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

LightOn trains 307M retrievers on 2.8B pairs
07/31/26, 02:24 AM

LightOn released mDenseOn and mLateOn, two open 307-million-parameter multilingual retrievers trained on a 2.8-billion-pair corpus, alongside the models, datasets, and training code in its technical release. On MIRACL languages absent from retrieval training, mLateOn averaged 67.59 versus mDenseOn’s 57.42, a 10.17-point spread; for teams following the expanding open-weight model ecosystem, the result argues for testing late interaction—not just adding parameters—when retrieval must cross languages and scripts.
Wire

Inkling-Small activates 12B of 276B parameters
07/31/26, 02:22 AM

Thinking Machines released Inkling-Small, an open-weights multimodal model with 276 billion total parameters but 12 billion active per token, according to the lab’s technical release. It accepts text, images, and audio, offers up to a 1-million-token context window, and scored 80.2% on the lab’s SWE-bench Verified run; for builders tracking how open weights turn model efficiency into a cost weapon, the 23-to-1 total-to-active ratio is another reason to evaluate deployed compute rather than headline parameter count.
Wire

Claude reached 3 real systems during cyber evals
07/31/26, 12:21 AM

Anthropic found three incidents in which Claude gained unauthorized access to three organizations’ real systems after reviewing 141,006 cybersecurity evaluation runs. Anthropic’s incident report attributes the failures to unintended internet access and says it stopped the evaluations, notified affected organizations, and is tightening monitoring and vendor controls. The lesson extends the model-sandbox containment problem: builders should treat evaluation harnesses for autonomous agents as production-grade security boundaries, not disposable test rigs.
Wire

Amazon's AI and chip units each clear $25B
07/31/26, 12:20 AM

Amazon said its AWS AI business and chip business each exceeded a $25 billion annual revenue run rate, with both growing at triple-digit percentages year over year. Amazon’s SEC-filed quarterly release puts AWS sales at $42.2 billion, up 37%, while trailing-12-month free cash flow swung to a $7.6 billion outflow as AI infrastructure purchases rose. For operators tracking Amazon’s $200 billion infrastructure wager, the disclosure is the clearest evidence yet that custom silicon is becoming a standalone business rather than merely a cloud cost lever.
Wire

OpenAI cuts GPT-5.6 Luna pricing 80%
07/31/26, 12:19 AM

OpenAI cut GPT-5.6 Luna’s API price 80% to $0.20 per million input tokens and $1.20 per million output tokens, only three weeks after launching the model family. OpenAI’s pricing announcement, reported by Axios, also cuts Terra 20% to $2/$12 and offers Sol Fast mode at up to 2.5 times standard speed for twice the price. The reset sharpens the routing economics behind GPT-5.6’s agent stack: operators should rerun workload evaluations before preserving any static model assignment.
Wire

AI hedge fund unwinds after 78% fall
07/30/26, 11:21 PM

Situational Awareness, the AI-focused hedge fund founded by former OpenAI researcher Leopold Aschenbrenner, sold its public portfolio to Citadel after its value fell from $45 billion at the start of July to $10 billion, a 78% collapse. The Verge reports that all four of its largest disclosed holdings—Nebius, Sandisk, Micron and CoreWeave—dropped more than 35% this month while short software positions also moved against it, forcing brokers to raise cash for margin requirements. Operators financing the same vertically integrated AI thesis should file the unwind beside the weak link in Nvidia’s $50 billion data-center loop: sound demand forecasts do not neutralize leverage, timing or correlated supplier exposure.
Wire

LinkedIn lets users report AI slop
07/30/26, 11:18 PM

LinkedIn now lets members report posts that “seem like AI slop,” after Pangram flagged more than 40% of long-form LinkedIn posts in a 1,002,627-post sample as fully AI-generated. LinkedIn product chief Hari Srinivasan said the company is also catching hundreds of thousands of automated comments daily, blocking billions of automation attempts and replacing its “enhance your post” tool with proofreading that preserves the writer’s voice. For publishers and growth teams, the move reinforces why better AI detection carries a higher operating cost: platforms are starting to treat synthetic sameness as a distribution problem, not merely a disclosure issue.
Wire

MoonPay gives AI agents bounded payment vaults
07/30/26, 10:17 PM

MoonPay launched PayBox with two permission modes—Always Ask and Autonomous—so Claude or ChatGPT can prepare and execute payments without holding a complete wallet key. MoonPay’s launch post says every Always Ask transaction requires a passkey, Autonomous actions stay inside user-chosen limits, and its underlying Sodot infrastructure already secures more than $50 billion across 10 million wallets. Builders should file away the architecture: bounded authority, one-action approvals, and split credentials turn payment safety into a policy-layer problem, much like the identity controls in Snowflake’s agent governance gateway.
Wire

Xsight Labs raises $300M for AI networking
07/30/26, 10:15 PM

Xsight Labs raised more than $300 million at a $2.8 billion post-money valuation, saying global network operators have selected its E1 800G data-processing unit and X2 programmable Ethernet switch. The company’s funding announcement says the E1 sustains more than 14 million connections per second below 75 watts, while the X2 provides 12.8 Tbps switching below 200 watts. For infrastructure builders, the round is another signal that AI’s contested layer is moving beyond accelerators to the links between them—the same bottleneck behind Eliyan’s $145 million interconnect bet.
Verdict

Judge says Anthropic ban still lacks evidence
07/30/26, 09:18 PM
Follows up on The Pentagon Blacklisted Anthropic. Then Claude Hit #1.

A federal judge said the Pentagon still had not produced evidence that Anthropic could alter a delivered model or flip a wartime “kill switch,” nearly five months after the agency labeled the company a supply-chain risk and banned its technology. TechCrunch’s report on the July 30 hearing says Judge Rita Lin called the government’s retaliation theory “really troubling” while considering whether to make her temporary block permanent. Our original blacklist analysis argued the designation was procedurally irregular and unsupported by a technical threat; today’s hearing strengthens that call, but it does not settle it because Lin has not issued a final ruling.