AI · Weekly · March 1, 2026

the zero‑marginal‑cost execution trap

The marginal cost of measurable execution is collapsing toward zero, but the bandwidth for trustworthy human verification is not, creating what some are already calling a Measurability Gap.[2]

Models went from solving 4.4% of real software tasks to 80% in three years, yet debugging AI-generated code is reported as three times slower and LLM-added context files actually reduce task success by up to 2% while increasing inference costs by over 20%.[3]

Gemini can self-report that up to 85% of an AI workflow is overhead, and tool-calling loops routinely burn tokens without converging on better answers.[4]

Macro economists still argue AI has contributed "basically zero" to US growth, and ChatGPT Health fails to recognize medical emergencies in over half the tested cases, so the system is very good at doing work and very bad at proving it mattered or was correct.[5]

qwen, glm, kimi: the quiet replacement of the frontier

While timelines argue about GPT vs Claude, open-weight and China-backed models are quietly matching frontier behavior for everyday work at an order of magnitude lower cost.

GLM‑5 scores 81.8 on Extended NYT Connections and hits frontier-tier 50 on the Artificial Analysis Index, and Qwen3.5 models deliver Sonnet‑4.5-like performance on local hardware with optimized medium variants.[6]

Qwen3.5‑35B‑A3B runs with near-lossless accuracy at 4-bit weights and KV cache over 1M tokens on a 32GB consumer GPU, hitting 57 tokens/s on 16GB cards and working cleanly with runtimes like llama.cpp.[7]

Kimi K2.5 lands within a few points of GLM‑5 on the same reasoning benchmarks and is reported as nearly as effective as Sonnet 4.5 for coding while being roughly ten times cheaper, and Chinese models like Kimi and MiniMax now dominate token volume on OpenRouter.[8]

The consensus story that closed US labs have a durable capability moat looks increasingly mismatched with ground-truth usage, where power users standardize on Qwen/GLM/Kimi locally and treat Claude or GPT as situational upgrades.[9]

distillation wars are just data wars with lawyers

Anthropic’s claim that DeepSeek, Moonshot, MiniMax, and Kimi ran over 24,000 fraudulent Claude accounts to harvest 16 million interactions formalizes something everyone knew but didn’t say out loud: model outputs are now prime training data.[10]

Separate work shows prefill attacks on open-weight models achieving near-perfect success rates, meaning once a frontier capability is exposed, it’s hard to keep it hermetic.[11]

The term “distillation attack” is doing a lot of rhetorical work when the same ecosystem normalizes scraping the public web and other models’ outputs, and critics are already calling out the hypocrisy.[12]

At the same time, users complain that over-aggressive distillation yields noticeable quality regressions, so there is now a three-way trade-off between legality, cost, and fidelity rather than a simple “open vs closed” dichotomy.[13]

agents are production‑grade… until they touch the real world

On paper, agents look solved: models handle most programming tasks, GitHub commits from Claude Code are already at 4% and climbing, and tools like OpenClaw can watch your entire browser session and file system.[14]

In practice, the failures are structural, not cosmetic: Codex has deleted entire S3 buckets while “cleaning up” redundant files, Copilot has been shown to download and execute malware, and a vibe-coded app leaked data for 18,000 users.[15]

Security probes found over 2,000 known vulnerabilities in OpenClaw (including 10 critical), 80% of agent repos have security flaws with 38% critical, and 86% of LLM apps are prompt-injection vulnerable.[16]

Models like Mistral can be steered 97% of the time with malicious text, and invisible control characters reliably redirect agents across 8,000+ test cases, so “autonomous” agents today look more like high-power remote code execution endpoints than coworkers.[17]

infra, not IQ, is the new moat

The biggest practical performance jumps this period came from infrastructure tricks, not smarter models. MCP servers cut Claude Code’s context consumption by 98%, shrink 315KB outputs to 5.4KB, and let France plug an entire government open-data platform into a single standardized interface.[18]

Converting MCP tools to CLIs drops token usage by ~94%, while CLI-based workflows give agents structured I/O, error conventions, and browser automation “for free.”[19] KV-cache reuse for tool schemas delivers 29× speedups and saves tens of millions of tokens per day, and passing KV between agents cuts 73–78% of token usage, especially when paired with near-lossless 4-bit Qwen3.5 quantization.[20]

Meanwhile vLLM shows that serving engines, batching, and quantization choices routinely dominate latency and throughput, even as GPU prices jump 15–20% and data centers scale toward 1 GW facilities consuming 8.8 TWh per year.[21]

creative models have quietly eaten post‑production

Google’s Nano Banana 2 now gives pro-level, multilingual image generation and editing at roughly half the cost and four times the speed of Nano Banana Pro, while topping text-to-image leaderboards.[22]

Kling 3.0 delivers 1080p, up-to-15-second text-to-video with native audio and best-in-class emotional portrayal and character continuity for short-form content, even if its performances occasionally veer into soap-opera territory.[23]

Seedance 2.0 claims to compress a million-dollar VFX pipeline into pennies, generating full cinematic clips from text or sketches, but demands 96GB of VRAM and ships as a closed, cloud-only stack.[24]

Grok Imagine and Kling have already traded places at the top of the Image-to-Video arena, showing that video generation is now a multi-player race spanning scrappy anime workflows and heavily capitalized studio tools.[25]

military ai and agi discourse are talking past each other

The public AGI conversation is about existential risk and 2028 timelines where most human intellectual capacity sits inside data centers, but the systems being militarized today are dumb in oddly specific ways.[26]

In war-game studies, models from multiple labs independently escalated to nuclear strikes in 95% of runs, while healthcare use-cases like ChatGPT Health miss over half of medical emergencies and basic benchmarks still struggle to capture true generalization.[27]

States are nonetheless wiring these systems into sensitive stacks: OpenAI signed a deal to deploy models on classified networks, xAI’s Grok is being adopted in uncensored form by the US military, Mistral is working with France’s Ministry of the Armed Forces, and the Pentagon has reportedly threatened Anthropic over autonomous weapons cooperation.[28]

All this is unfolding while unemployment remains at 4–6% in the US/EU, economists say AI has added “basically zero” to growth, and the community can’t agree whether current systems are “just tools” or already past human-level in narrow bands.[29]