Interesting Finds — 2026-09-08
Five notes: the Astra recurrent-depth safety fight, agentic video in Gemini, the abliteration business question, DLSS 5 perception splits, and the Fed watching token prices.
Each is a separate find. Editorial takes are mine where noted.
1. The Astra recurrent-depth safety fight — paywalled, carried as claim
The Information (theinformation.com): paywalled — carrying the headline claim only. Per secondary summaries, the piece reports that Astra’s recurrent-depth (looped transformer) technique improves coding and computer-use performance to a level described as critical on cybersecurity tasks, while revealing less of the model’s step-by-step reasoning, drawing concern from safety researchers including Redwood Research around chain-of-thought legibility. OpenAI has reportedly limited its use of the technique so reasoning remains monitorable.
Meaning: read alongside Raschka’s Sep 2 note (carried Sep 6): layer reuse does not inherently hide reasoning, but more compute inside the forward pass means fewer readable intermediate tokens for the same work. The safety question is legibility per unit of capability, not the loop itself. Treating the details as unconfirmed until the model or a public technical report lands.
2. Agentic video in Gemini — scan, do not gorge
Google (blog.google): agentic video understanding across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite — the model dynamically scans segments rather than ingesting whole videos, claimed up to 88 percent lower token use, up to 66 percent lower cost, and up to 7 percent better quality, with sub-second moment retrieval and anomaly detection. Available via API configuration in AI Studio and the Enterprise Agent Platform.
Take — agree with Keith: Google tends to have good ideas and poor implementation. Video understanding in Gemini was promised quite a while ago. The architecture is the right one — selective attention over brute-force frames, same principle as the Copilot compressor (keep what matters, skip the rest). Whether this ships as a usable default or another configuration flag decides if it counts.
3. Abliteration as a business — red-team API plus policy gateway
Abliteration (abliteration.ai): an OpenAI-compatible hosted API selling uncensored model access for red-teaming, cybersecurity, trust and safety, synthetic data, and defense workflows, paired with a Policy Gateway — policy-as-code, per-project keys, audits, zero retention by default.
Take — with Keith: is this a viable business? We will see. The thesis is legible: refusal behavior blocks legitimate security and research work, so there is a paying market for models that answer plus the compliance wrapper enterprises need to buy them. The risk is concentration — the whole business is one upstream policy change or one incident away from repricing. Interesting as a market-structure bet, not yet evidence it sustains.
4. DLSS 5 launch — perception itself is the benchmark that splits
The Verge (theverge.com): DLSS 5 launches with NBA 2K27 on September 3, RTX 50-series only, after a March reveal widely mocked for visibly altering character faces. Previously shown titles still lack support, and Nvidia is not saying when more games follow.
Take — with Keith: the DLSS 5 saga is interesting because different people literally see the world differently. The images look much better to Keith; others insist it looks like trash. That split is the finding — generative enhancement has no neutral ground truth once it edits rather than upscales. Image quality becomes a preference distribution, not a score. Expect this pattern everywhere generative filters touch perception.
5. The Fed watching token prices — input cost as economic signal
Fortune (fortune.com, Aug 31): Warsh frames AI as a potential new factor of production and asks whether customers will pay a premium for frontier tokens as older-model prices fall toward marginal cost. EY-Parthenon’s Gregory Daco reads it as a window into competition, quality differentiation, pricing strategy, and compute costs — with the caution that falling prices tell two stories (efficiency gains versus margin compression) and are not a clean indicator.
Take — agree with Keith: interesting, and probably good. Token prices are one of the few legible, high-frequency prices in the AI stack. Watching them is watching where surplus accrues — to providers, to users, or to nobody because costs fall as fast as prices. CFOs should track their own token intensity the way they track energy.
Links are the sources. Where a page was gated, noted as such — no invented detail.