhangxiaojun_concept_figures.py, switched generated diagrams to a Chinese+Latin-safe AR PL font, simplified diagrams to sparse teaching cards, regenerated 60 figures across EP130/132/133/134/135/136/137/138/139, recompiled those PDFs, re-rendered visual QA, and rechecked whitespace.

  • Completed EP129 Zhang Peng / Zhipu note: youtube/zhangxiaojun/ep129-9zSMTUUEfmU/ep129-notes.pdf (21 pages, quality ⭐⭐⭐, coverage check passes with figs=9 readfig=10 boxes=27 term_digest=2 summaries=14, visual QA checked; cover + simplified generated concept diagrams only, no repeated body speaker frames).

  • Continuing serial Zhang Xiaojun generation with EP128 Manus / Ji Yichao Peak (MW-ezf2RhVg): downloaded 720p work video, cover, and metadata. YouTube has no subtitles; faster-whisper large-v3 CUDA transcription is starting.

  • Completed EP128 Manus / Ji Yichao Peak note: youtube/zhangxiaojun/ep128-MW-ezf2RhVg/ep128-notes.pdf (21 pages, quality ⭐⭐⭐, coverage check passes with figs=9 readfig=8 boxes=28 term_digest=3 summaries=13, visual QA checked; cover + simplified generated concept diagrams only, no repeated body speaker frames).

  • Continuing serial Zhang Xiaojun generation with EP127 LLM quarterly report cross-year conversation (SG90aehV3vU): downloaded 720p work video, cover, and metadata. YouTube has no subtitles; faster-whisper large-v3 CUDA transcription is starting.

  • Completed EP127 Guangmi cross-year LLM quarterly report note: youtube/zhangxiaojun/ep127-SG90aehV3vU/ep127-notes.pdf (26 pages, quality ⭐⭐⭐, coverage check clean with figs=11 readfig=11 boxes=30 term_digest=5 summaries=11, visual QA checked; cover + simplified generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP127 done; next AI/internet selected target is EP125 (k82iFzvKFCQ) because EP126 is excluded from the filtered queue.

  • Fixed manifest workflow to scan figures/ directories so generated podcast concept diagrams are recorded as local visual assets.

  • Completed EP125 Freda / Altimeter note: youtube/zhangxiaojun/ep125-k82iFzvKFCQ/ep125-notes.pdf (28 pages, quality ⭐⭐⭐, coverage check clean with figs=11 readfig=12 boxes=33 term_digest=5 summaries=14, visual QA checked; cover + simplified generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP125 done; next AI/internet selected target is EP123 (qZbzFZ2R_Nw) because EP124 and EP126 are excluded from the filtered queue.

  • Completed EP123 ONE2X / Wang Guan note: youtube/zhangxiaojun/ep123-qZbzFZ2R_Nw/ep123-notes.pdf (21 pages, quality ⭐⭐⭐, coverage check clean with figs=11 readfig=10 boxes=25 term_digest=4 summaries=12, visual QA checked; cover + simplified generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP123 done; next AI/internet selected target is EP121 (2o281Zy5aZE) because EP122 and EP124 are excluded from the filtered queue.

  • Completed EP121 DeepMind Tan Jie robotics note: youtube/zhangxiaojun/ep121-2o281Zy5aZE/ep121-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=11 readfig=10 boxes=27 term_digest=3 summaries=10, visual QA checked; cover + simplified generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP121 done; next AI/internet selected target is EP120 (40qPt8R2uys).

  • Completed EP120 Xiaopeng / Liu Xianming Physical AI note: youtube/zhangxiaojun/ep120-40qPt8R2uys/ep120-notes.pdf (21 pages, quality ⭐⭐⭐, coverage check clean with figs=11 readfig=9 boxes=30 term_digest=2 summaries=11, visual QA checked; cover + simplified generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP120 done; next AI/internet selected target is EP119 (858HR43pegk).

  • Completed EP119 Kimi Linear / MiniMax M2 architecture survey note: youtube/zhangxiaojun/ep119-858HR43pegk/ep119-notes.pdf (22 pages, quality ⭐⭐⭐, coverage check clean with figs=11 readfig=10 boxes=30 term_digest=2 summaries=12, visual QA checked; cover + simplified generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP119 done; next AI/internet selected target is EP118 (RxXVq7-sJzM).

  • Continuing serial Zhang Xiaojun generation with EP118 Li Xiang second interview (RxXVq7-sJzM).

