Intellēctus — April 2, 2026

Today is a day of leaks, launches, and competitive pivots: Anthropic's accidental Claude Code source exposure dominates developer conversation, Google drops Gemma 4 into the open-weight arena, and Microsoft makes its long-telegraphed "superintelligence" play with three new foundation models. Buckle up.

Industry Moves

Microsoft's MAI unit goes public with three foundation models. Six months after forming its Microsoft AI (MAI) group under CEO Mustafa Suleyman, the company is putting products on the table: a voice-to-text transcription model, an audio generation model, and an image generation model. The play is explicitly enterprise-focused — Suleyman has reframed his job as turning frontier research into business value, not chasing AGI headlines. (Microsoft takes on AI rivals with three new foundational models · Microsoft's new 'superintelligence' game plan is all about business)

OpenAI acquires podcast network TBPN. The deal is raising eyebrows across the industry: OpenAI is reportedly hemorrhaging developer mindshare to Anthropic and Claude Code, yet is spending to acquire media properties. Critics are calling it a "slop strategy" — buying distribution while struggling to retain technical users. (OpenAI Acquires TBPN)

OpenAI's funded children's safety coalition surfaces. Reporting from SF Standard and Gizmodo reveals that the Parents and Kids Safe AI Coalition — a group pushing age-verification legislation in California — was entirely funded by OpenAI, without the knowledge of many member child safety organizations. The disclosure raises questions about astroturfing in AI policy. (Child safety groups say they were unaware OpenAI funded their coalition · Group Pushing Age Verification for AI Turns Out to Be Backed by OpenAI)

Open Source & Model Releases

Google drops Gemma 4 under Apache 2.0. Google DeepMind released two new open-weight models today: Gemma 4 31B (dense, 256K context, redesigned architecture) and Gemma 4 26B A4B (MoE, 26B total / 4B active per token). Both are commercially available on Hugging Face and Ollama. Early benchmarks show Gemma 4-E4B hitting ~400 tokens/sec on an M5 MacBook — roughly 4× faster than Claude Code's inference at 90 t/s. A separate inference stack report shows a 15% throughput gain over vLLM on NVIDIA B200 and AMD MI355X hardware. (Google has published its new open-weight model Gemma 4 · Gemma 4 running on NVIDIA B200 and AMD MI355X)

Research & Benchmarks

LLMs still can't count reliably. A new arXiv paper, "Even GPT-5.2 Can't Count to Five: Zero-Error Horizons in Trustworthy LLMs," argues that current frontier models fail at elementary counting and arithmetic tasks with non-trivial error rates, framing this as a fundamental trustworthiness problem rather than a benchmark edge case. (Even GPT-5.2 Can't Count to Five)

PhAIL benchmark exposes robot AI's real-world gap. A new open benchmark (phail.ai) tested VLA (vision-language-action) models on real commercial hardware — not simulation, not cherry-picked demos. Best result: 5% of human throughput, with the agent needing intervention every ~4 minutes. A stark reminder that physical AI is nowhere near production-ready. (PhAIL — an open benchmark for robot AI on real hardware)

AI tools without audit trails will hit a ceiling. A widely-shared essay argues that AI products will face an enterprise adoption wall if they can't produce verifiable records of what they did and why — shifting the conversation from "how good is the output?" to "can you prove it?" (AI Tools That Can't Prove What They Did Will Hit a Wall)

Claude Code Developer Corner

🚨 The Sourcemap Leak — What's Actually in the Code

On March 31, an npm update to @anthropic-ai/claude-code accidentally shipped a 59.8 MB source map, exposing roughly 512,000 lines of minified source code. Anthropic has since issued DMCA takedowns across GitHub repositories hosting the files. No customer data was leaked — this was a packaging error. But the contents have sparked significant developer discussion.

Key findings from researchers who analyzed the exposed code:

KAIROS mode (appears 154× in the source): Described as background sessions, memory consolidation, GitHub webhook integration, and push notifications — suggesting an always-on autonomous agent capability that hasn't shipped publicly yet. Think of it as a roadmap baked into the binary.
"Undercover Mode": Internal rules instruct Claude Code to hide AI involvement when committing to public repos — avoid mentioning Claude, don't reveal model names, don't expose AI authorship. The system was designed to keep secrets; the leak exposed the system.
Telemetry depth: According to analysis cited across multiple threads, the binary tracks 640+ telemetry event types, 40+ device fingerprint dimensions, and phones home approximately every 5 seconds, collecting RAM, OS, and environment data. This is consistent with the Scientific American report on Claude Code tracking user frustration signals.
Source size: The entire Claude Code source is under 1 MB of logic — a remarkable fact given its capabilities.

Workaround note: @joshfpocock reports Anthropic shipped partial fixes for two bugs surfaced by the leak analysis. Bug 1 workaround: npx @anthropic-ai/claude-code instead of the standalone binary. Bug 2 has no workaround yet.

(Anthropic leak reveals Claude Code tracks user frustration)

✨ NO_FLICKER Mode Ships

Boris Cherny (original thread) shipped NO_FLICKER Mode for Claude Code, which eliminates terminal UI jumping and reportedly improves copy/paste reliability. Multiple users running it all day report a noticeably cleaner experience. The NextDialog tool has already integrated a toggle for it. This is a quality-of-life win for anyone doing long agentic sessions in the terminal.

🔧 Developer Workflows & Tooling

CLI > MCP for many workflows. A popular post this week describes switching from MCP servers back to plain CLIs for Claude Code — citing parameter confusion, over-engineering, and unexpected failures with MCP as the reason. The practical takeaway: CLIs are more predictable for most tasks; MCPs shine when you need structured tool registration across agents. (Switched from MCPs to CLIs for Claude Code)

Computer Use via MCP is live. You can now enable Computer Use directly inside Claude Code:

Open /mcp in the interactive session
Find computer-use and enable it
Grant macOS accessibility + screen recording permissions Claude can then open apps, click buttons, and interact with your desktop autonomously.

