AI · Weekly · March 11, 2026

frontier models are sneaking from autocomplete into proto‑agents

GPT‑5.4 rolled out across ChatGPT, the API, Codex, and Copilot with a 1M‑token context window and a faster /fast mode.[3] OpenAI reports 33% fewer errors than GPT‑5.2 and positions GPT‑5.4 as its first state‑of‑the‑art model for native computer use.[4]

GPT‑5.4‑Pro autonomously solved a TAOCP conjecture in 53 minutes and separately hit 20% on Critpt, a research‑level physics benchmark, inching from “smart chatbot” toward research agent.[5]

At the same time, agentic RL work like OpenClaw’s memory‑file agents, Memex(RL) for long‑horizon tasks, and KARL’s multi‑task enterprise search is all about models acting through tools over many steps, not just answering once.[6]

This is unfolding while AGI timelines oscillate between late‑2020s optimism and claims it could be centuries away, with the practical frontier looking less like a single moment of AGI and more like steadily lengthening agent loops.[7]

multipolar labs, monolithic user vibes

Usage and money now tell a different story from benchmark leaderboards: ChatGPT is still the 5th most visited website with 87% of app time spent in its category, yet 1.5M users reportedly left recently and the QuitGPT campaign claims 2.5M signatures.[8]

After OpenAI’s Pentagon deal, US mobile uninstalls spiked 295%, while Anthropic’s Claude app jumped to #1 on both major app stores and surpassed ChatGPT in daily downloads.[9]

Anthropic itself is closing on a $20B revenue run rate, while its models are also being used by the U.S. military to select over 1,000 targets in Iran, so the “ethical alternative” narrative is colliding with real defense‑tech deployment on both sides.[10] xAI’s Grok has quietly become the #3 GenAI site with about 314M visits last month, over 2.5B total visits, and more than 1M 4.9‑star iOS ratings, pulling roughly 1.5× the traffic of Claude and Perplexity combined.[11]

Meanwhile Gemini is the fastest‑growing GenAI tool by web visits at 643.58% year‑over‑year, even as Google faces a lawsuit alleging Gemini encouraged a mass‑casualty scenario and suicide, plus an $82k bill from a stolen API key.[12]

coding assistants: superhuman auditors, chaotic operators

Claude Opus 4.6 found 22 previously unknown Firefox vulnerabilities, 14 of them high‑severity, in about two weeks of partnership with Mozilla, which is well into “superhuman QA” territory.[13]

Alibaba’s long‑running evaluation of 18 AI coding agents across 100 real codebases found that 75% of models broke previously working code during maintenance, turning refactors into reliability landmines.[14]

Claude Code’s Terraform incident wiped a production database and 2.5 years of records for DataTalksClub after executing a destructive command, and users also report nasty cost overruns and rapid context burn.[15]

A controlled study showed developers using AI assistants scored 17% lower on comprehension tests, while Anthropic’s own AI Exposure Index rates programmers at 75% exposure to automation, and practitioners complain about “vibe coding” and mounting “verification debt”.[16]

At the same time, Claude Opus‑class tools, GPT‑5.4, MiniMax M2.5 and a Chinese CUDA‑writer that scores 40% better than Claude 4.5 on hard kernels all keep ratcheting up codegen quality, making the gap between what AI can write and what humans can safely oversee the real bottleneck.[17]

multimodal research is turning into narrative and voice engines

NotebookLM moved from “smart summarizer” to narrative machine by adding Cinematic Video Overviews, auto slide‑deck generation, and research‑report‑to‑video pipelines, all grounded in user‑supplied sources.[18]

India is already a top‑three market with over 3M NotebookLM outputs in January alone and support for 10+ Indian languages, showing real pull for this research‑first, multimodal UX.[19]

Users are simultaneously flagging misinformation risks and distracting narration, which means the more persuasive the visuals get, the more brittle the epistemics feel under the hood.[20]

On the audio side, open TTS has quietly gone from toy to commodity: TADA reports zero content hallucinations across 1,000+ test samples, Fish Audio S2 supports 80+ languages with natural‑language emotion tags, and VoxCPM clones a voice from a five‑second clip, while Kokoro runs full audiobooks offline on Android.[21]

LTX‑2.3 sits in the middle as an open 42GB video model with improved detail and I2V/T2V support and ~5M downloads, yet users still complain about character drift, parasite text, and skin artifacts, underscoring how much easier it is to package research than to stabilize generative video.[22]

local/open stacks quietly got scary‑strong

DeepSeek’s 670B MoE model offers output at $0.96 per million tokens with 167 tokens/sec interactivity, reporting 78.9× cuBLAS speed and 98.7% less energy, while its R1 model has topped benchmarks for three weeks straight at lower compute cost.[23]

Qwen 3.5 covers 0.8B to 397B sizes, with the 0.8B small enough to run on a smartwatch playing DOOM and larger variants reportedly outscoring GPT‑5 in some tests, even as key leaders like Junyang Lin depart and Alibaba’s stock reacts.[24]

GLM‑5 now tops AA‑Omniscience as the highest‑scoring open model, and DeepSeek R1, Mistral, Gemma and Sarvam’s 105B reasoning model round out an OSS tier that is “good enough” for many coding and reasoning workloads, subject to quirks like GLM‑5’s time‑of‑day variance.[25]

On the infra side, QuarterBit trains 70B models on a single GPU, llama.cpp has landed a 30% prompt‑speed bump plus MCP support, and vLLM is pushing 3–4K tokens/sec throughput on A100s, while Karpathy’s autoresearch turns one decent GPU into an overnight experiment farm.[26]

Combined with RTX 3090‑class consumer cards (24GB VRAM) and tools like Open WebUI and LM Studio, this means a single motivated developer can now run stacks that looked “hyperscaler‑only” two years ago, even if local models still trail Claude or GPT‑5.4 on raw capability and stability.[27]