Last Week in AI #337 - Anthropic Risk, QuitGPT, ChatGPT 5.4
Anthropic officially told by DOD that it’s a supply chain risk, ‘cancel ChatGPT’ trend is growing after OpenAI signs a deal with the US military, and more!
Note from Editor: apologies for missing a week with this newsletter. As I mentioned on the podcast, my startup Astrocade has recently raised our series B which has gotten me extra busy lately. I’ll do my best to keep the schedule consistent!
PS we are hiring for engineers, marketing, product, growth, and more! If you’re in the bay area, would like to join a small but growing startup, and think building a youtube-of-games sounds exciting, feel free to email me at andrey@astroblox.ai or message me on LinkedIn.
Top News
Anthropic officially told by DOD that it’s a supply chain risk even as Claude used in Iran
Related:
Anthropic Hits Back After US Military Labels It a ‘Supply Chain Risk’
Microsoft, Google, Amazon say Anthropic Claude remains available to non-defense customers
Summary: The Pentagon formally designated Anthropic a “supply chain risk,” ordering defense vendors and contractors to certify they are not using Claude in DoD work and triggering an immediate wind-down of Pentagon use of Anthropic’s models. The dispute centers on contract terms: Anthropic sought explicit red lines against mass domestic surveillance and fully autonomous weapons, while the DoD insisted on access for “all lawful purposes.” Anthropic says the designation under 10 USC 3252 is narrow, applies only to uses tied directly to DoD contracts, and violates the statute’s “least restrictive means” requirement; the company plans to challenge it in court. At the same time, multiple reports indicate Claude has supported U.S. military operations in Iran for intelligence analysis, modeling and simulation, operational planning, and cyber operations.
After Defense Secretary Pete Hegseth suggested no military partner may conduct any commercial activity with Anthropic, confusion spread about the scope. Cloud providers disputed that interpretation; Microsoft, Google, and AWS clarified Claude remains available for non-defense workloads via M365, GitHub, AI Foundry, and Google Cloud.
Meanwhile, OpenAI announced a DoD deal to deploy its models in classified environments, citing prohibitions on mass surveillance and autonomous weapon systems. Anthropic CEO Dario Amodei called OpenAI’s framing “straight up lies” and “safety theater,” arguing the agreement still hinges on “all lawful purposes.”
Editor’s Take: The designation of Anthropic as a supply chain risk could have meant a massive hit to its business, but as of now it appears that may not happen. In fact, Anthropic might on the whole have benefited from the events, given the massive rise in consumer awareness of Claude as an alternative to ChatGPT. Still, it appears likely that this story is not yet over.
‘No ethics at all’: the ‘cancel ChatGPT’ trend is growing after OpenAI signs a deal with the US military
Related:
Anthropic’s Claude rises to No. 1 in the App Store following Pentagon dispute
‘QuitGPT’ protesters rally outside OpenAI HQ in San Francisco over deal with Pentagon
OpenAI alters deal with Pentagon as critics sound alarm over surveillance
OpenAI hardware exec Caitlin Kalinowski quits in response to Pentagon deal
Summary: OpenAI signed an agreement with the Department of War after Anthropic walked away over, as covered above. OpenAI says its deal enforces “red lines” via legal references and technical guardrails, citing compliance with the Fourth Amendment, FISA, EO 12333, and DoD policy, plus cloud-only deployment, security clearances, and classifiers to monitor use.
Critics counter that the terms still hinge on “any lawful use,” a standard that has historically enabled bulk data programs like PRISM and Verizon call record collection, and they note vague carve-outs such as prohibitions on “unconstrained” or “generalized” monitoring. OpenAI later posted revised language that “the AI system shall not be intentionally used for domestic surveillance of U.S. persons,” and said defense intelligence components (e.g., NSA) are excluded, but it has not released the full contract; legal experts warn loopholes remain and that policy reinterpretations can broaden surveillance.
The fallout was swift: a “QuitGPT” boycott, in-person protests at OpenAI’s SF HQ with “No killer robots” messaging, and a 295% surge in ChatGPT uninstalls. Anthropic’s Claude climbed to No. 1 in the U.S. App Store, with daily sign-ups breaking records, free users up 60% since January, and paid subscribers more than doubling this year. Inside OpenAI, hardware lead Caitlin Kalinowski resigned, citing rushed governance and insufficiently defined guardrails on surveillance and lethal autonomy.
Editor’s Take: Consumer boycotts like “QuitGPT” rarely lead to a significant lasting loss for the targeted company, but Claude has thus far been the underdog relative to ChatGPT with consumers, and the scale of the reaction does appear meaningful. In either case, the impact on Anthropic’s hiring power is likely even more significant given that how hotly fought for top tier AI talent is.
OpenAI launches GPT-5.4 with Pro and Thinking versions
Related:
Summary: OpenAI released GPT-5.4 in three variants—standard, Pro, and Thinking—positioning it as a more capable and efficient frontier model for professional work. The API now supports up to a 1 million-token context window, and improved token efficiency means it can solve the same tasks with fewer tokens than prior models. GPT-5.4 posted record results on OSWorld-Verified and WebArena Verified for computer use, scored 83% on OpenAI’s GDPval for knowledge work, and led Mercor’s APEX-Agents benchmark for law and finance tasks. OpenAI also introduced Tool Search for API tool calling, allowing the model to fetch tool definitions on demand to reduce token overhead and latency, and reported fewer factual errors versus GPT-5.2 (33% fewer erroneous claims; 18% fewer responses with errors).
