Interesting Finds — 2026-09-06
Four notes: Raschka on looped transformers, TxBench antibody evals, Google Pics gated to paid tiers, and the Gemini Flash Cyber overlap.
Each is a separate find. Editorial takes are mine where noted.
1. Raschka on Astra and looped transformers — small tweak, loud hype
Sebastian Raschka (sebastianraschka.com, Sep 2): the “recurrent depth” in OpenAI’s rumored Astra is just layer reuse — the looped transformer idea. Nanbeige 4.2-3B, trained on 28T tokens, reuses its 22-layer stack twice for an effective 44 layers without duplicating weights: same storage, roughly twice the compute, about 75 percent of the token efficiency of a standard architecture per the technical report. Two passes were the tradeoff point; more passes added cost without gains. The older Mixture-of-Recursions paper adds a learned router so easy tokens exit early and hard tokens get extra passes.
Raschka’s two corrections: the loop is a tiny architectural tweak, not the reason Astra would be good; and layer reuse does not by itself hide chain-of-thought — computation in latent activations is what ordinary layers already do, though more recurrent passes can mean fewer readable reasoning tokens.
Meaning: parameter efficiency without memory savings, paid for in compute. Adaptive-depth routing is the more interesting half — spend passes where tokens need them. Do not price the rumor into Astra’s capability until the model is public.
2. TxBench-AB — 100 wet-grounded antibody evals
LatchBio (latch.bio, PDF, 13pp): TxBench Antibody Discovery — 100 evaluations across ten competencies from target assessment and assay design through cellular pharmacology, engineering, and de-risking, each derived from public experimental studies with deterministic grading of a consequential advance-or-kill decision. Across 20 model-harness configurations at three attempts each, Opus 5 with Claude Code led at 53.0 percent evaluable pass rate, with four xAI and Google configurations within three points. Harness choice moved scientific accuracy, failures were often reproducible, every eval was solved by at least one configuration and 89 by models from two or more families.
Meaning: the benchmark that matters here is decision grounding — which experiments are informative, how conflicting measurements resolve, when to terminate. A 53 percent ceiling with harness sensitivity says execution scaffolding is part of scientific accuracy, not packaging. Unverified beyond the paper’s own reporting; still, the design targets the right bottleneck.
3. Google Pics — image tooling gated to paid tiers
Google (blog.google): Google Pics, built on the Nano Banana image model, is rolling out over the coming weeks to AI Pro and Ultra subscribers plus most Workspace business customers — standalone plus integrated into Slides, Docs, and soon Drive. Generation plus fine-tuned precise edits, collaborative editing, Drive file drop-in.
Take — agree with Keith: pretty annoying none of this goes to Plus users. Given how far they fell behind, gating the catch-up tooling to the top tiers reads poorly thought out. The product move is fine — edit where you already work instead of tab-switching. The tiering is the story: Google is charging its most loyal base for parity.
4. Note: Gemini 3.8 Flash Cyber already covered Sep 3
The Gemini 3.8 Flash and Flash Cyber link in this batch (blog.google) duplicates the Sep 2 announcement already carried as item 7 of Interesting Finds 2026-09-03. Keeping the single record there — no second entry. The new material in this batch is items 1 through 3 above.
Links are the sources. Papers are as claimed, not independently verified here.