Published on July 12, 2026 at 06:00 CEST (UTC+2)
Mesh LLM: distributed AI computing on iroh (147 points by tionis)
Mesh LLM: distributed AI computing on iroh
This article presents a distributed computing framework called Mesh LLM that pools GPUs and memory across multiple machines to run large language models. Instead of relying on centralized cloud providers, users can combine local hardware into a single OpenAI-compatible API. The goal is to give teams more control over model updates, data privacy, and costs. The approach challenges the dominant data-center model by enabling peer-to-peer and split execution of AI workloads.
We Know Simple Fluids Can Flow. Turns Out, Some Can Fracture (29 points by Anon84)
We Know Simple Fluids Can Flow. Turns Out, Some Can Fracture
Researchers have discovered that certain simple fluids, which were thought to only flow, can actually fracture under fast stretching. The work, using extensional rheology on viscous liquids like honey and crude oil, shows that non-elastic fluids can break apart. This finding questions previous assumptions that elasticity was necessary for fracture in fluids and has implications for materials science and industrial processes.
A pure scheme web programming tool (33 points by guenchi)
A pure scheme web programming tool (Goeteia)
Goeteia is a self-hosting web development tool written entirely in Scheme that compiles to WebAssembly in the browser. The page carries its own compiler (~70KB gzipped) and can recompile source code in ~15ms without a server. It features hygienic macros, real closures, tail calls, and a reactive web stack, all built on native Wasm GC objects. The project demonstrates a fully client-side, Scheme-based programming environment.
Show HN: Ant – A JavaScript runtime and ecosystem (206 points by theMackabu)
Show HN: Ant – A JavaScript runtime and ecosystem
Ant is a lightweight JavaScript runtime designed as an alternative to Node.js and Bun. It aims to provide a minimal yet capable environment for running JavaScript applications, though the article preview is brief and details are limited.
I Did Not Kill Stanley Lieber: How to Draw (With 9front) (27 points by c-c-c-c-c)
I Did Not Kill Stanley Lieber: How to Draw (With 9front)
This document is a tutorial on drawing and image manipulation using the 9front operating system's paint(1) program and associated tools. It covers hardware setup (mouse, tablet, keyboard), painting techniques, canvas management, and conversion methods. The author presents it as a guide for creating digital art on an obscure but powerful OS.
RISCBoy is an open-source portable games console, designed from scratch (78 points by mariuz)
RISCBoy is an open-source portable games console, designed from scratch
RISCBoy is a fully open-source handheld game console built from the ground up, including a RISC-V CPU, a raster graphics pipeline, PCB layout in KiCad, and other hardware infrastructure. It aims to be a "Gameboy Advance from a parallel universe where RISC-V existed in 2001." The design is written in synthesisable Verilog and targets iCE40 FPGAs.
A dock that wakes up reliably (43 points by ingve)
A dock that wakes up reliably
Fabien Sanglard critiques Thunderbolt docks that fail to wake laptops from sleep reliably, even after Thunderbolt 4 promised to fix the issue. He describes his long frustration with the CalDigit TS3 and TS4 docks, noting that the problem persists despite Intel's conformance testing. The article explores the underlying technical reasons and potential solutions.
The Energetic Costs of Cellular Computation (2012) (10 points by lioeters)
The Energetic Costs of Cellular Computation (2012)
This paper calculates the minimum energy required for a cell to compute the concentration of a chemical ligand, building on Berg and Purcell's classic problem. It shows that learning about external concentrations demands energy consumption and breaking of detailed balance. The findings suggest that energetic costs can constrain cellular networks, especially in resource-poor environments like bacterial spore germination.
Nvidia, CoreWeave, and Nebius: Inside the Circular Financing of the GPU Boom (185 points by adletbalzhanov)
Nvidia, CoreWeave, and Nebius: Inside the Circular Financing of the GPU Boom
The article investigates the financial dynamics behind the GPU boom, focusing on “neoclouds” like CoreWeave and Nebius that buy massive numbers of Nvidia GPUs. These companies rely on debt and Nvidia’s own investments to scale, creating a circular financing loop. While hyperscaler demand is high, the neoclouds are far from profitable, raising concerns about sustainability and a potential bubble.
A Erlang style pure Scheme Webserver and further (14 points by guenchi)
A Erlang style pure Scheme Webserver and further (Igropyr)
Igropyr is a fault-tolerant web server written in pure Chez Scheme, inspired by Erlang's "let it crash" philosophy. It uses supervised worker pools that automatically replace crashed or stuck handlers, with preemptive scheduling to prevent CPU loops from freezing the system. It also supports hot code swapping and graceful shutdown for live deployment.
Decentralized AI compute as a counterpoint to hyperscaler lock-in
Mesh LLM exemplifies a growing movement to pool local GPUs for LLM inference rather than relying on centralized cloud providers. This trend matters because it gives teams control over model versioning, data privacy, and costs. The implication is that as AI becomes ubiquitous, we will see more distributed, peer-to-peer infrastructure that rivals or complements today’s data-center model.
The fundamental energy cost of computation – a lesson for AI efficiency
The cellular computation paper shows that even simple biological computations incur a minimum energy cost due to thermodynamic constraints. This is directly relevant to AI: training and inference consume enormous energy, and understanding these fundamental limits can guide the design of more efficient algorithms and hardware. The takeaway is that energy-aware AI (e.g., spiking neural networks, low-precision compute) will become increasingly critical.
Circular financing in the GPU market signals potential overinvestment risk
The analysis of Nvidia, CoreWeave, and Nebius reveals that much of the GPU boom is fueled by debt and reinvestment from Nvidia itself. This matters because it questions the long-term financial health of neoclouds and the sustainability of hyperscaler demand. Investors and AI startups should monitor leverage ratios and cash flow – a correction in GPU prices or demand could destabilize the entire AI infrastructure ecosystem.
Rise of lightweight, self-contained runtimes for edge AI
Projects like Ant (JavaScript runtime) and Goeteia (Scheme → WebAssembly) indicate a trend toward minimal, portable runtimes that can run AI models at the edge. Coupled with Mesh LLM’s distributed approach, this suggests that future AI systems will be deployed on diverse hardware – from browser-based Wasm to embedded devices. The implication: developers should prioritize model portability and low-latency inference across platforms.
Fault-tolerant design patterns entering AI server infrastructure
The Igropyr web server brings Erlang-style supervision to Scheme, offering automatic crash recovery, hot code swapping, and preemptive scheduling. As AI services become mission-critical, we will see more adoption of such patterns (supervisor trees, bounded retries) in AI inference servers. The takeaway is that AI deployment stacks need to borrow from telecommunications and distributed systems to ensure reliability under load.
Open-source hardware and RISC-V as an alternative for AI acceleration
RISCBoy, while a games console, demonstrates that a full RISC-V system can be designed from scratch and run on FPGAs. For AI/ML, RISC-V’s open instruction set enables custom accelerator development. This trend matters because it could reduce dependency on proprietary GPU architectures, especially for specialized inference tasks. Expect more open-source AI chipsets based on RISC-V extensions.
The intersection of physics and computer science – fluid fracture as a metaphor for AI system limits
While not directly about AI, the discovery that simple fluids can fracture under stress mirrors how AI systems can “break” under extreme workload conditions (e.g., catastrophic forgetting, adversarial perturbations). Understanding these emergent failure modes in physical systems may inspire new robustness techniques in machine learning. The insight: we should study non-linear failure patterns across disciplines to build more resilient AI.
Analysis generated by deepseek-reasoner