  • Completed EP118 Li Xiang second interview note: youtube/zhangxiaojun/ep118-RxXVq7-sJzM/ep118-notes.pdf (22 pages, quality ⭐⭐⭐, coverage check clean with figs=12 readfig=11 boxes=36 term_digest=3 summaries=11, visual QA checked; cover + simplified generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP118 done; next AI/internet selected target is EP117 (zrvnoYYPaWQ).

  • Continuing serial Zhang Xiaojun generation with EP117 open-source paper exploration survey (zrvnoYYPaWQ).

  • Completed EP117 open-source paper exploration survey note: youtube/zhangxiaojun/ep117-zrvnoYYPaWQ/ep117-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=13 readfig=10 boxes=26 term_digest=3 summaries=7, visual QA checked; YouTube static cover plus generated concept diagrams, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP117 done; next AI/internet selected target is EP116 (khrOsS7YQn4).

  • Continuing serial Zhang Xiaojun generation with EP116 Wu Minghui 19-year history interview (khrOsS7YQn4).

  • Completed EP116 Wu Minghui / Minglue enterprise AI note: youtube/zhangxiaojun/ep116-khrOsS7YQn4/ep116-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=12 readfig=9 boxes=36 term_digest=3 summaries=7, visual QA checked; cover + generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP116 done; next AI/internet selected target is EP115 (gQgKkUsx5q0).

  • Continuing serial Zhang Xiaojun generation with EP115 OpenAI Yao Shunyu Agent interview (gQgKkUsx5q0).

  • Completed EP115 OpenAI Yao Shunyu Agent research note: youtube/zhangxiaojun/ep115-gQgKkUsx5q0/ep115-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=12 readfig=6 boxes=34 term_digest=4 summaries=8, visual QA checked; cover + generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP115 done; next AI/internet selected target is EP113 (ouG6jrkECrc) because EP114 is excluded from the filtered AI/internet queue.

  • Continuing serial Zhang Xiaojun generation with EP113 Yang Zhilin K2 / Agentic LLM interview (ouG6jrkECrc).

  • Completed EP113 Yang Zhilin / Kimi K2 Agentic LLM note: youtube/zhangxiaojun/ep113-ouG6jrkECrc/ep113-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=13 readfig=6 boxes=31 term_digest=4 summaries=7, visual QA checked; cover + generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP113 done; next AI/internet selected target is EP112 (6yExfoTuSWw).

  • Continuing serial Zhang Xiaojun generation with EP112 Guangmi LLM quarterly report (6yExfoTuSWw).

  • Completed EP112 Guangmi LLM quarterly report note: youtube/zhangxiaojun/ep112-6yExfoTuSWw/ep112-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=11 readfig=8 boxes=32 term_digest=1 summaries=10, visual QA checked; cover + generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP112 done; next AI/internet selected target is EP111 (JxEetUlV9RA).

  • Continuing serial Zhang Xiaojun generation with EP111 Li Yifan lidar entrepreneurship interview (JxEetUlV9RA).

  • Completed EP111 Li Yifan / Hesai lidar entrepreneurship note: youtube/zhangxiaojun/ep111-JxEetUlV9RA/ep111-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=12 readfig=11 boxes=33 term_digest=1 summaries=8, visual QA checked; cover + generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP111 done; next AI/internet selected target is EP110 (8dKBH4x0D9o).

  • Continuing serial Zhang Xiaojun generation with EP110 Kimi K2 report / ChatGPT Agent / Qwen3-Coder survey (8dKBH4x0D9o).

  • Completed EP110 Kimi K2 / ChatGPT Agent / Qwen3-Coder technical report note: youtube/zhangxiaojun/ep110-8dKBH4x0D9o/ep110-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=12 readfig=8 boxes=27 term_digest=3 summaries=9, visual QA checked; cover + generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP110 done; next AI/internet selected target is Lovart video special (biptonYq-ys).

  • Continuing serial Zhang Xiaojun generation with Lovart / Chen Mian video special (biptonYq-ys).

  • Completed Lovart / Chen Mian video special note: youtube/zhangxiaojun/special-biptonYq-ys/lovart-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=9 readfig=5 boxes=38 term_digest=2 summaries=10, visual QA checked; cover/title-frame plus generated concept diagrams, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark biptonYq-ys done; next AI/internet selected target is EP109 (pWY0HVUH8GA).

  • Completed EP109 Xie Chen / Guanglun embodied simulation and synthetic data note: youtube/zhangxiaojun/ep109-pWY0HVUH8GA/ep109-notes.pdf (24 pages, quality ⭐⭐⭐, coverage check clean with figs=11 readfig=13 boxes=33 term_digest=6 summaries=9, visual QA checked; cover plus generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP109 done; next AI/internet selected target is Puptr04av5g (Wang He embodied intelligence history/capital-chaos subtitle edition).

