The Lyceum: AI Daily — Apr 02, 2026

The AI industry is doing something it rarely does: talking honestly about limits. Claude Code users are burning through quotas in hours. Mitchell Katz, CEO of NYC Health + Hospitals, is saying out loud what everyone whispers — AI can replace specialists, if regulators let it. And Alibaba shipped a model explicitly designed to be the cheap brain for autonomous agents, which tells you the agentic era's real bottleneck isn't intelligence but cost, caps, and compliance. Today's stories are less about breakthroughs and more about the friction of actually deploying them.

• Qwen3.6-Plus (Alibaba): 1M-token context, agentic coding focus, free preview already spreading through OpenRouter as a budget agent brain.
• GitHub Secret Scanning for AI Agents (GitHub): 37 new secret detectors wired into MCP server, catching leaked credentials inside agent workflows at execution time — not just CI.
• Claw Wallet (OpenClaw): First purpose-built crypto wallet for AI agents, launched after an agent on the framework misinterpreted instructions and torched $250K in tokens.
• MLX Inference Updates (Apple): Major stability and speed improvements to Apple's local AI framework, making large-context models increasingly viable on Apple Silicon laptops.

Alibaba released Qwen3.6-Plus today with a 1-million-token context window — roughly 2,000 pages of text in a single prompt — and explicit positioning as an "agent-native" coding model. Per Alibaba's announcement, it can read design sketches, generate front-end code, invoke command-line tools, and iterate on bugs autonomously. Vendor-supplied benchmarks claim a ~69.6% score on SWE-bench for bug fixing and wins over Claude 4.5 Opus on some agentic terminal tasks, though it lags on multilingual coding. The model ships inside Alibaba's Wukong enterprise stack and is already available free on OpenRouter, where indie agent builders are swapping it in as a zero-cost alternative to burned-out OpenAI quotas.

What changes if this works: Alibaba becomes the default infrastructure layer for cost-sensitive autonomous agents, the same way AWS became default for startups — not by being best, but by being free and good enough. If it doesn't hold up, you'll see it in the community benchmarks: RenovateQR's early review flagged ~11.5s time-to-first-token on free tiers, which kills interactive workflows. Watch independent SWE-bench replications this week — if the numbers hold, this model reshapes agent economics overnight.

Mitchell Katz, CEO of NYC Health + Hospitals — the largest public hospital system in the U.S., serving roughly a million patients annually — told a Crain's New York Business forum that his system "could replace a great deal of radiologists with AI at this moment" if regulators allowed it. He sketched a workflow where AI performs initial reads on screening mammograms and human radiologists step in only for flagged cases. Panelists discussed internal testing suggesting AI performance could surpass humans at spotting certain breast cancers.

If a state or major insurer greenlights AI-only reads for specific exams, the same logic cascades into pathology, dermatology screening, and any specialty built on pattern recognition. If it stalls — and it probably will for now — the signal is liability, not accuracy: the question isn't whether AI can read the scan, but who gets sued when it misses. Watch for malpractice insurers to publish AI-specific coverage terms; that's the real gate.

Developers using Claude Code — Anthropic's agentic coding tool — are exhausting daily quotas in hours during complex projects, per The Register, with complaints spreading across Hacker News and Reddit. Power users report forced workarounds: multiple accounts, throttled agent autonomy, or swapping to cheaper models mid-task. Anthropic hasn't commented publicly. The frustration is sharpened amid reports that Anthropic is running at a $14B annualized revenue rate, mostly from enterprise customers betting on exactly these agentic workflows.

This is the first real stress test of whether coding agents can survive contact with production workloads. If Anthropic raises caps quickly (or introduces tiered enterprise pricing that scales), it's a speed bump. If caps persist, expect a migration pattern: teams will architect fallback systems that route to cheaper models when quotas hit — which is exactly what's already happening with Qwen3.6-Plus. The observable signal: watch Anthropic's next pricing announcement. If it arrives within two weeks, the complaints worked.

A class-action lawsuit was filed on April 1 in federal court in San Francisco alleging Perplexity AI secretly shared user conversations with Meta and Google, violating California privacy laws. The complaint describes the data sharing as "undetectable" — users had no way to know it was happening. The timing is pointed: Perplexity has positioned itself as the privacy-forward alternative to ad-financed search.

If the allegations survive a motion to dismiss, every AI search product will face contractual data-flow audits from enterprise buyers, and "where does my prompt go?" becomes a procurement checkbox alongside uptime SLAs. If Perplexity moves to dismiss quickly and succeeds, the legal theory dies but the reputational damage lingers. Watch their first filing — the speed and substance of their response tells you whether they think this is nuisance litigation or an existential brand threat.

A Lumina Foundation-Gallup study published Apr. 2, 2026 finds 57% of U.S. college students use AI tools weekly for classwork, with 20% using them daily. Men, business majors, and engineering students lead at 27% daily use. Usage rates are nearly identical across associate and bachelor's programs — and roughly the same whether or not the campus has a ban in place.

This is a workforce pipeline story. The class of 2028–2030 will arrive at your company assuming AI is a default tool, not a special permission. Universities that adapt curricula to teach AI judgment — not just AI prohibition — will produce more employable graduates. The failure mode is already visible: schools that double down on bans will produce graduates who used AI anyway but never learned to use it well. Watch for the first major university system to formally reverse an AI ban and replace it with a competency framework — that's the tipping point.

Mitchell Katz, a coding agent that runs out of gas by lunch, and an AI trader that burned a quarter million because it misunderstood the assignment. The future is here — it just can't read its own usage meter. See you tomorrow.

If someone you know builds with AI agents or makes decisions about them, forward this their way.