HackerNews Digest

Mesh LLM is a peer‑to‑peer platform that aggregates GPUs and memory from any number of machines into a single OpenAI‑compatible inference service (localhost:9337/v1). Each node runs an iroh endpoint, identified by a public key, which establishes authenticated QUIC connections (with NAT‑traversal and relay fallback) to other nodes without a central server. Requests are handled in three ways: locally on the node’s GPU, routed to a peer already hosting the model, or split across several nodes in a pipeline (“Skippy”) that partitions a model by layer ranges (e.g., layers 0‑15, 16‑31, …) so modest hardware can run very large models, up to 235 B‑parameter mixture‑of‑experts giants. The system ships with 40+ models, from laptop‑sized to massive, and uses a custom gossip layer to manage mesh membership, version compatibility, and trust. Installation is lightweight (~18 MB), supporting private or public meshes, and a forthcoming mobile app will use iroh’s Swift SDK and the ACP agent standard. Mesh LLM is open source and already deployed on hundreds of thousands of devices.

Researchers at Drexel University discovered that simple, low‑elasticity fluids—such as a hydrocarbon blend of hydrogen and carbon—can undergo brittle fracture when subjected to sufficient extensional stress. Using extensional rheology, the team repeatedly stretched the liquid and observed loud “pop” events captured by high‑speed cameras, revealing crack propagation speeds of 500–1,500 m s⁻¹, far faster than the ~0.07 m s⁻¹ seen in viscoelastic polymers. The fracture occurs at a critical tensile stress of roughly 2 MPa, proportional to the product of viscosity and strain rate, and appears independent of fluid elasticity. The mechanism aligns with earlier predictions that cavitation‑induced void formation can nucleate cracks in any liquid, with propagation unhindered by elastic energy absorption. Findings suggest that liquid fracture may be governed by cohesive intermolecular forces rather than elasticity, with potential implications for processes such as fiber spinning, inkjet printing, propeller design, and soft‑robotic materials.

The project implements a self‑hosting Scheme compiler written in the subset of Scheme it compiles. The compiler can rebuild itself, producing byte‑identical output verified on every change, and all tests run through both compilation stages. In WebAssembly it uses native GC objects: fixnums are unboxed i31refs, pairs and records are GC structs, and equality (`eq?`) is a single reference comparison, eliminating a shadow heap in JavaScript. Hygienic macros are supported via `syntax-rules` and `syntax-case` with fenders, nested ellipses, and `datum->syntax`, executed in a compile‑time interpreter. Functions have typed references with fast per‑arity entry points and generic closure entries; variadic calls and `apply` are cheap, and tail calls compile to `return_call`, enabling constant‑stack loops (e.g., 100 M iterations in ~150 ms). Continuations (`call/cc`) and `dynamic-wind` use the WebAssembly exception‑handling proposal, giving O(1) capture and minimal normal‑path overhead. The stack includes a reactive web framework (SX templates, fine‑grained signals, HTML renderer, JS FFI), a 3D layer built on Three.js and raw WebGL with S‑expression shaders, and a Scheme‑to‑Scheme RPC backend (Igropyr) that exchanges S‑expressions directly. Libraries follow R6RS‑style `library` forms with import resolution and dead‑code pruning.

The input consists solely of the title “Ant, a lightweight JavaScript runtime” with no additional content; therefore, no substantive summary can be generated.

The author experienced intermittent failures waking a docked laptop using a CalDigit TS3 Thunderbolt dock, where keyboard or mouse input, monitor power cycling, and undocking were required before the display reactivated. Thunderbolt 4’s specification (2020) promised reliable wake‑from‑sleep on docked systems, yet the CalDigit TS4 (released 2022) still showed occasional wake failures. After replacing an aging BenQ monitor with an ASUS ROG Swift PG27UCDM (240 Hz) in 2025, the wake‑up issue disappeared; the laptop now resumes and the desktop appears within a second. The author suspects the previous monitor’s firmware or latency caused the problem, though firmware updates to the MacBook Pro 16″ M1 Max and ThinkPad X1 Carbon Gen 9 may also have contributed. Configuration details are shared for users facing similar dock‑wake reliability issues.

RISCBoy is an open‑source, portable handheld console implemented entirely in synthesizable Verilog 2005 and designed for FPGA deployment. It features a RISC‑V‑compatible RV32IMC CPU that passes the official RISC‑V compliance suite and formal verification checks, along with M‑mode CSRs, exception handling, and vectored interrupts. The system includes a raster graphics pipeline, display controller, AHB‑lite/APB bus fabric, memory controllers, UART, GPIO, and other peripherals. The primary target hardware is a Lattice iCE40‑HX8k (7 680 LEs), though a reduced RV32I core can run on the iCE40‑UP5k, and experimental support exists for Lattice ECP5 boards using internal 256 KiB synchronous RAM. Build flow relies on the open‑source IceStorm toolchain, a custom RISC‑V GNU toolchain (RV32IMC/rv32i) and Xilinx ISIM for simulation. The repository is organized into submodules covering HDL, documentation, PCB KiCad files, and test suites. PCB Rev A is a 5 × 5 cm 4‑layer design compatible with iTead prototyping; Rev B is pending further gateware and bootloader maturation.

