TNP - 2026-11-17 It's Tuesday, November 17. Welcome to Trending Not Peak. Today, explore a street by listening, keep Python results between runs, and peek inside a console's display firmware. AI helped write this episode, and you're hearing a synthetic voice. Explore the stories at trending not peak dot com. Explore a neighborhood through speech and questions First up, StreetReaderAI makes street-level imagery something you can explore without seeing the pictures. Google Research has released this accessibility web application, built with and for blind and low-vision users. Keyboard shortcuts or voice commands move you through street images, while spoken descriptions and follow-up questions help you investigate the surroundings. The application combines panoramic imagery with OpenStreetMap routing and nearby points of interest. That gives accessibility researchers and interface developers a concrete starting point for virtual neighborhood exploration. StreetReaderAI lets you explore streets virtually, including moving along blocks and across intersections. It's an imagery-based preview; it doesn't establish current conditions on a real walk. The Apache-licensed code can run locally or load in Google AI Studio. A useful experiment would revisit a familiar block and ask about a particular landmark. Then examine whether the description gives you enough spatial context to ask a better next question. Help a fast drafter notice the picture Turning to inference, MMFlash investigates why speculative decoding gets awkward when an answer depends on an image. A lightweight model proposes tokens, and the larger target model checks them together. That saves time when the proposals survive verification. The researchers found acceptance falling from forty-six percent for the least visually dependent quarter of tokens to twenty-two percent for the most dependent quarter. MMFlash adjusts drafter training to give extra weight to visually dependent tokens it predicts poorly. For long videos, it also uses sparse attention to reduce redundant work. Across its evaluated multimodal benchmarks, the authors report lossless decoding speedups up to three point five eight times. Serving engineers could investigate this when a fast text drafter struggles with visual questions. Compare image-dependent answers with ordinary descriptive text, and measure the complete request. The interesting opportunity is faster generation that still lets the target model decide which proposed words belong. Keep Python function results after the process exits Closer to everyday Python, funkycache packages a persistent cache around a function decorator. Alex Chan built it after finding subtle bugs in the custom caching code behind a static site generator. Decorate an expensive function, and its result goes into SQLite, where another run can reuse it. Changes to the function's source file invalidate its cache. You can also declare dependencies on files and static values, such as a template or a library version. Build-tool authors could use funkycache to cache Python function results across repeated runs without writing retrieval and storage logic for every function. The best starting point is a deterministic operation with no side effects, like transforming an unchanged input file. Run it twice, then edit a declared dependency and check that the result rebuilds. That last step makes the experiment useful: saving time is lovely, provided yesterday's answer doesn't quietly survive today's input. Compare local gesture decisions across Apple's engines Over on Apple Silicon, d1 gestures turns webcam frames into typed decisions using Liquid AI's d1-two-B. Each frame gets questions about hand gesture, facial expression and their locations. One forward pass returns probabilities without generating a text response. Ivan Fioravanti's new comparison runs the same model through MLX, Core AI on the GPU, and Core AI on the Neural Engine. On one MacBook Pro M-five Max, he reports roughly sixty-six to sixty-seven milliseconds for the GPU paths, versus one hundred ninety-seven milliseconds on the Neural Engine, with substantially lower reported power on that slower path. Interface experimenters have a concrete way to compare responsiveness with power draw. The benchmark used video because starting the live camera itself increased Neural Engine activity. Keep that distinction when reproducing it. Begin with a repeatable gesture clip, then test the complete camera application. Your preferred engine may depend on how quickly the interface actually needs its next decision. Thanks to Intelligrit Between local model experiments and training supervision, thank you to Intelligrit, this episode's sponsor. Follow your next curious idea to intelligrit dot com. Let agents supervise an ongoing training run Back from the break, AutoRL gives agents a job that usually involves somebody watching curves and wondering what went wrong. Its Perceptors monitor training behavior, rollouts, evaluations and efficiency. When something stalls or degrades, Proposers diagnose the problem and develop interventions. An