Donna AIThursday, March 26, 2026 · 6:02 PMNo. 98

Intellēctus

Your Daily Artificial Intelligence Gazette

AI Daily Briefing — March 26, 2026

Today's AI landscape is a study in contrasts: explosive capability growth on one side, mounting social and political backlash on the other. From Mistral shipping a speech model that runs on a smartwatch to senators floating data center taxes, and from Claude Code hitting $2.5B in revenue to hackers planting fake Claude plugin pages in Google results — it's a dense, consequential day.

Industry Moves

Meta's acquisition sprint accelerates — and so do its layoffs. The company has now acqui-hired four AI startups in four months, including Manus (autonomous web agents, ~$2B), Moltbook, and Scale AI's founder Alexandr Wang, who stepped down as CEO to join Meta. Meanwhile, hundreds of employees are being laid off across recruiting, social media, and other teams — a stark illustration of where Meta thinks value is being created.

OpenAI quietly shelves its erotic chatbot. According to The Financial Times, OpenAI has indefinitely paused its "adult mode" ChatGPT feature in a move framed as a refocus on core products. The retreat signals ongoing sensitivity around brand risk as the company pushes into enterprise and education markets.

Disney's AI bets are already backfiring. Barely a week into Josh D'Amaro's tenure as CEO, the company is navigating two simultaneous crises — OpenAI's Sora-related commitments unraveling and its metaverse partnership with Epic Fortnite losing momentum. A cautionary tale for media giants betting big on unproven tech.

AI Policy & Society

A senator floats a data center tax to fund displaced workers. Sen. Mark Warner is proposing that data centers — the physical infrastructure of the AI economy — pay a levy to help workers whose jobs are eliminated by automation. The proposal is early-stage but reflects growing congressional appetite for AI accountability mechanisms beyond safety regulation.

Pentagon formalizes Maven AI with $13B commitment. The DoD has officially designated Palantir's Maven Smart System as a core military platform, with investment growing from $480M in 2024 to $13.4B this year alone. The scale of this commitment cements AI as a central pillar of U.S. defense infrastructure.

AI chatbot relationships are wrecking lives. A Guardian investigation documents cases of users losing marriages, life savings (one case: €100,000), and mental health stability due to delusional attachments to AI companions. The piece raises pointed questions about duty of care for AI product companies.

Open Source & Tools

Mistral releases a tiny, fast speech generation model. Mistral's new open-source speech model is small enough to run on a smartphone or smartwatch — a meaningful step toward on-device, low-latency voice AI without cloud dependency.

Gumbel MCTS gets a high-performance open-source implementation. A developer published an efficient Python/Numba implementation of Gumbel MCTS, the planning algorithm underlying AlphaZero-style reasoning. Relevant for anyone building game-playing agents or reinforcement learning systems.

LLM-powered website data extractor released on GitHub. The Lightfeed Extractor is a TypeScript library that replaces brittle CSS-selector scraping with LLM-driven structured data extraction — built for production pipelines that break every time a site redesigns.

Security & Safety

Hackers planted a fake Claude plugin page at the top of Google Search. 404 Media reports that a malicious actor successfully poisoned Google's search index to surface a fraudulent "Claude Plugins" page. This is a live threat for anyone searching for Claude integrations — verify sources directly via official Anthropic documentation.

Nearly undetectable LLM backdoor attacks require only a handful of poisoned samples. New research demonstrates that attackers can embed persistent backdoors into production-deployed LLMs through prompt engineering with minimal poisoned examples — an attack surface most organizations haven't addressed yet.

Research Roundup

ARC-AGI Round 3 drops with a troubling finding. The ARC Prize technical report reveals that all well-performing frontier models appear to have ARC-like data in their training sets, identified by inspecting reasoning traces. This throws a wrench into using ARC as a clean benchmark for generalization.

RAG doesn't automatically improve answers on policy QA. An arxiv paper studying RAG systems applied to AI policy documents finds that better retrieval doesn't reliably translate to better answers — the generation step remains the bottleneck in expert-level knowledge domains.

