Quick Hits — September 7, 2026
The midday round: a memory crunch starts setting phone prices, China's model labs move onto Tmall's shelves, and Cisco's numbers say a clean single-turn safety score tells you almost nothing.
Cisco tested 15 frontier models against attackers who get to come back for a second turn — and none of them held. Its researchers split harmful requests across a conversation rather than firing them off in one prompt, and multi-turn attack success rates landed between 7.89% and 88.30%, against 2.19% to 64.91% for the same models on single-turn tests. The orderings don't match: GPT-5.4 goes from 2.74% to 24.68%, Gemini 3 Pro from 18.10% to 73.35%, while Amazon's Nova 2 Lite inverts the pattern — worst-looking first (34.05% single-turn) but the best in the cohort under sustained pressure (7.89%). The sharpest finding is a settings toggle, not a model: Grok 4.1 Fast drops from 88.30% to 43.47% when reasoning is switched on. The take: if you picked a model on published refusal rates, you bought a number that moves by tens of points once someone is willing to chat — and no public benchmark currently reports that gap.
The memory shortage is now deciding which phones get made, not just what they cost. Apple is preparing roughly 10% more iPhone 18 Pro and Pro Max units for the September-to-December window than it allocated to last year's Pro models, concentrating its lineup at the high end — the only place with margin to absorb memory that Counterpoint says now runs about $181 for 16GB of smartphone DRAM, up from roughly $42 a year earlier. Micron's new New York fab won't produce meaningful output until 2030; its Idaho site starts wafers in mid-2027. Samsung's memory chief has said the company is prioritizing customers "who can guarantee committed future captive demand." The take: the AI buildout didn't create the shortage, but it converted consumer electronics into the customer that gets allocated last — and phones get more expensive before they get more capable.
Insilico Medicine says six independent aging clocks all point the same way on its AI-discovered lung drug. Using blood-protein data from 42 patients in a Phase IIa trial of rentosertib, six separately built proteomic clocks each showed biological age moving down against placebo, peaking at roughly three to four years of reversal at week four on the 30 mg twice-daily dose (up to six years on one clock). The lung-function signal is separate: the dose that best improved forced vital capacity was not the dose with the strongest age effect, which Insilico reads as the anti-aging activity not being a downstream result of treating the disease. Zhavoronkov is presenting it in Paris on September 8; the drug, a TNIK inhibitor found with the company's generative platform, is in Phase III in China for idiopathic pulmonary fibrosis. The take: 42 patients and a biomarker endpoint is a proof of concept, not a longevity drug — the real claim is that aging endpoints can be bolted onto ordinary disease trials.
Chinese model labs are now selling tokens on Tmall, next to the video memberships. Zhipu opened a flagship store on September 2 listing nothing but four token packages, and a day later Tmall launched an AI Space Station recharge center carrying subscriptions from Alibaba Cloud, Zhipu, Kimi, and MiniMax. The context is in Zhipu's interim results: first-half revenue of 954 million yuan, up nearly 400% year over year, with open-platform and API revenue up roughly 27-fold to 86.5% of the total while on-premise deployment revenue fell 54.6%. The take: a storefront is a distribution story, not a product one — the value of a token package still depends on a number nobody can tell you in advance, since the same 10,000 points is several small projects or one bad afternoon with a large codebase.
Optuna got a Rust rewrite and a major version at the same time. Preferred Networks released Rustuna, an implementation of the hyperparameter search library that keeps the familiar API but drops Python dependencies and adds an option to discard trial histories an algorithm no longer needs — which matters because Optuna keeps every trial in memory, so CMA-ES and NSGA-II runs slow down as they go. Its TPE sampler handles 10,000 trials in under a minute where Optuna's guidance was to stay under 1,000, and a production user, Matlantis, reports flat memory and per-trial time after migrating. Author c-bata conceded on Reddit that if each of your trials takes ten minutes, none of this matters. The take: the honest pitch is supply-chain hygiene and long-run stability, not raw speed — which is why the "no Python dependencies" line is the one worth reading twice.
Yandex Research is arguing that the runtime, not the model or the harness, is the missing lever on agent capability. The team's writeup proposes editing a model's inference state directly — the KV cache that holds what it has already attended to — to make systems interactive rather than turn-based, building on its earlier Hogwild! Inference and AsyncReasoning work and previewing a Qwen3.8-27B agent playing DOOM with the technique. The framing question is whether there's a layer between "swap the model," which is expensive, and "rewrite the harness," which is too abstract. The take: if writing to the cache while it runs is a legitimate control surface, agents stop being a loop of finished turns — and the safety assumptions built on "one request, one response" stop applying.
Unitree published what it calls the first fully autonomous humanoid combat footage, driven in real time by its UnifoLM-X2-1.0 world model. The video shows a G1 throwing combinations and recovering from hits with no teleoperation, which the company frames as evidence that world-model-driven humanoids can plan, decide, and act dynamically enough to be deployed at scale. The G1 boxing configuration is already a product page with motion-capture training, controller and voice control, and livestreamed events. The take: a world model that predicts far enough ahead to take a punch and stay upright is a genuinely useful robotics claim — the demo is fighting because fighting is the legible way to show it.
