On September 15, the call to not slow down AI rang out in unison from both industry and the administration. Nvidia CEO Jensen Huang told President Trump directly that he won't let AI slow down, and Trump himself criticized Anthropic CEO Dario Amodei by name. Meanwhile, a memoir by a DeepSeek engineer—who wrote that "in six months, AI will surpass the operators I write"—has been drawing attention, a sign that the wave of acceleration is undeniably reaching the development front lines as well.

CEO Huang: "We Won't Let AI Slow Down" — Trump Also Calls Out Amodei by Name

According to TechCrunch, Nvidia CEO Jensen Huang told President Trump regarding a slowdown in the pace of AI development: "We're not going to let that happen." With Elon Musk and Sam Altman backing Amodei's call to slow down, Huang has clearly staked out a different position.

Before that, Trump posted on social media that the only guardrails AI needs are "a strong and smart president," emphasizing sweeping criminal and regulatory powers over AI companies. He criticized Amodei by name for "pretending to be a perfect little angel" and denounced moves to obstruct AI development as a conspiracy to benefit China (ITmedia).

On the other side, voices in Congress are growing louder in demanding crackdowns. Former FTC (Federal Trade Commission) chair Lina Khan, citing precedent from 1934, argued that AI company CEOs need "handcuffs"—that is, criminal liability (The Register). Senator Bernie Sanders has also reportedly put forward a proposal to impose 20-year prison sentences on developers of superintelligent AI.

Why it matters: This goes a step beyond the previous stage of "Trump rejecting slowdown requests"—the CEO of Nvidia, the top GPU supplier, has now explicitly voiced opposition aimed at the administration itself. The temperature gap between the executive branch and industry pushing for "acceleration" and Congress plus some executives pushing for "regulation" looks likely to shape the course of future legislation.

A DeepSeek Engineer's Memoir: "Burying Talent in Yesterday"

A memoir penned by a DeepSeek engineer has been drawing attention in Reddit's LocalLLaMA community (the original is said to be an English translation of a Chinese-language blog).

The author is the person who handled the core attention operators in DeepSeek v4.1. Precisely for that reason, their retrospective observation carries weight: a year ago, AI was a "little assistant" that looked up documentation, read code, and found bugs; now it reads CUDA, PTX, and SASS, analyzes per-instruction stall cycles with specialized tools, and optimizes operators on its own—an "expert."

"AI thinks at 300 tokens per second, types a command in 0.5 seconds, and finishes writing code in 20 seconds. I can't do that." The author reckons that within six months to a year, AI-written operators will match or exceed their own, and while they won't lose their job, they will likely have to leave the field of operator design they loved and retrain as a "mech pilot for agents."

Especially memorable is the metaphor of a hand-knit sweater artisan facing the arrival of automated knitting machines. Even if the work remains, the joy of quietly working the needles by the window gets crushed by the roar of the machines. Even so, they write—as their reason for staying on the front line—"if I'm going to be revolutionized, I want the one doing the revolutionizing to be me."

Occamy-1.0 — A 35B Open Model Maximizing "Intelligence per Dollar"

Featured on Hugging Face Daily Papers, Occamy-1.0 is a 35B-parameter open model that plants its flag on the "Pareto frontier." According to its developers, agents invoke a model hundreds of times for a single task, which makes "intelligence per dollar" a metric every bit as important as "intelligence per call."

With an eye on agents doing real work (co-work), the goal is to deliver practical intelligence on modest compute—the tagline being "less compute, serious work." Detailed benchmark figures remain limited to the public summary at this point, but this should add a promising option for running agents under tight cost constraints.

Qwen3.8-Flash-Next at ~20 tok/s on 12GB VRAM — The Frontier of Local Inference

A validation report posted on LocalLLaMA is practical and fascinating: it runs Qwen3.8-Flash-Next, a 125B-A6B MoE model (including a 51B n-gram table), on a mid-range rig of an RTX 4070 (12GB VRAM) + 64GB DDR5-5600 + Gen4 NVMe.

Generation speed that started at around 6 tok/s reportedly climbed to the 19–20 tok/s range after stacking 4.27bpw quantization, offloading the n-gram table to SSD, and applying MoE improvement patches. Combining an unmerged MTP (multi-token prediction) PR with a compact MTP head then broke through the 20 tok/s barrier, with prediction acceptance rates of 77–96%.

The poster rates it as "the most intelligence this machine can deliver." On a rig where a lack of VRAM rules out 27B dense models, MoE plus offloading that leverages fast RAM is the counterintuitive play. It's a validation worth referencing for users in low-VRAM, large-RAM environments.

ReactHuman — Robot Brains Fail at "Split-Second Evasion" One Time in Three

Published on Hugging Face Daily Papers, ReactHuman is the first physics-grounded benchmark to cast an embodied multimodal LLM (MLLM) as the brain of a humanoid in simulation and measure its reactions to sudden household hazards.

Catching a slipping plate, dodging a falling knife—reaction abilities like these can't be measured by passive, video-Q&A-style evaluations of physical understanding. ReactHuman makes 17 event families and more than 1,000 scenes fully reproducible with 240Hz rigid-body simulation and derives annotation-free ground truth. It also includes adversarial objects whose appearance contradicts their physics, such as a "foam anvil" or a "steel apple."

The results of evaluating seven representative MLLMs are brutal. Roughly one time in three, they fail to deal with a hazard; they act on fixed tendencies rather than the scene they observed; they trust appearances over motion; and even when they pick the right action, they miss the interception point by whole meters. What's more, these failures did not shrink with model scale-up. It's a study that presents, as a fine-grained diagnosis, the challenges standing in the way of practical household robots.

"AI Bot Traffic Is Out of Control" — The Cry of a Web Operator

On Hacker News's Ask section, a plea from the operator of a web app with 50,000 statically generated pages has stirred debate.

According to the post, Facebook's crawler alone generated more than 3 million requests to the same page over two days. Even after blocking as much as possible on Vercel and Cloudflare, AI scrapers send requests through residential proxies and even create accounts to get inside the authentication gate protected by Cloudflare Turnstile. Hosting costs have skyrocketed, and the operator, worn down, confesses, "I can no longer trust anything I read or see on the web." As a last resort, they plan to try switching to hCaptcha next.

Countermeasures against generative AI training and search bots are increasingly outgrowing the realm of robots.txt and IP blocking. The cat-and-mouse game between content providers and crawlers looks set to shape the cost of keeping the web running.