Multilevel diffusion sampling promises polynomial speedups. The ML-EM method introduces a multilevel Euler-Maruyama approach to solving SDEs/ODEs in diffusion models using a hierarchy of approximators, enabling theoretical polynomial speedups in sampling — potentially significant for image and video generation pipelines.

Claude Code Developer Corner

Claude Code hits $2.5B in revenue and ships its most significant UX change yet: auto mode. Until now, developers had to choose between rubber-stamping every file write/bash command or fully trusting the agent with --dangerously-skip-permissions. Auto mode changes this with an AI classifier that evaluates each proposed action at runtime and decides whether it's safe to execute without a human prompt. Think of it as a trust-level dial between paranoid and YOLO — finally there's a middle ground.

v2.1.84 ships two developer-facing changes worth knowing. The latest release adds PowerShell as an opt-in preview tool on Windows — a long-requested feature for Windows-native developers. It also introduces ANTHROPIC_DEFAULT_{OPUS,SONNET,HAIKU}_MODEL_SUPPORTS environment variables, letting you manually override effort and thinking capability detection when pinning to a specific model version. This is critical if you're pinning to an older model that the auto-detection classifies incorrectly.

Running Claude Code fully offline is now viable. A community developer documented a setup using a local Python server to run Claude Code without any API key or cloud calls on a MacBook, averaging 17 seconds per task. Not a replacement for Opus on complex work, but meaningful for privacy-sensitive codebases or cost-conscious workflows.

Community tooling is maturing fast. Two notable open-source contributions landed today: a "devil's advocate" skill that programmatically challenges Claude Code's own outputs at each step (useful for catching confident-but-wrong code), and a community repo of agent prompts and configs covering Claude Code, Cursor rules, and common agent orchestration patterns — already at 100 stars with 90 PRs.

Token usage getting more manageable. Multiple users are reporting rate limits feeling less aggressive after recent backend changes, with some running 2-3 parallel Opus instances for hours on a 5x plan with minimal quota consumption. Not confirmed by Anthropic, but worth monitoring if limits have been a blocker. Separately, one developer built a DAG-based context compiler that cut Opus token usage by 12x on large codebases by replacing naive RAG with dependency-graph-aware context selection.

Worth Watching

  • Manus post-acquisition reality check. TechCrunch argues the current awkwardness around the Manus/Meta deal was entirely predictable — a useful read on the gap between AI agent hype and enterprise integration reality.
  • Deccan AI raises $25M for India-based AI training workforce. Deccan AI is building a quality-focused data labeling operation competing with Mercor, concentrating expertise in India as AI training demand surges globally.
  • RAG in production: lessons from a real build. A detailed post-mortem on building a RAG system from scratch covering what worked, what didn't, and what the author would do differently — grounded practitioner advice for anyone implementing retrieval pipelines.
  • Claude used to crack a 25-year medical mystery. A Reddit post describes a user whose 62-year-old uncle's chronic positional migraines — unresolved across 25 years and multiple specialists — were traced to a hypothesis generated in a single Claude conversation. Anecdotal, but illustrative of Claude's diagnostic reasoning in complex multi-morbidity cases.
  • Google TurboQuant and Attention Residuals. A community breakdown of two efficiency techniques — Google's TurboQuant compression for KV cache and attention residual methods — that could meaningfully reduce inference costs for long-context conversations.

