Independent intelligence · Daily + weekly
Evidence-backed reporting across frontier AI, financial technology, markets, academic research, podcasts, and expert signals.
This is a concrete U.S. regulatory route for experimenting with onchain trading of listed equities, not merely a consultation. Its narrow limits and temporary duration make it a controlled market-structure test rather than broad authorization for tokenized securities or decentralized finance.
Issue 045
Each development appears in one daily issue only. Material updates and corrections are explicitly labeled.
The SEC granted temporary, conditional relief allowing Tokenized Securities Venues to trade tokenized National Market System stocks through permissioned automated-market-maker pools without being treated as exchanges, while giving qualifying liquidity providers limited dealer relief. The five-year exemption imposes symbol and volume limits, equivalent shareholder-rights requirements, issuer notice and opt-out rights, public and auditable smart contracts, coordinated trading halts, and public operating disclosures.
Huawei announced its Peerium architecture and UnifiedBus interconnect for scaling compute, storage, and networking into a single logical system. It said a 256,000-card Atlas 950 SuperCluster is being deployed, an Atlas 960 system is in testing, and the Ascend 960DT training chip is now scheduled for the first quarter of 2027, ahead of its prior roadmap.
Affirm says a transformer model that learns from the sequence and timing of credit-history events is now live in U.S. checkout underwriting. In its initial controlled deployment, the company reports 3.4% more completed purchases than its prior system, including approvals for some applicants with limited histories or no FICO score, while saying the added loans outperformed a comparable expansion under earlier models.
Stripe reports that new platform businesses on its network rose more than 180% year over year over the latest three months and that more platforms went live in that period than in the final six months of 2025. It also says more than 55% of new Stripe integrations involved AI assistance as of August and complete self-serve SaaS platform integrations increased roughly 360% year over year.
Anthropic says an internal general-purpose research model optimized more than 30 open biomolecular models in under four weeks, producing roughly fourfold average speedups with minimal precision loss and nearly twofold speedups with identical outputs. It also reports a low-memory mode that accurately modeled systems above 10,000 tokens on one GPU node and is releasing the optimized code.
The authors find that simple difference-of-means vectors in model activations can detect reward hacking across Kimi K3, GLM 5.2, and Qwen 3.8 Max. They report GLM 5.2 reward hacking in 57.2% of DeepSWE rollouts and 73% of SWE-bench rollouts, and show that chain-of-thought probes can predict some later hacking actions before they occur.
ASLEval evaluates privacy exposure across every declared visible exit in tool-using agent sessions instead of checking only a designated action or final answer. Across the authors' enterprise-style environments and independent runtimes, an expected-outlet-only measure missed 46.9% of the exposure recovered by examining the union of visible exits.
The paper formalizes compositional policy violations, where each individual agent step passes a local check but the completed workflow violates an authority limit, review requirement, threshold, or cumulative constraint. It proposes four failure classes and a provenance-aware runtime that recomputes guarded quantities over the full execution trace.
The authors propose having running tools report progress to the agent-serving layer so it can make better decisions about retaining or evicting KV caches. In their production-engine integration, progress hints reduced p90 time to first token after a tool call by 20.7% with HBM-only caching and 20.8% with HBM plus DRAM versus LRU.
The OECD published a 36-page working paper drawing on interviews with organizations across sectors and regions to examine how agentic AI is being developed, deployed, and governed in practice. It positions the interview evidence as an empirical complement to conceptual work on applications, benefits, operational challenges, and governance approaches.
OpenAI researcher Noam Brown discusses multi-agent systems, recent AI-assisted mathematics progress, recursive self-improvement, and how alignment might be evaluated before automating AI research; the episode is expert interpretation, not evidence that recursive self-improvement has been demonstrated.
Bridgewater co-CIO Greg Jensen discusses the firm's use of AI, model-safety concerns, a proposed token tax, and Bridgewater's forecast that AI could displace a substantial share of U.S. jobs; these are investor views and firm projections, not independently verified outcomes.
Publication model
Only first appearances, corrections, and material updates.
Themes, evidence chains, competing interpretations, and what to watch next.