Influence functions: what you did, and where to restart

Covers Wed 9 Sep – Wed 16 Sep 2026 (weeks 37–38). Generated Thu 17 Sep 2026 (week 38), 09:45 PDT on the Intel MacBook Pro. Sources: Codex and Claude Code session logs, the live Google Doc, GitHub, check-ins, Tickertape, Gmail and earlier reports. The M5 was asleep, so its local session logs were not read.

Bottom line

• Two sittings, a lot built. On Wed 9 Sep and Sat 12 Sep (week 37) you and your agents made 3 prep reports, 2 sets of notes, an MNIST ASTRA lab, 4 interactive tutorial pages and an MNIST influence atlas.
• The two outputs it was all for haven't happened. You haven't applied, and you haven't written to Roger Grosse. Gmail since 8 Sep has no application confirmation and nothing to or from him. I found no draft of the blog post anywhere.
• Your goal, in your own doc: research influence functions → blog post → apply to Anthropic in week 38, which is this week. On Wed 16 Sep you said the application is "maybe… pushed to Friday."
• 🛑 Don't send the 9 Sep draft to Roger as written. It pitches the Wick/Isserlis factorization. Tuesday's mini-lab found that result already in Martens & Grosse 2015 (Appendix A, Lemma 4), and measured it at 0.08%.
• 🛑 ~/git/influence-functions still has no git history. It is 2.8 GB, on the Intel only, and has been flagged daily since Tue 15 Sep.

Where everything lives

The sessions

Day by day

• Wed 9 Sep. At 09:27 you told your coach you'd apply that day. The 13:00 block went to preparation instead. Claude's plan was: apply, reply in Roger's "catch up?" thread, then do the math. Codex then built the lab and the Kaczmarz tutorial. At 21:13 you said «большой шанс, что меня не возьмут туда» ("there's a big chance they won't take me").
• Thu 10 Sep. At 07:38: "I want to get into Roger Grosse's influence functions… a tutorial." The afternoon went to launching the MNIST competition.
• Fri 11 Sep, Kateryna Peters (session 5). The deliverable became a tutorial blog post on how the mechanism works: about 16 h, 3 done. Plan: 2 days × 4 h, then send it to Roger, raw if need be, on Sat, Wed and Thu, before Fri 18 Sep. You named the blocker yourself: shame at showing a professor something raw.
• Sat 12 Sep. You started the project doc, and the tutorial pages and atlas were built. At 15:44 you emailed Tim Salimans that Roger was recruiting you. He replied (Netherlands, occasional SF visits). You asked on Sunday when he's next in SF, and he hasn't answered.
• Mon 14 – Tue 15 Sep. At 06:45 Monday: "I should just do a mini lab with MNIST." The overnight pass ran it. On Tuesday you told Jason Yosinski you were writing "an influence functions tutorial so I can send it to Roger Grosse."
• Wed 16 Sep. A planned 13:00 influence-functions slot collided with Ainekko, and Anthropic got 0 h. You added a Friday 11:00–16:00 "anthropic" block, which overlaps Kateryna Peters at 11:00 and Tau Robotics at 13:30.

What the work found

1. ASTRA solves more accurately, but no method predicted retraining. The lab ran the 11.97M-parameter Ciresan MLP through EK-FAC → ASTRA → PCG and 24 half-data retrains. Mean LDS: EK-FAC 0.014, ASTRA −0.033, PCG 0.030, and every per-query 95% interval includes zero. The Wednesday Modal bill was $0.34. SOURCE was not implemented.
2. At this scale the ground truth is noise. In the mini-lab, two seeds of the same leave-one-out run agree at Spearman −0.007. Score against PBRF or dattri's pre-retrained MNIST models, not cold-start retrains.
3. Wick/Isserlis is neither new nor large. The cross-covariance is the layer's mean gradient, so it vanishes at convergence (0.08%).
4. What survives: K-FAC's error at convergence is the basis, not the eigenvalues. On a small MLP's last layer, relative error is EK-FAC 0.671, best rank-1 Kronecker 0.557, rank-2 0.434, and 45% of the mass lies where EK-FAC can't reach. Hong, Eschenhagen, Mlodozeniec & Turner (2509.23437) say the opposite, but their own share falls 60 → 58 → 41% with training. Runa Eschenhagen is on Roger's team. A per-layer, per-epoch sweep settles it in about an hour on the laptop. Kronecker sums have prior art (Koroko et al., 2201.10285).
5. ⚠️ The 9 Sep "half-hour falsification test" is a trap as written. MNIST's constant border pixels give a fake ρ₁ ≈ 0.55.
6. The atlas uses a CNN at 0.39% test error. It shows the top-5 training digits for the 100 most ambiguous test digits. It's a good picture for the post, but nothing validates it.

What you planned, and where it stands

Restart: suggested order

1. Commit the lab (10 min). Run git init, add the two toy scripts from ~/.codex/attachments/…, and push. The existing .gitignore should keep .venv, node_modules and checkpoints out.
2. Apply (20 min): Research Engineer / Scientist, Alignment, San Francisco. It doesn't need the post, and the post has already slipped twice.
3. Write the post from what exists, and send it raw in Roger's "catch up?" thread. - Intuition from the Kaczmarz pages, cost from The Price of Influence, one atlas example. - Then the honest MNIST result: better solves, no LDS gain, noisy ground truth. - End on the basis-vs-eigenvalue question as your next experiment. Leave Wick out.

Corrections made this pass: the lab tutorial is 2,760 words, not the "20,000" the mini-lab report gave (that was its size in bytes). The 9 Sep report now marks its Wick pitch and draft message as superseded.

Not checked: the M5's local logs (ssh timed out), the transcripts of the claude.ai chat and the Fable session (web only), and the atlas page itself (sign-in).

Total life satisfaction

This is one of the two things this week is for. The research part is finished. What's left is a 20-minute application and one raw email, and on 11 Sep you said yourself that the blocker is showing raw work, not the work. Sending both before Kateryna Peters tomorrow turns eight days of preparation into the result you set out to get.