Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

NEXTDC prices A$1.1B for AI data centers
09/10/26, 05:31 AM

NEXTDC priced A$1.1 billion of subordinated convertible notes to fund its Australian data-centre development pipeline, with a 1.75% coupon and a 2031 maturity, according to the company’s ASX release. It expects about A$1.006 billion net after capped-call costs, so operators should treat this as committed financing—not commissioned megawatts—and keep delivery gates on power, grid connections, land, and customers alongside Figure’s staged Nscale compute reservation.
Wire

Wiz finds 9.6% of LiteLLM instances exposed
09/10/26, 04:27 AM

Wiz found 294 of 3,074 internet-facing LiteLLM instances—9.6%—accepted a default master key or required no authentication, and says patches are available for the disclosed flaws. Wiz’s advisory says one related CVE entered CISA’s Known Exploited Vulnerabilities catalog, turning an AI gateway’s configuration into a cloud-security concern because it can hold provider credentials and connect to MCP tools. Teams running LiteLLM should inventory public exposure, patch, and enforce unique credentials before expanding gateway scope; MCP’s stateless migration does not remove this deployment risk.
Wire

Stilla joins Meta after eight months
09/10/26, 04:26 AM

Stilla, an AI agent startup launched 8 months ago, is joining Meta, keeping its service live while its team and technology move into Meta’s business-AI effort. Stilla’s announcement says its agents have handled work for companies serving millions, but gives no purchase price or customer-count disclosure. Operators should file this as talent-and-product consolidation—not proof that Meta’s business agents have shipped a measurable capability—alongside Meta’s local-agent bet.
Wire

Claude Marketplace launches with six partners
09/10/26, 04:25 AM

Anthropic launched Claude Marketplace with 6 enterprise partners—GitLab, Harvey, Lovable, Replit, Rogo, and Snowflake—letting buyers apply part of an existing Anthropic spend commitment to partner tools. The official Claude Marketplace page says Anthropic handles invoicing, but does not disclose what share commitments can move, so procurement teams should treat it as a consolidation option, not a new budget line. That matters alongside Claude Code’s control-plane shift: compare contract portability with actual workflow lock-in before renewing.
Wire

Q2D-Web tests agents on 190M documents
09/10/26, 03:32 AM

Q2D-Web introduces an agentic-RAG retrieval benchmark built on a 190-million-document web corpus and 70,000 search queries in 10 languages. The paper benchmarks 13 lexical, dense, and late-interaction retrievers and reports that a subcorpus containing one-third of the documents preserved full-corpus model rankings, while absolute Recall@1000 rose only 3–7 points in the full evaluation. Retrieval teams should add agent-reformulated, multilingual queries to acceptance tests rather than trust human-query leaderboards, alongside the search API cost ledger when sizing production RAG.
Wire

Alibaba prices Accio at $3.69 for 107 tasks
09/10/26, 03:30 AM

Alibaba.com says its Accio commerce agent completed a 107-task benchmark for an estimated $3.69, versus $9.27 for OpenAI Codex and $9.51 for Anthropic Claude Code. The company attributes the gap to commerce-specific models, cache reuse, context compression, and coordinated agent execution; its benchmark release says Commerce Agent Bench draws on 1.6 million conversations and 200,000 execution traces from 10 million active SMB users. Small-business builders should read this as a vendor-run routing and workflow result, not a universal model price list, and compare their own task mix against the price-per-point lesson from Gemini 3.7 Flash before switching.
Wire

AWS benchmark leaves AI security below 10%
09/10/26, 03:28 AM

AWS’s new Deception Benchmark finds that none of 12 tested model configurations kept both false-positive and false-negative rates below 10%. The benchmark covers 14,822 samples in 16 languages and more than 70 CWE categories; AWS reports direct prompting caught up to 95% of real vulnerabilities while flagging 41–99% of safe code, and proof-of-exploit prompting lowered false positives but missed 7–44% of real flaws, according to AWS’s benchmark release. Security teams should treat model-generated vulnerability findings as triage until they measure both error rates on their own codebase, extending the agent-evaluation governance case rather than shipping an accuracy-only scorecard.
Wire

Listen Labs walks from $125M round amid Salesforce talks
09/10/26, 02:29 AM

Listen Labs, a voice-AI market-research startup, walked away from a signed $125 million Series C term sheet at a $1.5 billion valuation as Salesforce reportedly discussed a roughly $2 billion acquisition. TechCrunch’s account says Listen has about $30 million in annualized revenue, making the reported acquisition talk roughly 67 times revenue; neither financing nor deal is finalized. For builders buying voice-led customer research, hybrid voice-AI pilots remain a workflow decision, not proof that CRM consolidation has set a reliable price: treat the $2 billion figure as reported negotiation context, not a market-clearing value.
Wire

OpenAI backs four California AI-safety bills
09/10/26, 01:27 AM

OpenAI announced support for 4 California bills covering independent AI-safety assessments, auditor standards, protections for young people, and safeguards against AI-enabled biological threats. In a Reuters report carried by KFGO, the company says SB 813 would build an independent-assessment framework and AB 1405 would set registration, independence, transparency, and accountability requirements for AI auditors; it also says SB 1119 covers companion chatbots and AB 1864 gene-synthesis screening. For builders, the number to file is four: California’s state-by-state AI law gap is becoming an implementation problem for any product that touches high-risk decisions, minors, or biological workflows.
Wire

OpenAI deepens Samsung chip partnership
09/10/26, 01:25 AM

OpenAI says its joint research and production work with Samsung now covers next-generation chips, following its June 2026 launch of Jalapeno, its first custom inference accelerator. A Reuters report carried by The Standard says Samsung and SK hynix had already signed letters of intent to supply memory for Stargate, while Samsung declined to confirm customer details. For operators, this is a capacity signal rather than a performance result: pair the Samsung work with Arm’s push into data-center silicon and wait for node, yield, and volume disclosures before treating a second supply path as delivered capacity.