- The paper describes how to create digital drawings on the 9front operating system using its native program **paint(1)**. - Any input device that can generate mouse events works: mouse, touch screen (left‑click only), finger‑painting, Wacom‑style pens, and external tablets (e.g., GAOMON S620). Keyboard shortcuts supplement missing mouse buttons. - paint(1) stores images in the internal **image(6)** format (`*.pic`). Files are saved with **w**, opened with **r**, and can be piped through external commands (e.g., `rotate(1)`) using `|`, `<`, or `>` redirection. - The canvas is defined by the furthest placed pixels; it expands automatically and can be trimmed with crop operations. Background colour is set with the fill tool by clicking outside the drawn region; it does not alter canvas size. - No pressure sensitivity or layers are supported; drawing behaves like paper with a single raster layer. - Core shortcuts: `0‑9` select brush size, `f` fill, `c` clear, `u` undo last continuous stroke, `-`/`+` zoom, `Escape` show full canvas. - For larger artworks, multiple paint(1) windows can share a file, allowing palette consistency across sessions.

The paper examines how cells expend energy when performing biochemical computations, focusing on the classic Berg‑Purcell problem of estimating external ligand concentration. Using a simple two‑component network that implements a noisy version of the Berg‑Purcell strategy, the authors calculate the thermodynamic cost of “learning” the concentration. They show that accurate inference requires breaking detailed balance, which necessarily consumes free energy, and that the amount of energy dissipated scales with the precision of the inferred concentration. The analysis highlights that energetic constraints can limit the design and operation of cellular signaling pathways, especially in resource‑limited contexts such as bacterial spore germination networks. The work provides a quantitative link between information acquisition and metabolic cost in cellular systems.

Igropyr is a fault‑tolerant web server written entirely in Chez Scheme, separating a minimal core (exposed via `http‑listen`) from an optional Express‑like layer (`create‑app`, `app‑get`, `send‑json!`). It runs thousands of lightweight green processes on a single OS thread, using continuation‑based preemptive scheduling; processes communicate via pure message passing (`spawn`, `send`, `receive`, `link`, `monitor`) with no shared state. A fixed worker pool under a supervisor automatically restarts crashed workers and retries tasks up to three times, killing workers stalled >30 s. Failure hooks can return structured JSON instead of plain 500 responses. Key features include: - Per‑dialogue green processes that preserve state across multi‑request conversations, with guaranteed rollback on crash or TTL. - Hot code swapping of handlers/routes without dropping connections. - WebSocket (RFC 6455) support, streaming/SSE via `res‑begin!`/`res‑write!`/`res‑end!`, and a full OTP‑style toolkit (`gen‑server`, process registry, PubSub). - Robust JSON parser/writer, form and cookie handling, middleware suite (sessions, CORS, security headers, access logging). - Non‑blocking Redis/MySQL clients, outbound HTTP/WebSocket clients, async file reads via libuv, gzip compression, transparent chunked decoding. - Rate limiting, global error handling, Prometheus metrics (`/metrics`), runtime introspection (`http‑stats`), graceful shutdown, and multi‑process scaling via `SO_REUSEPORT`. Benchmarks report ≈35 k req/s at 500 concurrent connections on an Apple Silicon laptop.

- Neocloud providers CoreWeave and Nebius are scaling AI infrastructure by securing large GPU allocations and optimizing utilization, targeting 1.7 GW (CoreWeave) and 0.8‑1 GW (Nebius) of active power by 2026‑2027. - Microsoft and Meta have pledged roughly $122 billion in long‑term capacity contracts with neoclouds, a commitment comparable to ~90 % of AWS’s trailing‑12‑month revenue; additional AI labs (OpenAI, Anthropic) add further undisclosed commitments, bringing total potential contracts above $145 billion. - These firms gain rapid access to Nvidia’s latest GPUs—Nvidia has invested $2 billion in each company and holds stakes dating back to 2025—creating a circular financing loop where Nvidia funds neocloud expansion and backs unsold GPU capacity (e.g., a $6.3 billion purchase obligation for CoreWeave). - CoreWeave’s 2025 Q1 results show $2.08 billion revenue versus $7.7 billion capex, generating –$4.71 billion free cash flow; debt rose to $24.9 billion and a $3.5 billion senior note issuance is planned. - Nebius holds $9.37 billion cash with $8.45 billion debt, but its midpoint 2026 capex guidance of $22.5 billion implies a $6‑7 billion funding gap, likely to be filled via debt or equity programs. - Both companies rely heavily on GPU‑backed delayed‑draw term loans, some rated investment‑grade due to contracts with hyperscalers; rising interest rates increase financing costs, pressuring profitability. - The overall model hinges on hyperscalers shifting capex to opex, neoclouds’ ability to convert contracted power to active capacity, and sustained Nvidia financing amid widening cash‑flow versus capex gaps.