Scam losses hit roughly $20 billion in 2025 while banks detect only 10% to 15% of losses and recover 2% to 5%, so prevention is the only viable defense. The report flags misaligned liability across banks, social platforms, and telecoms as the reason prevention stays underfunded.
Victim losses were estimated at roughly $20 billion in 2025, up 26% year over year.
Read the source →European financial regulators warn Q-Day could arrive before quantum computers are commercially useful and want post-quantum cryptography adopted by end of 2026. Harvest-now-decrypt-later attacks make encrypted data stolen today a future liability.
Regulators urge adoption of post-quantum cryptography by the end of 2026.
Read the source →A settlement network for tokenized deposits across 25 major U.S. banks, connected to existing rails like RTP and CHIPS, is a concrete step toward regulated on-chain money. It reaches participating banks in the first half of 2027, setting a shared standard banks will have to plan around.
The Clearing House processes over $2 trillion daily.
Read the source →An eighteenfold revenue surge to $375M annualized on a shift to hybrid synthetic-plus-expert datasets confirms frontier model development still runs on proprietary, verified data, not public web scraping. Data provenance and quality assurance are emerging as a distinct spend layer in AI infrastructure.
Things happen · 23
Policy & regulation 7
Research · 51 analyzed, top 8
State-of-the-art LLM assistants complete less than 60% of grounded, multi-turn Indian retail banking tasks reliably (success on all three repeated trials), and at-least-once success overstates dependable performance by 11 to 21 percentage points.
BEARS ON 1 · RIGOR HIGH
Whether an LLM agent avoids duplicate side effects on failed writes is determined by the fault type: frontier models handle read-back-resolvable failures well (0.5% duplicates) but duplicate in 56-74% of ambiguous cases, where only tool-contract mechanisms like idempotency keys (dropping duplicates from 28% to 4%) can fix it, while the harness barely matters.
BEARS ON 2 · RIGOR HIGH
Post-training quantization of TTS models to the same bit width produces wildly model-specific outcomes because the sensitive component differs unpredictably across architectures, but a staged ablation plus calibration procedure can identify and repair it.
BEARS ON 1 · RIGOR HIGH
What builders are reading · 3 reads
Tooling & models 1
Automatic project memory for Claude Code; pattern for persistent coding-agent context worth piloting internally
Security & governance 2
NSA is spending billions testing AI models; signals government model-security bar regulated firms may be measured against
Court upholds Anthropic supply-chain-risk designation; may constrain vendor approvals if you build on Claude
Blind spots
Nothing the desk knows it missed today.
Snorkel raised $350 million at a $3.5 billion valuation with $375 million in annualized revenue.
The Zhenwu V900, clusterable into units of 500,000, is a concrete non-US vendor capability claim backed by a 2032 global capacity target. The open question is whether Alibaba can sustain that capex without US supply chains.
Alibaba targets 20 gigawatts of data center capacity by 2032.
Read the signal →Business banker openings at major competitors jumped 105% as banks chase business owners for cross-selling. BofA is pairing aggressive hiring with AI tools to speed underwriting, and targeting 30% client share, up from 27.6%.
Job openings at major competitors rose 105% over the past year.
Read the source →The column
Let us close the week with an accounting theme: everywhere you look, someone is spending money to defend against a bill that has not arrived yet.
The scam economy has an allocation problem. The sharpest number in the Thomson Reuters report is not the $20 billion in losses, it is the 2% to 5% recovery rate. When recovery is functionally zero, fraud stops being a reimbursement expense and becomes an industry, which is what generative AI has made it. The report's diagnosis, misaligned liability among banks, platforms, and telecoms, is really a claim that nobody owns the externalized cost, so nobody invests in preventing it. What would strengthen that reading: any regulatory move assigning liability upstream, say to platforms that host the lures. What would weaken it: banks quietly improving detection well above that 10% to 15% range, which would suggest the market can fix this without a mandate.
Everyone is prepaying future liabilities, including cryptographers. European regulators warn in the Q-Day advisory that harvest-now-decrypt-later attacks mean data stolen today is already compromised in 2032, so they want post-quantum cryptography adopted by end of 2026. It is the same structure as the fraud story in reverse: pay now to avoid a bill later, with the payment due before anyone can prove the threat is real. Meanwhile the actual plumbing is being rebuilt too. The Clearing House's tokenized deposit network, spanning 25 banks and bridging to RTP and CHIPS by mid-2027, is the rare infrastructure story where the standard arrives before the fight over the standard. That coordination is worth more than the tokenization itself.
The data layer keeps getting paid. Snorkel's $3.5 billion raise is the week's clearest evidence that verified training data is a distinct line item, not a byproduct. Eighteenfold revenue growth to $375 million annualized tells you enterprises will pay for provenance, which supports the broader view that training-data IP liability is now a quantifiable balance-sheet cost for AI vendors. If verification is a cost center buyers demand, the vendors who cannot document clean data get excluded from procurement, and the industry segments on paperwork. The counterpoint is that synthetic data could eventually make the whole verification layer cheap to substitute; Snorkel's own hybrid model hedges that, which is telling.
One honest wrinkle: all this prepaying assumes the spending produces revenue. The same week brings Alibaba's Zhenwu V900 and a 20-gigawatt bet against a backdrop where announced aggregate capex materially exceeds what will be deployed. Betting on 2032 capacity is the ultimate pay-now-for-maybe-later trade.
The banks fighting over small-business bankers with AI underwriting tools at least bill someone this year.
Research & evidence 4
Banks & payments 1
Elsewhere 1
Under a hard evaluation budget, honest confidence intervals for a benchmark score on a fixed grid need replication, and the optimal interval width follows sharp laws: Θ([M(t+1)]^-1/2) when every task must be observed, and Θ([M(t+√M)]^-1/2) when tasks may be omitted; a disagreement-aware 'Joint' interval converts this into practical finite-budget gains.
BEARS ON 1 · RIGOR HIGH
A sequential policy (SCALE) that selectively queries a noisy LLM judge and escalates only some AI-scored items to human review can achieve a statistically valid population-level hypothesis test at near-optimal first-order cost, cutting total evaluation cost by up to roughly 36 to 39 percent versus the cheaper single-source (human-only or AI-only) baseline in a benchmark instance.
RIGOR HIGH
Sophisticated cache-eviction algorithms provide little to no benefit over simple LRU for LLM prefix caching because session-paced reuse makes recency unusually predictive, and gains should instead come from targeted additions like quick demotion and compute-aware eviction.
RIGOR HIGH
Long-context model performance on corpus tasks depends on the task's complexity class: tasks requiring quadratic or more oracle calls (high CTC) degrade far faster with corpus size than linear tasks, and reverse standard modeling conclusions about efficient attention variants.
BEARS ON 1 · RIGOR MEDIUM
AI-moderated interviews match human-moderated interviews in qualitative depth and cover more themes and customer needs per dollar, but the extra adaptivity of AI moderation adds no quantitative predictive benefit for digital twins over static interviews.
RIGOR HIGH