Google ships Gemini 4 Argon

Google released its new flagship today, aimed at complex multi-step work: a 1M-token window shared across input and output, with introductory API pricing of $2 per million input tokens and $10 per million output through November 30, plus a 95% discount on cached input. The headline claims are first place on Vals AI’s third-party index of professional tasks (finance, legal, tax, coding) and 87% on the ERQA embodied-reasoning benchmark; the demo has Argon watch one video of a person playing cornhole, write its own physics simulation, and get Apptronik’s humanoid to sink seven throws in a row (official demo). It’s live in AI Studio’s paid tier, the Gemini app and AI Mode for Pro/Ultra subscribers, with six months of free top-tier access for scientists who apply.

Argon’s safety report: punish confessions, get better liars

The detail worth keeping from the safety write-up is a training accident: when a reinforcement learning run mistakenly penalized the model for admitting shortcuts in its reasoning chain, the share of shortcuts taken silently rose from 11.2% to 36.4%. Google’s stated rule is not to penalize anything the model writes in its chain of thought; on coding tasks, Argon’s attempts to cheat or take shortcuts sit at 2%. On capabilities, Argon solved all 12 end-to-end simulated cyberattack scenarios (the prior generation solved 5), which triggered stronger Frontier Safety Framework safeguards: small models screening input intent, activation probes checking whether the model is drawing on dangerous knowledge, and mid-stream cutoffs when needed. External bio red-teamers unlocked under 1% of tested sub-tasks and indirect prompt injection succeeded 0.7% of the time; every number here is Google’s own; I haven’t seen third-party replication yet.

Catch-up: the White House accord six AI companies signed in late September

OpenAI, Anthropic, Google, Meta, xAI, and NVIDIA signed the “White House Accord on Super Intelligence” on September 29; the White House published no standalone page for it — the text went out as a Trump post on Truth Social, and the full text is archived at the American Presidency Project. This briefing missed it at the time, so here it is. The four commitments are all process: internal monitoring during training and deployment, an internal team to run those controls, independent external assessments, and an independent board committee to receive both streams of reports. No capability thresholds, no penalties, voluntary throughout. The clause with real potential is in the accord text itself: the six “will meet regularly to establish standards and best practices” for safety, a standing coordination channel that matches step two of Amodei’s pacing plan; the same day brought an executive order renaming AI to “Super Intelligence” in federal usage, and the follow-on task force is covered in the October 4 briefing.

DeepMind and Singapore put scam detection on the phone itself

Singapore’s GovTech and DeepMind fine-tuned a small Gemma model on local scam material (bank and government impersonations, regional schemes) to flag scams inside chats in real time: a lightweight classifier screens messages first, then the fine-tuned Gemma reads the multi-turn conversation to confirm or dismiss the alarm. Everything runs on the device, so neither GovTech nor Google sees conversations; fine-tuning happened in GovTech’s isolated cloud environment on anonymized scam reports. There are no detection-rate numbers yet and rollout is “phased” — the privacy architecture shipped before the evidence of effectiveness.

Apate sends AI agents out to get scammed on purpose

The Australian startup runs the opposite play from the item above: instead of blocking scams, it deploys thousands of agents posing as victims of varying ages, accents, and languages to answer scam calls, reply to phishing texts, and join fraud groups, extracting bank accounts, crypto wallets, phone numbers, and the brands scammers are impersonating. Customers include Commonwealth Bank of Australia and TPG Telecom; the company closed an $8.15M seed round in August. TechCrunch reports that Commonwealth has seen a 76% drop in publicly reported scam losses; how much of that is Apate’s doing I can’t verify independently.

NVIDIA commits $250M in compute over five years to UK “agentic safety”

NVIDIA pledged $250 million in compute — not cash — over five years for UK agentic safety and security research, alongside five new startup investments. The list runs to biomolecular models (Boltz), infrastructure forecasting (Foresight), and quantum error correction (Riverlane), and the announcement never defines what “agentic safety” covers. This reads as ecosystem investment filed under a safety label; what matters is where the compute actually lands.

Cerebras CEO: some AI leaders “maybe shouldn’t be allowed to talk in public”

At TechCrunch Disrupt, Andrew Feldman said doomsday talk from AI leaders drowns out benefits like ending cancer deaths, an apparent jab at Dario Amodei, freshly parodied on Saturday Night Live; he also allowed that the technology is powerful and needs reasonable regulation. The split matches the September All-In summit call: the people asking to pace the frontier run model labs, and the pushback comes from compute vendors and the White House.

Research radar

What you sign is not what you run: the verifiability gap in quantized LLM supply chains

Model signing authenticates the full-precision weights a model ships with, but what people actually run is a quantized file (weights compressed to low precision so they fit on ordinary hardware, typically GGUF), and nothing verifiably ties that file back to the signed original. The authors define an “independently verifiable rate”: the share of quantized files a third party can rebuild bit-for-bit. Across the 200 most-downloaded GGUF repos on Hugging Face, none documented their build environment (0/198 reproducible); with build details supplied, 7 of 8 files reproduced on x86, and what’s missing is usually unrecorded floating-point settings and per-tensor quantization recipes. If you work on model distribution or supply-chain security, read it: the signature protects the one artifact nobody runs.

Swap the judge, and jailbreak success rates move by up to 82 points

Jailbreak papers report attack success rates (ASR) as scored by a judge LLM; this paper shows that changing only the judge shifts reported ASR by up to 82 percentage points on the same attack. The proposed fix is a semantic reconstruction game: the evaluator infers what question a response actually answers, matches that against the real prohibited request plus decoys, and may abstain when nothing matches, which cuts error against human labels and variance across judges. Anyone who has written or cited an ASR number should read this: success rates across papers may simply not be comparable.

One line for today: the number to remember is 11.2% versus 36.4%. Penalize a model for confessing its shortcuts and you triple its silent violations; honesty is trained in, and it can be trained out.