  • EP106/Puptr04av5g transcription command correction: initial run used legacy --output-json/--output-srt flags; current tool requires --out-prefix. Re-running with --out-prefix.

  • EP106/Puptr04av5g transcription issue: single full-video faster-whisper run stalled after writing partial output around 01:08 despite process remaining alive. Root cause likely one long/awkward decode chunk rather than CUDA failure. Switched plan to chapter-based audio chunks and merge outputs.

  • EP106/Puptr04av5g transcription issue 2: per-chunk shell loop also stalled on 3-minute chunk01 during model initialization/decode. Next approach: load WhisperModel once in a custom batch script and transcribe all chapter WAV chunks in one process.

  • EP106/Puptr04av5g GPU runtime anomaly: CUDA/CTranslate2 calls left unkillable processes even after SIGKILL, so switched to testing CPU int8 transcription for chapter chunks rather than continuing to use the broken CUDA path.

  • EP106/Puptr04av5g CPU fallback: chunks 01--03 transcribed successfully; generated provisional chapter-transcripts.md with later chunks marked pending while CPU transcription continues.

  • EP106/Puptr04av5g CPU fallback: chunks 01--04 transcribed successfully; chunk05 running. Updated chapter-transcripts.md with completed chunks and pending markers for the rest.

  • EP106/Puptr04av5g CPU fallback: chunk05 completed; updated chapter-transcripts.md with chapters 01--05. Remaining chunks 06--11 still running/pending.

  • EP106/Puptr04av5g CPU fallback transcription completed for all 11 chunks and merged into transcript.zh.json/srt/txt; preserved the earlier stalled GPU partial transcript as .gpu-partial files for diagnosis.

  • Completed EP106/Puptr04av5g Wang He embodied intelligence subtitle edition note: youtube/zhangxiaojun/ep106-Puptr04av5g/ep106-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=12 readfig=18 boxes=25 term_digest=3 summaries=9, visual QA checked; cover + generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark Puptr04av5g done and original EP106 9DnjiuN6Yg0 duplicate/covered by the subtitle edition.

  • Continuing serial Zhang Xiaojun generation with nuclear-fusion interview pVuE4J5cn98; metadata and 360p work video downloaded. YouTube exposes no subtitles despite title marker 含字幕; using fixed-duration CPU transcription chunks because CUDA cleanup was unreliable in the previous item.

  • Completed nuclear-fusion interview note pVuE4J5cn98: youtube/zhangxiaojun/special-pVuE4J5cn98/fusion-notes.pdf (21 pages, quality ⭐⭐⭐, coverage check clean with figs=10 readfig=12 boxes=33 term_digest=3 summaries=9, visual QA checked; cover + generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark pVuE4J5cn98 done; next AI/internet selected target will be discovered from queue.

  • Continuing serial Zhang Xiaojun generation with EP104 Rokid / Zhu Mingming (qW-kgogQwJc); metadata and 360p work video downloaded. YouTube exposes no subtitles; chapter-based CPU transcription is starting.

  • EP104 CPU fallback transcription completed for all 20 fixed-duration chunks and was merged into transcript.zh.json/srt/txt plus chapter-transcripts.md.

  • Completed EP104 Rokid / Zhu Mingming note: youtube/zhangxiaojun/ep104-qW-kgogQwJc/ep104-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=9 readfig=8 boxes=33 term_digest=5 teacher_voice=7 summaries=7, visual QA checked; cover plus generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP104 done; next AI/internet selected target should be discovered from ai-internet-queue.json.

  • Checked next AI/internet queue item EP103 (Xo7TxXkNsoA) and found it is a duplicate upload of the already completed Lovart / Chen Mian video special biptonYq-ys: same title family, same 1h45m interview, same opening transcript and main chapter content. EP103 was transcribed locally before the duplicate was confirmed; no second note was generated. Queues were updated to mark EP103 duplicate, covered by youtube/zhangxiaojun/special-biptonYq-ys/lovart-notes.tex.

  • Completed EP102 Zhang Xiangyu multimodal research note: youtube/zhangxiaojun/ep102-vWrYHvSRz0s/ep102-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=13 readfig=12 boxes=26 term_digest=6 teacher_voice=4 summaries=8, visual QA checked; cover plus generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP102 done; next AI/internet selected target should be discovered from ai-internet-queue.json.