Orchestrator reviews the recommendation, then applies approved changes through a live update or a checkpoint restart. The paper reports experiments with Qwen-three and LongCat-Flash-Lite that outperform static reinforcement-learning recipes and a general agent. Training researchers could investigate this architecture for routine supervision of a long experiment, especially when the next adjustment depends on what the run is doing now. The authors identify expensive validation as a continuing obstacle. A proposed change still needs evidence that it helps. A focused experiment would give the supervisor a known training fault, record its diagnosis, and compare recovery with a fixed recipe under a disclosed compute budget. Rust nightly puts more cores into frontend compilation Next, Rust's compiler frontend now runs in parallel by default on nightly. That includes work such as type checking and borrow checking, which previously ran sequentially by default even though the LLVM backend already used multiple threads. The change has been enabled since the November fourteenth nightly, with the compiler team seeking real-project issues before stabilization. In its benchmark suite, eight frontend threads reduced mean clean cargo-check time for individual leaf crates by about forty-four percent. Incremental rebuild gains averaged a much smaller three to five percent, with the team acknowledging weak benchmarks for that case. Rust developers with slow checks have a useful trial available today. Compare clean checks and your ordinary edit-check cycle separately on a test checkout. Those are different waiting problems, and this release gives you a chance to discover which one extra frontend parallelism actually improves. Prove that a valid trade can actually finish For a verification example, Harmonic has launched Rust support for Aristotle, alongside a case study with Manifest's open-source Solana order book. The company reports twenty-eight proved properties, including conditions under which a valid swap completes and an accounting of where swap errors originate. It also describes two findings that Manifest corrected. One involved rounding: a fill could meet the requested target, then fail during a subtraction that should no longer have happened. Another involved a conversion that could fail on a cancellation path even though its result wasn't used. Smart-contract developers and verification learners get a concrete example of asking whether valid operations make progress, alongside checking that funds remain protected. Each result includes a Lean statement the team can inspect. Follow the rounding example from its arithmetic failure to the corrected stopping rule. The useful habit is making successful completion an explicit property you can examine. Run a Git forge from one binary Back to project infrastructure, rbit is a newly introduced self-hosted Git forge written in Rust. Its current site identifies version zero point two point two and describes one binary backed by SQLite and an embedded store, with no additional services to run. Git works over HTTPS and a built-in SSH server. Teams, deploy keys, releases, mirrors and webhooks give a small project room to grow beyond a bare repository directory. Developers maintaining a private lab or a small collaboration could use rbit to host Git repositories while exploring a compact deployment. The site also lists two-factor sign-in, scoped tokens and an audit log; those features still deserve hands-on checking. Start with a disposable repository and another collaborator account. Push a branch, limit access, and follow a webhook through to its receiver. That gives you a practical tour of the pieces your own forge would need. Follow a cropped Wii picture into an eight-bit controller And finally, GaryOderNichts has gone digging into why Wii games can look worse through a Wii U's HDMI output than on an original Wii. The new account examines forced cropping and color bleed inside Wii Mode. Behind the picture sits a little eight-bit controller that translates the old Wii display registers into commands for the newer graphics hardware. To investigate its firmware, Gary first wrote a processor specification for Ghidra, drawing on related processors the tool already understood. The reverse engineering revealed a viewport configured to simulate overscan, cropping up to sixteen pixels. Changing the firmware bytes changes those viewport settings. Console-preservation tinkerers and reverse-engineering students get a wonderfully concrete path from a visible annoyance to the register values causing it. Follow that path through the writeup. Sometimes the missing edge of a game sends you all the way into a tiny processor hiding behind the television picture. A few useful places to begin: StreetReaderAI to explore streets virtually, funkycache to cache Python function results, and rbit to host Git repositories. Find the projects and sources at trending not peak dot com. This is Trending Not Peak. Have fun following a small question into a project, even if that question starts with a strangely cropped Wii game.