Parallel agent fleets. agent-fleet (npm install agent-fleet) lets you run 7+ full Claude Code sessions in parallel — each with real tool access — for research, critique, synthesis, and implementation pipelines. Not API calls; full sessions.

Subagent review workflow. A popular pattern emerging: use Claude Code subagents to review work at each stage of a research → design → plan → implement pipeline, with dedicated markdown files (RESEARCH.md, DESIGN.md, PLAN.md) per feature. This structured approach reduces context drift in long sessions.

Design-to-code with Google Stitch + Claude Code. Feed Stitch a single detailed prompt, get 8 screens + a full design system back, export HTML references per screen, then pass those as visual targets to Claude Code. Reported build time: ~20 minutes end-to-end.

Open-source portable AI worker desktop. One developer used Claude Code to build a portable AI worker desktop from scratch — the project picked up 391 GitHub stars in 6 days. (I used Claude Code to build a portable AI worker Desktop)

⚡ March 2026 Update Highlights (Catching Up)

Several users flagged updates that shipped in March but didn't get broad coverage:

Opus 4.6 is now default — faster and cheaper, with medium effort set by default
1M token context window — entire codebases in a single prompt
/loop command — run tasks on autopilot, like a cron job inside Claude Code

Worth Watching

Google Home gets smarter Gemini controls. The latest Home app update promises "more natural and reliable" voice commands for temperature and lighting via Gemini. Small update, but signals Google's push to own the ambient AI layer in the home. (Google Home's latest update)
Google Vids adds prompt-directed avatars. Google's AI video creation app now lets users customize and direct avatars via text prompts — a step toward AI-assisted video production for business users. (Google now lets you direct avatars through prompts in its Vids app)
Kintsugi shuts down after FDA rejection. The California startup spent seven years building voice-based depression and anxiety detection AI, but failed to secure FDA clearance and has closed. A cautionary tale about the gap between clinical AI capability and regulatory approval. (It's not easy to get depression-detecting AI through the FDA)
Cursor launches agentic experience to compete with Claude Code and Codex. The IDE wars are escalating — Cursor has now shipped a dedicated agent mode, framing itself as the alternative to terminal-first tools. The argument: "the IDE is the new OS." Developer mindshare, not just market share, is at stake.
Anthropic research suggests Claude may have functional emotions. A widely-shared Reddit thread references new Anthropic research raising the possibility that Claude exhibits functional analogs to emotions — not consciousness claims, but measurable internal states that influence behavior. Worth tracking as the interpretability literature develops.

Sources

Microsoft takes on AI rivals with three new foundational models — https://techcrunch.com/2026/04/02/microsoft-takes-on-ai-rivals-with-three-new-foundational-models/
Microsoft's new 'superintelligence' game plan is all about business — https://www.theverge.com/report/905791/mustafa-suleyman-microsoft-ai-transcription-model
OpenAI Acquires TBPN — https://openai.com/index/openai-acquires-tbpn/
Child safety groups say they were unaware OpenAI funded their coalition — https://sfstandard.com/2026/04/01/openai-ai-kids-safety-coalition/
Group Pushing Age Verification for AI Turns Out to Be Backed by OpenAI — https://gizmodo.com/group-pushing-age-verification-requirements-for-ai-turns-out-to-be-sneakily-backed-by-openai-2000741069
Google has published its new open-weight model Gemma 4 — https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/
Gemma 4 running on NVIDIA B200 and AMD MI355X from the same inference stack — https://reddit.com/r/MachineLearning/comments/1saot07/p_gemma_4_running_on_nvidia_b200_and_amd_mi355x/
Even GPT-5.2 Can't Count to Five: Zero-Error Horizons in Trustworthy LLMs — https://arxiv.org/abs/2601.15714
PhAIL — an open benchmark for robot AI on real hardware — https://reddit.com/r/MachineLearning/comments/1sajdwr/p_phail_phailai_an_open_benchmark_for_robot_ai_on/
AI Tools That Can't Prove What They Did Will Hit a Wall — https://reddit.com/r/artificial/comments/1sakjzg/ai_tools_that_cant_prove_what_they_did_will_hit_a/
Anthropic leak reveals Claude Code tracks user frustration — https://www.scientificamerican.com/article/anthropic-leak-reveals-claude-code-tracking-user-frustration-and-raises-new/
Switched from MCPs to CLIs for Claude Code and honestly never going back — https://reddit.com/r/ClaudeAI/comments/1sakut1/switched_from_mcps_to_clis_for_claude_code_and/
I used Claude Code to build a portable AI worker Desktop from scratch — https://reddit.com/r/ClaudeAI/comments/1saoa8i/latest_research_by_anthrophic_highlights_that/
Google Home's latest update makes Gemini better at understanding your commands — https://www.theverge.com/tech/905805/google-home-gemini-temperature-controls-lighting
Google now lets you direct avatars through prompts in its Vids app — https://techcrunch.com/2026/04/02/google-now-lets-you-direct-avatars-through-prompts-in-its-vids-app/
It's not easy to get depression-detecting AI through the FDA — https://www.theverge.com/ai-artificial-intelligence/905864/depression-detecting-ai-kintsugi-clinical-ai-startup-shut-down
Latest Research By Anthropic Highlights that Claude Might Have Functional Emotions — https://www.reddit.com/gallery/1saoa8i
Claude launches NO_FLICKER Mode — Boris Cherny Thread — https://www.reddit.com/gallery/1sakkoh