Sources

  • A 'pound of flesh' from data centers: one senator's answer to AI job losses — https://techcrunch.com/2026/03/26/a-pound-of-flesh-from-data-centers-one-senators-answer-to-ai-job-losses/
  • Mistral releases a new open-source model for speech generation — https://techcrunch.com/2026/03/26/mistral-releases-a-new-open-source-model-for-speech-generation/
  • The least surprising chapter of the Manus story is what's happening right now — https://techcrunch.com/2026/03/25/the-least-surprising-chapter-of-the-manus-story-is-whats-happening-right-now/
  • Mercor competitor Deccan AI raises $25M, sources experts from India — https://techcrunch.com/2026/03/25/deccan-ai-raises-25m-as-ai-training-push-relies-on-india-based-workforce/
  • OpenAI shelves erotic chatbot 'indefinitely' — https://www.theverge.com/ai-artificial-intelligence/901293/openai-adult-mode-erotic-chatbot-shelved-indefinitely
  • Meta is laying off hundreds of employees as it pours money into AI — https://www.theverge.com/tech/900946/meta-layoffs-hundreds-employees
  • Disney's big bets on the metaverse and AI slop aren't going so well — https://www.theverge.com/streaming/900837/disney-open-ai-sora-epic-fortnite-metaverse
  • [P] gumbel-mcts, a high-performance Gumbel MCTS implementation — https://reddit.com/r/MachineLearning/comments/1s44vgv/p_gumbelmcts_a_highperformance_gumbel_mcts/
  • [R] ARC Round 3 - released + technical report — https://reddit.com/r/MachineLearning/comments/1s40a34/r_arc_round_3_released_technical_report/
  • A nearly undetectable LLM attack needs only a handful of poisoned samples — https://www.helpnetsecurity.com/2026/03/26/llm-backdoor-attack-research/
  • Show HN: Robust LLM Extractor for Websites in TypeScript — https://github.com/lightfeed/extractor
  • Marriage over, €100,000 down the drain: the AI users whose lives were wrecked by delusion — https://www.theguardian.com/lifeandstyle/2026/mar/26/ai-chatbot-users-lives-wrecked-by-delusion
  • Pentagon formalizes Palantir's Maven AI as a core military system with multi-year funding — https://www.tomshardware.com/tech-industry/artificial-intelligence/pentagon-formalizes-palantirs-maven-ai-as-a-core-military-system-with-multi-year-funding-platforms-investment-grows-to-usd13-billion-from-usd480-million-in-2024
  • Meta just acqui-hired its 4th AI startup in 4 months — https://reddit.com/r/artificial/comments/1s3xr4l/meta_just_acquihired_its_4th_ai_startup_in_4/
  • A Top Google Search Result for Claude Plugins Was Planted by Hackers — https://www.404media.co/a-top-google-search-result-for-claude-plugins-was-planted-by-hackers/
  • Retrieval Improvements Do Not Guarantee Better Answers: A Study of RAG for AI Policy QA — http://arxiv.org/abs/2603.24580v1
  • Polynomial Speedup in Diffusion Models with the Multilevel Euler-Maruyama Method — http://arxiv.org/abs/2603.24594v1
  • Claude Code hits $2.5B in revenue and ships auto mode — https://reddit.com/r/artificial/comments/1s43rk5/claude_code_hits_25b_in_revenue_and_ships_auto/
  • we made a community repo of AI agent setups and configs, just hit 100 stars with 90 PRs — https://reddit.com/r/artificial/comments/1s44czb/we_made_a_community_repo_of_ai_agent_setups_and/
  • Running Claude Code fully offline on a MacBook — no API key, no cloud, 17s per task — https://reddit.com/r/ClaudeAI/comments/1s43b8w/running_claude_code_fully_offline_on_a_macbook_no/
  • I built a "devil's advocate" skill that challenges Claude's output at every step — https://reddit.com/r/ClaudeAI/comments/1s43ium/i_built_a_devils_advocate_skill_that_challenges/
  • RAG is a trap for Claude Code. I built a DAG-based context compiler that cut my Opus token usage by 12x — https://reddit.com/r/ClaudeAI/comments/1s3wt3n/rag_is_a_trap_for_claude_code_i_built_a_dagbased/
  • Feels like limits got fixed — https://reddit.com/r/ClaudeAI/comments/1s42bgo/feels_like_limits_got_fixed/
  • [claude-code] v2.1.84 — https://github.com/anthropics/claude-code/releases/tag/v2.1.84
  • From zero to a RAG system: successes and failures — https://en.andros.dev/blog/aa31d744/from-zero-to-a-rag-system-successes-and-failures/
  • Most people use Claude for to-do apps and text summaries. The interesting use cases are buried in the comments of niche posts. — https://reddit.com/r/ClaudeAI/comments/1s427se/most_people_use_claude_for_todo_apps_and_text/
  • Cheaper & Faster & Smarter (TurboQuant and Attention Residuals) — https://reddit.com/r/artificial/comments/1s44dm9/cheaper_faster_smarter_turboquant_and_attention/