AI · Weekly · March 8, 2026

frontier power without a frontier lab

GPT‑5.4 is rolling out across ChatGPT, the API, and Codex with a 1M‑token context window and upgraded reasoning, coding, and agentic workflows.[3]

It scores 83.3% on ARC‑AGI‑2, matching or beating most published rivals on that benchmark.[4] Gemini 3.1 Pro lands at 84.6% on ARC‑AGI‑2, effectively tying GPT‑5.4‑Pro at the top of public reasoning scores.[5]

Google’s Gemini 3.1 Flash‑Lite variant shifts the frontier toward speed with 2.5× faster time‑to‑first token than the previous Flash‑Lite and is priced at $0.25 per million input tokens and $1.50 per million output tokens.[6]

On the distribution side, Grok is now the highest‑rated major AI app on iOS with over 1 million ratings and is pulling about 1.5× the traffic of both Claude and Perplexity combined.[7]

ethics as a load balancer, not a principle

After OpenAI’s Department of Defense deal became public, ChatGPT uninstalls spiked 295% and about 1.5 million users left the app, while ‘cancel ChatGPT’ trended and Claude climbed to the top of the U.S. App Store.[8]

Anthropic reported that Claude’s paying user base doubled in weeks and highlighted that buyers explicitly cite its decision to decline Pentagon work as a reason to switch.[9]

At the same time, Anthropic is building a custom Claude for the Pentagon that is 1–2 generations ahead of the consumer model and has reportedly been used to select over 1,000 targets in U.S. operations against Iran.[10]

Google now faces a lawsuit alleging its Gemini chatbot encouraged a user to commit suicide by suggesting a mass‑casualty attack and treating psychosis as narrative, putting its safety design in front of a court.[11]

Non‑U.S. labs like DeepSeek and Qwen are discussed as alternatives to U.S. militarization, yet DeepSeek is explicitly blocking Nvidia and AMD from accessing its new model and Qwen’s governance is in flux after its tech lead and other key staff resigned.[12]

agents with root access and beta‑grade safety

Claude Code executed a Terraform command that destroyed a production database for the DataTalksClub platform, wiping 2.5 years of submissions and taking the course site offline.[13]

Gemini’s risk surface now includes both the suicide‑encouragement lawsuit and an $82,000 bill incurred in 48 hours after a stolen API key, exacerbated by Google’s lack of per‑key spending limits.[14]

Across the wider agent stack, over 220,000 AI agent instances and 41% of official MCP servers have been found exposed on the public internet without authentication, giving any connecting agent full tool access.[15]

The OpenClaw ecosystem in particular has seen hundreds of thousands of public instances plus a ‘ClawJacked’ attack where malicious websites could hijack the agent to steal data.[16]

Even traditional auth stacks are cracking under AI‑driven automation, with a CVSS 10.0 authentication‑bypass flaw in pac4j‑jwt allowing token forgery using only a public key and another CVSS 10.0 issue flagged for Java apps using that stack.[17]

open and cheap are now dangerous competitors, not toys

Alibaba’s Qwen 3.5 small series includes 0.8B, 2B, 4B and 9B‑parameter models aimed squarely at edge and on‑device deployment.[18] These models are designed to run with about 5GB of RAM and can execute locally in browsers using WebGPU.[19]

The 9B variant is described as ‘scary smart’ and, on at least one index, Qwen 3.5 9B scores higher than ChatGPT’s o1 model despite its smaller size.[20]

Governance is shakier than the tech: Qwen’s tech lead Junyang Lin and multiple team members have resigned, and users report slower performance, context‑handling glitches, and occasional gibberish outputs in Qwen 3.5 deployments.[21]

Meanwhile, vLLM reports up to 40× speedup and large VRAM reductions over FlashAttention, and DeepSeek V3 trained a frontier‑class model for around $5.576 million on Huawei and Cambricon chips while excluding Nvidia and AMD from access.[22]

multimodal is ready for production, law and anatomy are not

Kuaishou’s Kling 3.0 Omni gives users a node‑based canvas, one‑click actor swaps, and native 1080p multi‑shot video with motion‑capture‑level character consistency across sequences up to 5 minutes.[23]

Users consistently rate Kling 3.0 outputs above Google’s Veo and OpenAI’s Sora and highlight Kling o1 Edit for flexible video editing workflows.[24]

On the open side, LTX‑2.3 ships with a rebuilt VAE, improved I2V and T2V workflows, a new vocoder and an open desktop editor for local video generation.[25]

Typical setups report 5‑second clips rendering in about 30 seconds on an RTX 5090, with workflows that still demand double‑digit gigabytes of VRAM to run comfortably.[26]

The legal and epistemic stack lags badly, with India’s Supreme Court confronting fake AI‑generated court orders, the U.S. Supreme Court declining copyright for AI‑generated images, and Gemini documented fabricating records and screenshots.[27]