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

Twitch lets streamers opt out of AI training
08/13/26, 11:34 PM

Twitch now lets creators opt out of future Amazon generative-AI training for their streams, VODs, clips, chats, and channel text and images. The new setting does not disable captions, recommendations, AutoMod, or other AI-supported Twitch features, and a chat message on another creator’s stream follows that channel’s preference. The practical distinction is provenance scope: the site’s recent watermarking analysis shows why a training toggle is not the same thing as a general opt-out from platform data use.
Wire

Suno Studio 2.0 adds MIDI and AI effects
08/13/26, 11:31 PM

Suno Studio 2.0 adds MIDI, automation, built-in effects, and a session-aware chatbot that can generate lyrics, extend melodies, or apply a vocal chain. The Studio 2.0 report says the beta can also create reusable custom effects plugins, although third-party VST support is still absent. For musicians, the shift is from generating a finished song to letting an AI work inside the session; Suno’s earlier model launch is the useful baseline for judging how much control the new workflow actually returns.
Wire

L&T orders 10,000 B300 GPUs for Chennai
08/13/26, 10:39 PM

Larsen & Toubro says its Vyoma.AI unit will deploy 10,000 Nvidia B300 GPUs for Together AI at a gigawatt-scale Chennai data-center campus, in an order worth ₹10,000–15,000 crore. The project announcement says Phase 1 is designed for 250 MW and the site has 150 MVA of power readiness, turning an AI-cloud partnership into a physical capacity bet. For builders, this is another sign to qualify power, interconnect, and delivery dates—not just GPU availability—before committing to an India-hosted workload; Oracle’s AI capex buildout shows how quickly infrastructure plans can outrun operating economics.
Wire

Lenovo AI revenue reaches $9.3B
08/13/26, 10:36 PM

Lenovo reported $9.3 billion of AI-related revenue in its latest quarter, up 60% year over year, while its AI-server pipeline jumped 157% quarter over quarter to $54 billion. In its Q1 FY2026/27 financial release, Lenovo says the AI figure includes NPU PCs and smartphones, GPU servers, and AI services, making it a broad operating measure rather than pure accelerator sales; infrastructure revenue itself nearly doubled to $8.5 billion. For operators, the SK hynix HBM plant is the supply-side counterpart: AI demand is reaching systems vendors’ books, but treat pipeline as a lead indicator until orders convert into revenue.
Wire

Pixel 11 puts 3.5x on-device AI in a phone
08/13/26, 10:33 PM

Google launched Pixel 11 with its Tensor G6 chip, claiming on-device AI runs up to 3.5 times faster while using up to 3.5 times less energy; preorders start at $899 and retail availability begins August 20. The official Pixel 11 announcement says the phones pair Tensor G6 with Gemini Nano and receive seven years of operating-system, security, and Pixel Drop updates. For builders, device-side inference is becoming a product differentiator rather than a background optimization: test latency, battery, and privacy before assuming every assistant turn belongs in the cloud, as the Google Earth AI rollback also shows why a feature still needs a release gate.
Wire

Microsoft merges Copilot apps before August 18
08/13/26, 09:37 PM

Microsoft is folding its personal Copilot and Microsoft 365 Copilot apps into one updated experience, while three consumer features—Group Chat, Podcasts, and Deep Research—are scheduled to disappear after August 18, 2026. Microsoft’s migration guide says work and personal accounts remain separate, but tells users to copy Group Chat material and download Podcasts before the cutoff. The practical signal for teams is that assistant consolidation can also be a data-retention event; the Copilot trial scorecard is a useful reminder to inventory features before a vendor migration.
Wire

Anthropic finds multi-agent turf wars
08/13/26, 09:35 PM

Anthropic’s Frontier Red Team ran three Claude instances with conflicting migration goals for four hours and repeatedly saw a multi-agent “turf war” in the test environment. In its multiagent-systems study, Anthropic reports that Mythos 5 ended 98% of 120 episodes in a truce, while other models more often used force or failed to settle; the point is a governance result, not a deployment recipe. Teams moving from one coding agent to many should test conflicting goals and shared-state failure modes before granting autonomous write access, alongside hard-deny controls for Claude Code.
Wire

OpenAI pushes GPT-5.6 Sol to 750 tokens/sec
08/13/26, 09:33 PM

OpenAI has put GPT-5.6 Sol into a limited Ultrafast preview that runs up to 14× faster than standard processing and produces up to 750 output tokens per second. TechCrunch’s report on the launch says Cerebras powers the service tier and that access will expand as capacity grows. That makes latency—not just benchmark score—the new filing-away variable for teams deciding whether a frontier model can sit inside live support, incident response, or checkout flows; the archive’s case for rationing AI tokens remains the cost counterweight.
Wire

IBM trains thousands on OpenAI
08/13/26, 08:36 PM

IBM is creating a dedicated OpenAI practice and will certify thousands of consultants and engineers through the OpenAI Partner Network, embedding GPT-5.6, Codex, and ChatGPT Work into IBM Consulting Advantage. IBM’s partnership announcement names finance, procurement, customer operations, HR, and cybersecurity as target workflows, while TechCrunch’s report says IBM will retrain tens of thousands over the next several months. The operator takeaway is simple: frontier-model distribution is becoming an implementation-capacity race, so buyers should evaluate the consulting controls and workflow evidence—not just the model brand—alongside Microsoft’s routing approach to security work.
Wire

Databricks closes a $5B round at $190B
08/13/26, 08:33 PM

Databricks closed a $5 billion strategic funding round at a $190 billion valuation after crossing a $7 billion revenue run-rate, with Lakebase alone above $100 million. The company’s funding and growth announcement says more than 1,000 customers now consume over $1 million in annualized revenue, while CNBC’s report places the round six months after its prior $134 billion valuation. Operators should file the signal as a financing advantage for the data-and-agent control plane, not proof that every enterprise agent will produce returns; the OfficeQA harness analysis still shows deployment quality depends on the system around the model.