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

Thrive’s $516M AI fund now tops $3.7B
08/15/26, 05:44 AM

Thrive Capital’s $516 million 2022 early-stage fund is now worth more than $3.7 billion, a roughly 7.2× gross multiple, after early bets on OpenAI, Anduril, and SpaceX. TechCrunch’s account of Thrive’s first investor letter says the firm has $60 billion in assets under management and that about 90% of each fund’s capital goes into its top 15 investments, while noting the estimate comes from Bloomberg. Builders tracking the archive’s AI rollup thesis should file the return as a concentration-and-access outcome, not a repeatable seed benchmark: the evidence for a new fund is realized liquidity, follow-on rights, and operating value—not a single marked portfolio.
Wire

Aisot closes a CHF 2M seed extension
08/15/26, 05:41 AM

ETH Zurich spin-off Aisot Technologies closed a CHF 2 million seed extension to expand its AI-driven portfolio-intelligence platform for institutional investors. In the company’s announcement, Aisot says its stack combines quantitative analysis, machine learning, and LLM-based news sentiment for forecasting, portfolio construction, and risk management; former Partners Group partner Felix Haldner joined existing investors, family offices, and angels. Wealth-tech builders should read the round as a distribution signal for explainable AI inside investment workflows—not as evidence of returns—alongside the archive’s accounting-agent rollup analysis, where domain operations remain the test of model value.
Wire

OpenAI loses its AI ethics lead
08/15/26, 05:38 AM

OpenAI’s AI ethics lead Chloé Bakalar is reportedly leaving after 12 months at the company, according to Computerworld’s summary of the Financial Times report. Bakalar previously built ethics programs at Meta and focused at OpenAI on how people interact with AI and questions around machine consciousness; the departure follows other senior exits as OpenAI prepares for an IPO. Teams following OpenAI’s model-release gate should file the staffing change as a governance-capacity signal, not proof of a policy failure: ask who owns ethics review, escalation, and release authority when the named specialist is gone.
Wire

DeepSeek opens a four-mode agent harness
08/15/26, 04:40 AM

DeepSeek has released the developer preview of dsh, an open-source agent harness with 4 operational modes: standard, code, creative, and minimal. The official GitHub repository says its architecture treats every component as a plug-in, but warns that compatibility-breaking changes are still coming; reporting on the launch frames it as DeepSeek moving beyond models toward the agent layer. Builders should file this as an open orchestration surface to watch—not a production framework to standardize on until the preview settles—alongside the wider economics of China’s open-model push.
Wire

GitHub funds 50 projects for AI-era security
08/14/26, 10:42 PM

GitHub’s Secure Open Source Fund invested more than $500,000 across 50 projects, giving each maintainer team $10,000 plus a year of security guidance, tools, and cloud credits. The Session 4 report says AI-assisted workflows sped investigation and remediation, but maintainers still supplied the judgment and accountability for what shipped. That is a useful supply-chain counterweight to stacked agent-authored pull requests: more generated code makes funded, human-led security maintenance infrastructure—not just faster review—the scarce layer.
Wire

Vals AI raises $40M for real-world model tests
08/14/26, 10:39 PM

Vals AI raised a $40 million Series A at a $400 million valuation, while reporting 8× revenue growth versus 2025 and a doubled customer base. In Vals’s funding announcement, the evaluator says its private, work-like benchmarks are already cited in model cards from OpenAI, Anthropic, Google, Meta, and xAI; a16z’s account says results can turn around within hours of model access. For teams still choosing models from public leaderboards, the agent-evaluation governance case is the practical companion: benchmark the workflow and its controls, not just the model’s exam score.
Wire

SMIC raises wafer prices as AI demand surges
08/14/26, 10:36 PM

China’s top foundry, SMIC, crossed $3 billion in quarterly revenue as AI-related orders lifted wafer shipments 14% and average selling prices 5.7% sequentially. In Reuters’ account of SMIC’s earnings call, management said third-quarter wafers will carry higher prices, with fabs already running at 93.7% utilization. The signal for SK hynix’s US HBM buildout is upstream: AI scarcity is repricing the supporting chips and foundry capacity around the accelerator, not only the accelerator itself.
Wire

Natural gas could triple AI power costs
08/14/26, 08:49 PM

Natural-gas prices could rise above $10 per million BTUs at some U.S. hubs, up from roughly $2–$4.50 today, as AI data-center demand meets slower supply growth and rising LNG exports, according to energy firm Noreva’s forecast reported by TechCrunch. Because fuel is about half the cost of electricity from a large gas plant, the projected move would turn “bring your own power” from a shortcut into a material token-cost risk; the Noreva market platform tracks the underlying capacity and fuel markets. AI infrastructure buyers should put gas-price sensitivity, hedges, and grid fallback into campus contracts now, extending the archive’s household-bill case from public backlash to operator margin.
Wire

Alibaba puts a 10T-parameter supernode on sale
08/14/26, 08:46 PM

Alibaba Cloud has put its Lingjun Zhenwu M890 supernode instance on sale in Ulanqab, with an 8-node, 64-card unit that can handle mixture-of-experts models above 10 trillion parameters, according to Alibaba Cloud’s product bulletin. Alibaba says the M890 delivers 3× the performance of its Zhenwu 810E predecessor; its cloud overview adds 800 GB/s inter-chip bandwidth and more than 560,000 Zhenwu chips shipped. The practical signal is availability: China’s domestic accelerator story is moving from chip announcements toward rentable, rack-scale inference, giving teams operating under export constraints another reason to benchmark the whole software stack—not just compare peak silicon claims—against the archive’s analysis of China’s AI price war.
Wire

Cursor closes its $60B SpaceX acquisition
08/14/26, 08:43 PM

SpaceX completed its $60 billion all-stock acquisition of Cursor on August 14, issuing 389,289,254 Class A shares and converting roughly 44.4 million options, according to the company’s closing 8-K. Cursor says the combination gives it access to SpaceX’s GPU fleet and will put Grok 4.6 inside a coding product that already has developer distribution; Cursor’s announcement frames the deal as a path to stronger, cheaper models. Builders should file the closing as a change in vendor-control risk, not just an acquisition headline: the earlier analysis of SpaceX’s AI infrastructure now has a live developer surface attached to it, so model neutrality, telemetry, and an exit path deserve written answers before deeper editor lock-in.