Agentic Brew Daily
Your daily shot of what's brewing in AI
Fresh Batch
- Robinhood handing Claude and ChatGPT keys to a brokerage account landed the same week five labs rushed out agent-security patches, proving the money layer arrived before the trust layer did.
- SK Hynix posting a 72% operating margin while Nvidia sits at 65% just flipped the AI-stack pricing power story — memory makers, not GPU vendors, hold the choke point.
- Pope Leo XIV staging Anthropic's Chris Olah on the Magnifica Humanitas dais while DuckDuckGo iOS installs spike 70% shows institutions and consumers are pushing back on frontier labs from opposite ends.
Bold Shots
Today's biggest AI stories, no chaser
On May 25, Pope Leo XIV released Magnifica Humanitas, the first papal encyclical dedicated entirely to AI, deliberately signed on the 135th anniversary of Rerum Novarum. The 42,300-word document declares AI must be 'disarmed,' not merely regulated — demanding decision traceability, meaningful human control over lethal autonomous weapons, and worker protections against algorithmic surveillance. He presented it at the Vatican Synod Hall alongside Anthropic co-founder Chris Olah, while OpenAI, Google DeepMind, and Microsoft stayed quiet. Detector tool Pangram flagged 40-100% of certain paragraphs as AI-generated, sparking debate about whether the text was drafted with Claude.
Why it matters: This is the Catholic Church mobilizing the moral authority of 1.4 billion believers to reframe AI as an ethics problem on the scale of the Industrial Revolution. The choice to share the Vatican stage exclusively with Anthropic — and the irony that detectors flagged the text itself as AI-written — has reshaped the public legitimacy game for frontier labs.
President Trump understands that unnecessary regulation is the biggest threat to innovation in America. Winning the AI race means not only beating China but also clearing bureaucratic hurdles thrown up by state legislatures and woke politicians in DC
It is doubtful that the Pope's encyclical will have much real impact on how companies develop AI or how countries regulate it. But it's still getting a lot of attention—and criticism.
Micron became the first U.S. pure-play memory firm to cross $1T on May 26 after a 19% single-day jump triggered by a UBS upgrade that tripled its price target from $535 to $1,625. SK Hynix followed one day later as shares jumped 12% to a record high. Both companies have their entire 2026 HBM production sold out; hyperscalers are signing 5-year deals with 10-30% cash prepayments. SK Hynix posted a Q1 2026 operating margin of 72% on $35.5B in revenue — higher than Nvidia's 65%.
Why it matters: Wall Street has finally repriced memory makers as AI infrastructure rather than commodity DRAM, minting two trillion-dollar companies in 48 hours. SK Hynix's 72% operating margin — beating Nvidia's — flips the conventional wisdom about where pricing power sits in the AI stack and reframes HBM as a structural choke point.
On May 25, HiSilicon president He Tingbo used his ISCAS 2026 keynote to propose the Tau (τ) Scaling Law as a successor to Moore's Law, shifting optimization from transistor geometry to RC signal delay. LogicFolding stacks digital, analog, and memory layers at design time, pushing density from 155 to 238 MTr/mm² (+53.5%) on the same node. Huawei claims 381 chips already designed under the τ framework and projects 1.4 nm-equivalent density by 2031. Fall 2026 Kirin chips will be first commercial silicon. Markets responded: SMIC +18%, Hua Hong hit its 20% daily limit.
Why it matters: Tau scaling reframes chip progress around signal delay rather than transistor size — a clever workaround to U.S. EUV sanctions whose physics still trails TSMC by years. If even partially credible, it changes the political ROI of export controls; if it isn't, it's still a strong domestic-substitution rallying flag.
Chinas Huawei can't access EUV. So they wrote their own scaling law. The leverage of US export controls erodes. Huawei just presented the Tau (τ) Scaling Law at IEEE ISCAS...
Huawei claims its new LogicFolding approach can help narrow the gap with $TSM by improving chip density, latency & power efficiency without relying only on smaller transistors.
Robinhood launched Agentic Trading and an Agentic Credit Card on May 27, opening MCP servers that let approved AI agents place equity trades and run card purchases from a sandboxed, user-funded account. Approved at launch: Claude Code, Claude Desktop, ChatGPT, Codex, Codex CLI, Cursor — plus any MCP-compatible client. Beta is equities-only; options, crypto, futures, and prediction markets are on the roadmap. The Gold virtual card earns 3% cash back on agentic purchases, with guardrails including sandboxed accounts, push notifications, real-time P&L, and optional manual approval.
Why it matters: Robinhood is the first regulated U.S. broker to license Claude and ChatGPT to trade stocks and swipe a credit card on your behalf. This is the first concrete answer to 'what does it mean to let an agent touch real money?' for a mass retail audience — and a template every other broker will now have to react to.
DuckDuckGo reported a sustained surge in U.S. installs and noai.duckduckgo.com traffic the week after Google I/O 2026 made AI Overviews and AI Mode the default with no consent screen. U.S. app installs were up 18.1% WoW on average May 20-25, six consecutive days of growth, peaking at 30.5% on May 25. iOS installs in the U.S. jumped 33% WoW on average, peaking at 69.9% on May 25. Visits to the AI-free noai.duckduckgo.com page rose 22.7% WoW. Third-party Apptopia confirmed +29% U.S. daily downloads.
Why it matters: This is the first measurable consumer revolt against mandatory AI Search. Users actively sought out an AI-free experience — a clean rejection of Google's bet that AI Mode wins on default. It sets up 'AI-free' as a deliberate product positioning category and pressures every search vendor to expose a kill switch.
