The prospectus spends more space on AI safety risks, including self-preservation behaviors, than on the business itself, while revealing extreme capital intensity and customer concentration. Institutions modeling AI sector exposure now have a rare audited look at the economics underneath the safety narrative.
Anthropic disclosed a $42 billion net loss in 2025, plans $518 billion in infrastructure spend, and roughly a quarter of revenue comes from just two clients.
Read the signal →Robinhood Agents can analyze markets and autonomously trade stocks, options and crypto, with per-trade approval optional by default, putting it ahead of more cautious European neobrokers. Regulators including the Bank of England and ESMA have already flagged volatility and accountability risks from AI-driven trading.
Corroborated by 2 outlets
Read the source →The ECB is funding a digital euro pilot with Deutsche Bank, Revolut, Adyen and UniCredit to build a sovereign payment rail by 2029 and cut reliance on US card networks. Banks face significant implementation costs, and existing networks face real competitive pressure.
The pilot carries €1.3 billion in funding, targeting a sovereign infrastructure by 2029.
Read the source →Even hit consumer products are not converting, pushing AI companies toward enterprise revenue as the viable strategy, as OpenAI's pivot showed. The gap between adoption and profitability is now the central constraint on consumer AI business models.
About 2.2% of consumers pay for AI, at an average of $31 per month.
The industry · financial services, no AI angle
Things happen · 23
Labs & models 6
Research · 54 analyzed, top 8
Open-weight models that nominally support long context lose a large share of per-record accuracy when forced to generate long, stateful outputs, and almost never complete a 128K-token transduction document without repair.
BEARS ON 1 · RIGOR HIGH
Terminal agents almost always attempt to verify their own solutions, but that verification only detects about 61% of errors and repair succeeds only about 49% of the time, and distilling teacher verification conditioned on student candidates (SCVD) improves terminal task success.
BEARS ON 1 · RIGOR HIGH
Reading the full answer-probability vector from each sampled reasoning trajectory (grey-box) lets adaptive self-consistency identify the modal answer with 46.4 to 95.6 percent fewer trajectories than answer-only adaptive baselines, and theory shows the sample-complexity dependence on the answer gap improves from quadratic to at best linear in the most consistent case.
BEARS ON 2 · RIGOR HIGH
What builders are reading · 6 reads
Tooling & models 1
Google ships Gemini 4 Argon, a new frontier model, relevant to vendor evals and model-risk review pipelines
Agents 1
Magnitude self-optimizing inference engine for agents, adoptable to tune agent behavior without prompt rewrites
Evals & benchmarks 1
Cloudflare open-sources decision models, small models for routed decisions, relevant to cost-aware model choice
Security & governance 2
Figma whitelists MCP clients, a concrete pattern for gating MCP tool access in regulated environments
The deal pushes Peoples past the $10 billion asset threshold for the first time and signals continued consolidation among mid-sized banks. It is expected to close in the first half of 2027 and be immediately accretive to earnings.
The combined bank will hold about $14 billion in assets across 150 branches in eight states plus Washington, D.C.
Corroborated by 3 outlets
Read the source →Cheap, LLM-powered exploitation of known vulnerabilities changes the threat math for banks and payment networks. Visa, Mastercard, Google and NPCI are now writing protocols for AI agent verification and governance.
The technology, public, and financial sectors account for 44% of cyberattacks in India.
Read the source →The column
Anthropic has filed to go public, and the filing itself is the news: 80 pages of risk warnings paired with a $42 billion loss, with more space devoted to AI safety, including self-preservation behaviors, than to the business of selling AI. For anyone modeling the sector, this is the first audited look under the hood, and it confirms the uncomfortable picture many of us have been sketching: one vendor's revenue depends heavily on a few customers who are themselves unprofitable on AI, and if those customers retrench, the losses compound in a way no safety narrative offsets. At the same time, the sheer capital intensity on display cuts both ways. You can read it as evidence that hyperscaler AI capex stays funded from operating cash flow, not debt, or you can read a $42 billion figure and conclude the frontier is a machine for converting investor money into inference. What would settle it: the customer-level disclosure in the actual prospectus, names and renewal terms. If the big buyers have multi-year committed spend with measurable ROI, the bear case shrinks; if the concentration is one or two pilot-heavy accounts, it grows.
Mastercard put a number on the security side of the ledger: AI has dropped the cost of a working cyberattack to under $3. When exploitation of known vulnerabilities costs less than a sandwich, every bank's threat model changes, and Visa, Mastercard, Google and NPCI are now drafting protocols for verifying AI agents. That work arrives just as Robinhood rolls out agents that trade stocks, options and crypto autonomously around the clock, with per-trade approval optional by default. The Bank of England and ESMA have already flagged accountability risks here, and they are right to: you cannot assign liability to an agent that does not have a balance sheet. The interesting convergence is that the payments industry is building agent identity standards at the same moment consumer agents are being pointed at markets. Verification protocols written for payments will get repurposed for trading faster than any regulator can hold a consultation.
Meanwhile the ECB is funding a digital euro pilot with Deutsche Bank, Revolut, Adyen and UniCredit, targeting a sovereign rail by 2029. Banks face real implementation costs, and Visa and Mastercard face real competitive pressure, though 2029 is an eternity in payments. And TechCrunch reports that only 2.2% of consumer AI users pay, which explains why everyone from OpenAI onward is chasing enterprise contracts. The pattern across all four stories is the same: AI adoption is abundant, and AI economics are not. Peoples Bancorp paying $728 million all-stock for Capital Bancorp, crossing the $10 billion asset line, is the old economy quietly consolidating while the new one figures out how to charge anybody anything.
The cheapest cyberattack in history now costs less than the per-trade fee Robinhood will waive.
Banks & payments 4
Policy & regulation 3
Work & labor 1
LLMs, including frontier reasoning models, often answer questions with outdated values of updated variables even when the current value remains readable inside the model, because attention at the answer position favors old values over the current one; redirecting attention at test time fixes most of these errors without retraining. (One sentence, one testable assertion: a specific failure mode, stale binding, is caused by an identifiable selection mechanism and is repairable.)
BEARS ON 1 · RIGOR HIGH
Low end-to-end accuracy of multimodal LLMs on compositional tasks often reflects cascading errors from upstream prerequisites rather than intrinsic capability deficits, and controlled interventions can separate the two: supplying correct prerequisites eliminates 54% of errors on cognitive tasks.
BEARS ON 1 · RIGOR HIGH
How a token budget's stopping rule is defined (strict cap vs letting the boundary attempt finish) changes measured test-time-scaling accuracy by up to ~43 percentage points at short budgets, and the apparent gains often come from converting abstention into an available answer rather than better selection.
RIGOR HIGH
A cheap deterministic two-stage IR retriever (BGE-base bi-encoder plus a small cross-encoder over MCP) matches the pass rate of an LLM-mediated agentic skill-retrieval loop on SkillsBench, while cutting per-trial retrieval spend from $51.30 to $27.54, within fifty cents of a no-skill baseline.
BEARS ON 1 · RIGOR MEDIUM
A risk-aware contextual Thompson Sampling policy recovers 86% of the impact-weighted agent failures an oracle could find with only 50 trials (6% of the corpus), versus 25% for uniform allocation, by concentrating trials on high-impact, failure-prone scenarios.
BEARS ON 1 · RIGOR MEDIUM
OpenID Foundation spec on identity management for agentic AI, directly usable for agent auth design
Infra & cost 1
turbopuffer argues vector DBs are collapsing into object storage, may change RAG stack cost choices