Liu Gong’s read of the week: the story in AI moved from “how smart is the model” to “how long can it work, and can we verify who it is.” Google’s Gemini 4 Argon pushed the frontier from single-turn answers toward 24-hour autonomous execution and shipped a risk-threshold report before launch. Meanwhile Character.AI’s bankruptcy put user data in front of 40 state attorneys general, major sites started blocking agents outright, and a verified badge amplified a deepfake. Liu Gong thinks the winners of the next procurement cycle will be whoever turns long-horizon delivery and verifiable identity into purchasable engineering, not just higher benchmark scores.
Signal 1: Google ships Gemini 4 Argon, aiming long-horizon reasoning at hard science (Oct 13). The benchmark table is not the point; the 24-hour continuous autonomous execution spec is. Paired with 2M-token input and 200K-token output — roughly double the single-shot output ceiling of its frontier peers — Argon can emit a full module of code or a long experiment report in one pass instead of stitched continuations. Liu Gong expects at least two top labs to headline “continuous autonomous working hours” within six months, and static benchmark suites to be sidelined in enterprise procurement within a year.
Signal 2: Gemini 4 Argon puts Google back in the frontier’s first tier (Oct 13). Argon will not move market share in three months, but Liu Gong thinks it likely returns Google to the top three of enterprise agent shortlists before Q1 2027 — provided Vertex AI’s price sheet and latency hold up, not just launch-day scores. The underrated line is the pledge to cut hedging and repetition: longer outputs raise downstream parsing cost, and post-training against verbosity moves safety from phrasing back into capability, which compounds across multi-step agent workflows.
Signal 3: DeepMind publishes a Gemini 4 Argon safety report and risk thresholds (Oct 13). The content matters less than the placement: it landed before the model went live. Liu Gong predicts at least two leading labs will fold comparable evaluation summaries into their release pipeline within six months; if that fails to happen, competitive pressure still outweighs compliance pressure. Against DeepMind’s own Frontier Safety Framework — three evaluation categories in May 2024, four-plus by February 2025 — the doubled taxonomy in a year shows the risk list is still mutating, and verifiability remains thin.
Signal 4: Character.AI goes bankrupt as 40 state AGs demand user-data protection (Oct 14). Within six months, any buyer of Character.AI’s user data will almost certainly have to attach data-isolation, purpose-limitation and deletion clauses to the deal to clear state review. What makes this more than a bankruptcy squabble is that a companion-AI product is being forced to answer, in court, whether conversation data counts as a liquidatable asset. Chat histories, character cards, affinity scores and memory vectors are tightly coupled: split them apart and much of the value evaporates — the same arithmetic Liu Gong sees in industrial inspection, where re-labelling and de-identification across factories often cost more than the defect images are worth.
Signal 5: Nvidia dials back DLSS 5’s face generation, launching next month with Hogwarts Legacy (Oct 13). Within six months DLSS 5 will most likely not be on by default in AAA titles, but offered as an optional strength or developer preset — Nvidia first has to prove it does not alter character identity. In early demos a 15-year-old was rendered older, suggesting the model mistook “realistic skin texture” for image-quality gain. Lowering intensity means down-weighting the generative prior so the super-resolution network leans more on source pixels and on engine-supplied geometry and materials. Industrial inspection shows the same failure mode: beautifying a defect texture causes misses outright.
Signal 6: Microsoft tunes Windows for Nvidia’s first RTX Spark laptops (Oct 15). Over the next six months the shipment centre of gravity should sit with developer and creator machines above 15,000 yuan, not volume consumer laptops; the real verdict waits until the second-generation platform and the Windows scheduler have bedded in, around H2 2027. Unified memory is the solid part: per the published DGX Spark specs, GB10 offers 128GB of unified LPDDR5X, 273GB/s of bandwidth and 1 PFLOPS of sparse FP4. Against Apple’s M4 Pro at 273GB/s and M4 Max at 546GB/s, the first RTX Spark has no bandwidth edge — its hand is operator coverage and CUDA compatibility.
Signal 7: Hugging Face releases an Indic pretraining recipe with 2.8T tokens (Oct 16). Within 12 months the number of open Indic base models will double, but fewer than five will see real industrial use — and the deciding factor is reproducible data pipelines, not architecture. Treat 2.8 trillion tokens with a discount: split across 22 languages, per-language yield may still be under a tenth of a comparable English model. The real breakthrough is the open pipeline, letting researchers re-mix by language. Mix ratio beats raw volume — in inspection projects, tripling labels while skewing class balance has pushed accuracy down.
Signal 8: Major sites block AI agents en masse, stalling automated checkout and ticketing (Oct 17). Within 12 months, holding a site-issued verifiable identity will matter more to an agent’s usefulness than its model scores. General-purpose agents without signed agreements and paid terms will likely fall below a 50% success rate in high-value contexts such as e-commerce checkout and ticketing, while the few players with authorisation deals get a de facto exclusive door. Web Bot Auth turns “who am I” into a cryptographically verifiable claim through HTTP message signing; without an identity layer, behaviour detection is all that is left — and its marginal cost always exceeds signature verification.
Signal 9: France deploys agentic computing across industries, extending Nvidia’s reach to aerospace (Oct 13). The value here is less about model capability than about turning agents from demo material into a purchasable engineering component. Liu Gong expects at least two European countries to copy Paris’s three-part formula — summit pledges, a domestic ten-thousand-GPU cluster, anchor industry customers — within 12 months, though most will manage only the first two parts. Agentic workloads look nothing like chat: one task becomes dozens of tool calls and state rollbacks, multiplying KV-cache and scheduling overhead. GB300’s 288GB of HBM3e and roughly 8TB/s per card, over triple the H100’s 80GB and 3.35TB/s, is not a bonus — it is the bar for finishing long-horizon tasks on a single node.
Signal 10: Mia Goth accuses Instagram of letting AI forgeries of her spread (Oct 13). Verification badges are degrading from identity credential to decoration, and platforms’ reliance on them will convert into measurable litigation and compliance costs within 12 months. The heart of the Goth case is not synthesis but that forged content wore an official verification mark — the platform effectively endorsed the fake. The industry keeps framing this as “detect what is real,” yet synthetic quality has already crossed the naked-eye threshold, so the engineering path is provenance: C2PA-style credentials signed at the capture device or generation tool and checked at distribution, rather than guessed after the fact.