People aren't just complaining about Google's AI search overhaul, they're leaving. Yesterday alone, our week over week installs surged 30% in the U.S. Momentum is growing. It's time to Fire Google.
DuckDuckGo installs are up 30% as users reject being 'force-fed' Google's AI Search
Slow Drip
Blog reads worth savoring
Exclusive hiring data shows AI engineering listings up 50-100% YoY at Apple/Google/TikTok while Stripe and Atlassian quietly out-hire Big Tech, with concrete percentages by company and region.
First production-ready 2-bit KV cache quantization that stays within 1.42 points of BF16 on RULER-NIAH at 128K, where KIVI, QuaRot, and TurboQuant+ all collapse.
Async RL trick that shrinks 1TB checkpoints to ~20GB sparse safetensors deltas by exploiting bf16's 99% non-update rate, enabling cross-cloud RL with no direct network link between trainer and vLLM.
Concrete lethal-trifecta breakdown showing how Copilot's unapproved self-emails plus rendered external images plus pre-auth OneDrive links chain into a prompt-injection file exfiltration.
The Grind
Research papers, decoded
Apple researchers ran frontier reasoning models (o3-mini, DeepSeek-R1, Claude 3.7 Sonnet Thinking) through controllable puzzle environments instead of contaminated math/code benchmarks. Three regimes: standard LLMs beat reasoning models on low-complexity tasks, LRMs win in the middle, and both collapse completely above a complexity threshold. Counterintuitively, LRMs reduce reasoning effort as they approach failure, even with token budget left. Even handing the model the exact algorithm doesn't fix the collapse. Test agents against complexity-graded puzzles, not just AIME/MATH — there's a hard ceiling where 'thinking longer' gives up.
SkillOpt treats an agent's natural-language skill document as trainable external state for a frozen LLM, applying optimizer discipline — learning-rate-bounded edits, rejected-edit buffer, validation gating, epoch-wise meta updates — to evolve the skill from scored rollouts. Across 52 (model × benchmark × harness) cells covering GPT-5.5, Codex, and Claude Code, it beats hand-written, one-shot LLM, Trace2Skill, TextGrad, GEPA, and EvoSkill baselines: +23.5 pts in direct chat, +24.8 in Codex, +19.1 in Claude Code on GPT-5.5. Optimized skills transfer across model scales and execution environments. The most concrete recipe yet for improving skill files without fine-tuning.
PiD replaces the standard VAE decoder with a conditional pixel-diffusion module that fuses decoding and upsampling into one generative pass. DMD2 distillation gets it down to 4 inference steps. Decodes 512×512 latents to 2048×2048 in ~210ms on a GB200 or under 1s on a consumer RTX 5090 with 13GB peak memory — roughly 6× faster than cascaded super-resolution pipelines, with better fidelity. A near drop-in path to 4×-8× upscaled output without a separate SR stage.
Adds an offline 'sleep' phase to hybrid attention/SSM models: N recurrent passes consolidate recent context into the SSM block's fast weights via a learned local rule, then the KV cache is cleared. Inference latency stays the same because the extra compute is paid offline. Synthetic Rule-110 cellular automaton accuracy jumps from ~10% to >30% with 3-4 sleep loops at depth 32; Jet-Nemotron-2B gains 11% on 8-op GSM-Infinite problems. A credible path to long-horizon reasoning without paying quadratic attention costs at wake time.
HumanEgo learns manipulation policies from ~30 minutes of Aria-glasses human egocentric video per task, with no robot data. Pipeline: inpaint out human arms, render virtual grippers to close the appearance gap, encode hand-object relationships as Interaction-Centric Tokens (ICT), then train a flow-matching policy. Hits 92.5% average success on four real-world tasks; 15 minutes of human video beats 30 minutes of robot teleoperation (75% vs 51%); transfers zero-shot to Trossen, Franka, and UR10 without retraining. Robot teleop is the most expensive part of the data pipeline — a person wearing glasses for 30 minutes can replace it.
The Mill
Builder tools ground for action
The Counter
Voices from the AI bar today
2.3K viewsRoman Yampolskiy lays out an evidence-based case that superintelligence containment is mathematically impossible and the near-term risk window is closing fast.
20K viewsA working multi-agent system that scrapes SEC Form 4 and 13F filings, Fed speeches, and on-chain data to generate trade signals via an automated n8n/LangGraph workflow.
4.5K viewsThe DoD vs. Anthropic standoff over autonomous-weapons and domestic-surveillance use of Claude, and what the policy fallout means for every frontier-lab usage policy.
8.5K engagementPolymarket flags FBI labeling AI and data-center protesters 'anti-tech extremists' as moratorium odds reach 91%.
3.8K engagementThe AI ROI reckoning conversation lands on mainstream feeds as companies pull back when costs outpace human labor.
2.6K upvotes · 135 commentsFree, certificate-bearing courses on agentic AI, Claude Code, and MCP.
2.2K upvotes · 359 commentsMeta is moving legally against the tool that strips safety guardrails from Llama, igniting an open-weights-control debate.
Roast Calendar
Your AI week, day by day
Last Sip
Parting thoughts
The through-line today is timing. Robinhood opened a brokerage account to Claude the same week Simon Willison documented a working Copilot exfiltration chain, and the same week SK Hynix booked a 72% operating margin selling the memory those agents need to think. The interesting question isn't whether agents touch real money — they already do. It's who picks up the bill when they trade wrong, who owns the rails when memory is sold out for the year, and who signs the moral document the next time a Pope writes one. Worth chewing on.