  • Completed EP101 YouWare / Ming Chaoping Agent application entrepreneurship note: youtube/zhangxiaojun/ep101-a04POJEknCY/ep101-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=13 readfig=14 boxes=33 term_digest=4 teacher_voice=4 summaries=9, visual QA checked; cover plus generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP101 done; next AI/internet selected target should be discovered from ai-internet-queue.json.

  • Completed EP100 Mercedes-Benz / Ola Källenius transformation note: youtube/zhangxiaojun/ep100-9Yjws_rt378/ep100-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=9 readfig=8 boxes=27 term_digest=4 teacher_voice=4 summaries=6, visual QA checked; cover plus generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP100 done; next AI/internet selected target should be discovered from ai-internet-queue.json.

  • Checked next AI/internet queue item EP99 (aoaSwGAJW6M) and found it is covered by the already completed nuclear-fusion subtitle edition pVuE4J5cn98: same guest/topic/title family. No second note generated. Queues were updated to mark EP99 duplicate, covered by youtube/zhangxiaojun/special-pVuE4J5cn98/fusion-notes.tex.

  • Started VLA paper-survey投屏版 eiQFomOuCJs: metadata, cover, 1080p work video, 34 chapter-aligned slide/frame candidates, and contact sheet were acquired under youtube/zhangxiaojun/special-eiQFomOuCJs/.

  • VLA paper-survey CPU fallback transcription completed for all 22 fixed-duration chunks and was merged into transcript.zh.json/srt/txt plus chapter-transcripts.md (4215 segments). Next step is to write the slide/frame-backed VLA course note.

  • Completed VLA paper-survey投屏版 eiQFomOuCJs: youtube/zhangxiaojun/special-eiQFomOuCJs/vla-notes.pdf (24 pages, quality ⭐⭐⭐, coverage check clean with figs=26 readfig=17 boxes=23 term_digest=1 teacher_voice=1 summaries=6, visual QA checked; real投屏 frames used as visual spine).

  • Updated Zhang Xiaojun queues to mark eiQFomOuCJs done; next AI/internet selected target should be discovered from ai-internet-queue.json.

  • Checked next AI/internet queue item EP98 (3jI6F3M2ocU) and confirmed it is the same VLA paper-survey content as the completed投屏版 eiQFomOuCJs: same duration, upload date, 34 chapter titles/timestamps, and description storyline. Queues were updated to mark EP98 duplicate, covered by youtube/zhangxiaojun/special-eiQFomOuCJs/vla-notes.tex.

  • Started EP97 YshXmh_q_Q4 Q1 2025 large-model quarterly review: metadata, cover, description, and 360p work video downloaded; YouTube exposes no subtitles or auto captions. Audio was split into 16 fixed-duration chunks and CPU int8 transcription is in progress.

  • Completed EP97 Q1 2025 large-model quarterly review note: youtube/zhangxiaojun/ep97-YshXmh_q_Q4/ep97-notes.pdf (26 pages, quality ⭐⭐⭐, coverage check clean with figs=13 readfig=12 boxes=43 term_digest=5 teacher_voice=8 summaries=12, visual QA checked; cover plus generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP97 done; next AI/internet selected target should be discovered from ai-internet-queue.json.

  • Started EP96 qtugoE1xQZk autonomous-driving interview: metadata, cover, and description downloaded. YouTube exposes no subtitles or auto captions; queue status set to in_progress.

  • Completed EP96 Lang Xianpeng autonomous-driving technical interview note: youtube/zhangxiaojun/ep96-qtugoE1xQZk/ep96-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=13 readfig=12 boxes=30 term_digest=7 teacher_voice=7 summaries=7, visual QA checked; cover plus generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP96 done; next AI/internet selected target should be discovered from ai-internet-queue.json.

  • Started EP95 VdWEE6vOYRw Manus founder Xiao Hong interview: metadata, cover, and description downloaded. It is distinct from EP128 Peak/Manus final interview; YouTube exposes no subtitles or auto captions; queue status set to in_progress.

  • Completed EP95 Manus founder Xiao Hong interview note: youtube/zhangxiaojun/ep95-VdWEE6vOYRw/ep95-notes.pdf (20 pages, quality ⭐⭐⭐, coverage check clean with figs=12 readfig=10 boxes=33 term_digest=4 teacher_voice=6 summaries=8, visual QA checked; cover plus generated concept diagrams only, no repeated body speaker frames).

  • Updated Zhang Xiaojun queues to mark EP95 done; next AI/internet selected target should be discovered from ai-internet-queue.json.