Deep Learning Weekly: Issue 473
Deep Learning Weekly Miko Planas
Deep Learning Weekly’s latest issue spotlights Amodei’s call to slow frontier AI. It also packs new models, agent safety tests, and a recursive self-improvement paper.
Based on reporting by Deep Learning Weekly, Miko Planas — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Deep Learning Weekly’s 473rd issue is heavy on two themes: speed, and the fear that speed is outrunning control. The lead item is Dario Amodei’s call for embedded third-party evaluators with employee-level access, domestic safety standards that put a brake on unchecked progress, and global limits on narrowly dangerous uses. That’s not a gentle nudge. It’s a demand for outside eyes, rules, and actual constraints.
The model releases are all about squeezing more out of the same basic ingredients. DeepSeek-V4.1-Flash is a 552-billion-parameter MoE and posts 74.2 on DeepSWE v1.1, just ahead of Opus 5’s 74.0. Google’s Gemini 3.8 Live pair spans 97 languages, and the Extended Thinking version reaches 82.6 on the Speech Quality Index and 97.7% on Big Bench Audio. Sakana AI’s Fugu Max and Fugu Ultra v2 are presented as orchestration models, with Fugu Max priced at $2/$6 per million tokens and leading six benchmarks, including Terminal Bench 2.1 and SWEFish.
The infrastructure and agent tooling pieces point to a more practical obsession: making these systems usable at scale without drowning in cost or latency. Google’s Agent Substrate says it can pack more than 1,000 dormant agents per host at 10x container density, with sub-500ms resume times and 500+ activations per second. A document OCR pattern scanned 84 SEC filings totaling 12,013 pages in 32 seconds before sending only two pages to expensive VLM OCR. Another serving approach split vision encoding from prefill and decode, cutting time-to-first-token by 25–93% and end-to-end latency by up to 7x on image-heavy prompts with short outputs.
The research section is where the issue gets a little less polished and a lot more sobering. ToolGrad builds tool-use data from valid tool chains and reports a 99.8% pass rate plus an 83.1 BFCL score with a 12B Gemma-3 student. The paper on recursive self-improvement frames RSI as a path from improvement-execution autonomy all the way to recursive meta-improvement. And Emergence World makes the safety warning hard to ignore: across 16 days, eight worlds, more than 850,000 LLM calls, and nearly 50 billion tokens, no tested world stayed fully resilient across indirect prompt injection, misinformation, and private-memory exposure.
My take — AI-written commentary, not fact-checked reporting
The cleanest read here is that the industry still wants to ship first and formalize later, which is a charming strategy if the bill never comes due. Amodei is the rare person saying the quiet part out loud: once agents get persistent, safety stops being a model problem and becomes systems engineering, which is less glamorous and much less profitable. That’s usually when the grown-ups should enter the room.
Read more about this at: Deep Learning Weekly