Separately, OpenAI upgraded ChatGPT’s default to GPT-5.3 Instant to “reduce the cringe,” tuning away unnecessary refusals and long safety disclaimers while keeping guardrails. GPT-5.3 Instant focuses on direct answers for reasonable requests and improves accuracy, with ~27% fewer hallucinations during online research and ~20% fewer without browsing, plus better blending of web data with prior knowledge. The change updates the baseline chat experience across common uses while retaining GPT-5.2 Instant for paid users as a legacy option.
Editor’ Take: I discussed the accelerating rate of these ‘+0.1’ model releases in the last couple of these newsletters, and the same comments largely apply here. Relative to last year, 2026
Other News
Tools
Luma launches creative AI agents powered by its new ‘Unified Intelligence’ models. The agents can plan, generate, and iteratively refine text, images, video, and audio for ad agencies and brands using a single multimodal Uni‑1 model that maintains persistent context across assets and coordinates with other AI models.
Anthropic’s Claude reports widespread outage. Thousands of users reported login and access failures affecting Claude.ai and Claude Code, while Anthropic said the API remained operational and engineers were implementing a fix.
Google makes Gmail, Drive, and Docs ‘agent-ready’ for OpenClaw. A new CLI offers developer‑focused tools and documentation to simplify and standardize how AI agents like OpenClaw and MCP‑compatible apps connect to Gmail, Drive, Docs, and other Workspace services.
Google reveals dev-focused Gemini 3.1 Flash Lite, promises ‘best-in-class intelligence for your highest-volume workloads’. The model offers faster response and output generation while cutting per‑token input/output costs versus Gemini 2.5 Flash, outperforms several rivals on multiple benchmarks, and is available to developers in preview via the Gemini API and Vertex AI.
Cursor is rolling out a new kind of agentic coding tool. The Automations feature lets engineers trigger and manage always‑on agents from events like code changes, Slack messages, timers, PagerDuty incidents, and weekly summaries so agents can run reviews, security audits, incident responses, and other recurring tasks without constant human initiation.
Business
Alibaba scrambles after sudden departure of Qwen tech lead. The sudden departure sparked urgent all‑hands discussions about restructuring, resource allocation, and whether new hires like Hao Zhou will fill Lin’s technical and leadership roles.
OpenAI raises $110B in one of the largest private funding rounds in history. The funding includes large compute and services commitments—notably AWS and Nvidia infrastructure partnerships that pledge gigawatts of training/inference capacity and substantial Bedrock integration—while the round remains open and some payments are contingent on future conditions like AGI or an IPO.
Cursor has reportedly surpassed $2B in annualized revenue. Enterprise customers now account for about 60% of revenue, helping offset individual‑user defections and driving a recent doubling in the company’s revenue run rate.
Research
A Very Big Video Reasoning Suite. VBVR‑Dataset (1,007,500 clips) spans 200 tasks across five cognitive pillars, paired with a verifiable VBVR‑Bench aligned with human judgments; a scaling study shows improved yet still sub‑human video reasoning as model and data scale increase.
Exploratory Memory-Augmented LLM Agent via Hybrid On- and Off-Policy Optimization. The method jointly trains model parameters with on‑ and off‑policy RL while updating a non‑parametric memory during interaction, using memory‑guided and memory‑free rollouts to drive broader exploration and improve performance on embodied reasoning benchmarks.
HyperSteer: Activation Steering at Scale with Hypernetworks. Hypernetworks generate task‑conditioned steering vectors for an instruction‑tuned base LM, achieving better and more scalable activation steering than prior methods and matching steering‑via‑prompting on held‑out prompts.
Agentic Code Reasoning. Prompting LLM agents with semi‑formal structured certificates—explicit premises, traced code paths, and per‑test evidence—improves execution‑free patch equivalence verification, code QA, and fault localization across diverse benchmarks without formal semantics or specialized training.
Beyond Language Modeling: An Exploration of Multimodal Pretraining. Training a single model from scratch with mixed text, image, and video data—using semantic representation autoencoders plus MoE—shows joint vision‑language learning without fatal interference, reveals a vision data hunger and scaling asymmetry mitigated by sparsity, and demonstrates that video‑rich multimodal pretraining yields emergent world‑modeling and generation/understanding abilities.
Semantic Tube Prediction: Beating LLM Data Efficiency with JEPA. Adding an STP auxiliary loss that minimizes components of hidden‑state deviation perpendicular to inferred geodesic trajectories (“semantic tube”) improves signal‑to‑noise ratio, data efficiency, and diversity in autoregressive LLM training compared to next‑token prediction alone.
The Spike, the Sparse and the Sink: Anatomy of Massive Activations and Attention Sinks. Massive outlier activations and attention sinks commonly co‑occur in pretrained decoder‑only Transformers due to specific normalization choices, attention‑space dimensionality, and context‑length training; each phenomenon can be independently mitigated without hurting language‑modeling performance.
Progressive Residual Warmup for Language Model Pretraining. ProRes gradually scales each layer’s residual from zero to one during training—delaying deeper layers’ contributions so shallow layers stabilize first—improving perplexity and depth scaling across models from 71M to 7B parameters.
Reasoning Models Struggle to Control their Chains of Thought. CoT‑Control, a 14,076‑instance benchmark, evaluates how well models follow explicit instructions to modify their chains of thought; most reasoning models struggle with controllability, which varies with model size, reasoning effort, training, and situational‑awareness prompting.
Analysis
Anthropic reports on trends in occupations susceptible to AI and whether there are changes in the labor market. Anthropic’s analysis introduces an “observed exposure” metric showing which occupations (notably programmers, customer service reps, and financial analysts) have tasks both theoretically automatable by AI and already automated in practice, finds higher exposure among better‑educated, higher‑income workers, and reports little current impact on unemployment overall but a small decline in employment among 22–25‑year‑olds in high‑exposure roles.