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

Small-business AI use reaches 61% in borrower survey
07/31/26, 03:21 PM

AI use reached 61% among 805 surveyed U.S. small businesses with working-capital loans, up from 58% in the first quarter, while 91% of users reported a positive business impact. The OnDeck and Ocrolus report pairs the June survey with cash-flow data from more than 3.76 million financing applications, though its borrower sample should not be read as representative of every small business. The three-point quarterly rise broadens the adoption side of the enterprise AI value divide: vendors selling to smaller operators should bring workflow-specific outcome evidence, because self-reported use is spreading faster than independently measured return.
Wire

Parlance brings AI-briefed transfers to 400 health systems
07/31/26, 03:19 PM

Parlance added AI-generated call summaries and verified patient context to human transfers across a voice platform that serves more than 400 health systems and has answered over 1.9 billion calls. Its product announcement says Webex Contact Center agents receive the caller’s identity, verification status, intent, requested provider, and options already discussed; managers can also change greetings, outage messages, and FAQ answers without opening a support ticket. For teams watching consumer health AI outrun the clinical privacy perimeter, the operational lesson is narrower but useful: judge healthcare voice automation at the handoff, where context fidelity and minimum-necessary data exposure matter more than a polished synthetic voice.
Wire

AZIO commits $2.4M to wire a 500MW AI campus
07/31/26, 03:17 PM

AZIO AI committed about $2.4 million to AT&T fiber services for its planned 500-megawatt Texas AI data-center platform. The company’s agreement announcement says the network is intended for training, inference, GPU cloud, colocation, and other high-performance workloads, but warns that financing, utility availability, and deployment timelines remain uncertain. Set beside Nvidia’s $50 billion Texas capacity loop, the contract is a useful reminder for infrastructure planners: securing megawatts is only one dependency; campuses also need contracted low-latency paths before that power becomes sellable compute.
Wire

RLPF lifts runnable code from 11.1% to 54.6%
07/31/26, 02:19 PM

RLPF fine-tuning raised Qwen3-32B’s correct-and-runnable rate on PerfCodeBench from 11.1% to 54.6% and its relative efficiency from 8.1% to 38.6%, according to the new reinforcement-learning study. The method ranks failed programs by execution progress, then ranks correct programs by runtime improvement toward an expert reference, supplying a useful reward before and after correctness. For teams building the verified loops described in the coding-agent stopping-rule debate, the operator implication is narrow but important: passing tests should be the gate, not the final reward, when generated systems code has a performance budget.
Wire

DeepSeek V4 Flash lifts agent scores 13 points
07/31/26, 02:17 PM

DeepSeek released V4-Flash-0731 with a Terminal Bench 2.1 score of 82.7, up 20.9 points from the preview’s 61.8 and 10.6 points above V4 Pro Preview, according to the MIT-licensed model card. DeepSeek also reports 54.4 on DeepSWE versus 7.3 for the preview, while leaving the API model name unchanged. The update closes the agentic deficit identified when DeepSeek V4 first paired near-frontier capability with a 98% price discount, but operators should rerun their own harness before treating vendor benchmarks as a drop-in production verdict.
Wire

LG opens a 750B-parameter Korean model
07/31/26, 02:15 PM

LG AI Research released K-EXAONE 2.0 with 750 billion total parameters, 37 billion active per token, and a 262,144-token context window under Apache 2.0, according to the official model card. LG says the sparse model averaged 70.1 across 24 benchmarks versus its predecessor’s 63.3, while its serving recipe calls for two nodes carrying 16 H200 GPUs—so it is a sovereign-scale artifact, not a workstation model. For operators comparing Korean models after Solar Open 2 put a 250B agent model on two H200s, K-EXAONE widens the capability ceiling but also makes infrastructure cost the first gate in any self-hosting evaluation.
Wire

Revolut bundles ChatGPT Go for 75M customers
07/31/26, 01:22 PM

Revolut will offer its more than 75 million customers three to 12 months of ChatGPT Go at no extra cost, depending on their plan. The companies’ partnership announcement says paid-plan access starts now, while the three-month Plus and Standard promotion will arrive over the coming weeks. For consumer-AI builders, the bundle reinforces the argument that distribution can matter more than the assistant slot: OpenAI just acquired a bank-sized trial funnel without asking users to discover another subscription.
Wire

Unitree opens a 40.45M-share robotics IPO
07/31/26, 01:20 PM

Unitree Robotics opened the IPO process for 40.45 million shares, or 10% of its enlarged capital, after shipping more than 5,500 humanoid robots in 2025. Xinhua reports that the Chinese company expects first-half 2026 revenue of 1.052 billion to 1.128 billion yuan, up 35.62% to 45.41% year over year. For robotics builders, the offering puts public-market scrutiny behind the hardware side of the 20x simulation-to-hardware testing funnel: commercialization will now be measured in shipped machines and revenue, not demonstrations alone.
Wire

Anthropic campus seeks a $15B Google backstop
07/31/26, 12:32 PM

Nexus Data Centers is reportedly in advanced talks for a $15 billion loan to build an Anthropic campus in Hubbard, Texas, with Google backstopping Anthropic using its investment-grade credit. CNBC’s account of the financing says Morgan Stanley is leading the talks, while Anthropic, Nexus, and Morgan Stanley declined to comment. After Nvidia’s circular $50 billion Texas lease showed vendors reserving their own capacity, operators should file this as another sign that frontier-model supply now depends as much on a sponsor’s balance sheet as on model quality.
Wire

MiniMax H3 generates 15-second 2K video
07/31/26, 11:16 AM

MiniMax launched H3, an omni-modal model that generates up to 15 seconds of 2K video with native stereo audio from combinations of text, images, video, and audio. The company’s H3 announcement says 2K output costs less than one-third as much per second as unspecified mainstream models and promises open weights in the coming days, while the live API documentation confirms 4-to-15-second generation and 2K output. For creative teams comparing it with FLUX 3’s gated 20-second audiovisual model, H3 is worth testing now, but its self-reported price advantage and future-tense weight release still need independent verification.