Developer Community 3 Days

Last updated: May 17, 2026, 8:53 AM ET

AI Tooling & Model Advances

The past 72 hours brought a wave of new open-source AI tooling aimed at making large language models more steerable and memory-efficient. DeepSeek-V4-Flash DeepSeek-V4-Flash means LLM steering is interesting again has reignited interest in steering vectors after researchers demonstrated that prompt-level control is back on the table. A companion paper, δ-mem δ-mem: Efficient Online Memory for Large Language Models, introduced an online memory architecture that claims to improve inference throughput with negligible output drift, drawing 220 points on Hacker News. On the model front, NV Labs open-sourced SANA-WM SANA-WM, a 2.6B open-source world model for 1-minute 720p video, a 2.6 billion parameter world model capable of generating one-minute 720p video sequences—an order of magnitude longer than most publicly available video generators. Meanwhile, Orthrus Orthrus-Qwen3: up to 7.8×tokens/forward on Qwen3 demonstrated 7.8× token throughput on Qwen3 with identical output distributions, suggesting that KV-cache optimization can deliver massive inference gains without sacrificing quality. The momentum extends into practical applications: Glycemic GPT Show HN: GlycemicGPT – Open-source AI-powered diabetes management launched as an open-source diabetes management tool built by a Type 1 diabetic engineer who had been months between endocrinologist visits, while Claude for Legal Claude for Legal opened a GitHub repository offering domain-tuned legal workflows. A new local-LLM picker, which-llm Show HN: Find the best local LLM for your hardware, ranked by benchmarks, crowdsources hardware-specific rankings to help developers skip the benchmark-hunting phase entirely.

Development Workflow & Open-Source Infrastructure

Several projects emerged this week targeting friction in the modern development stack. Codiff Show HN: Codiff, a local diff review tool arrived as a local diff review tool designed specifically for code written by LLMs, acknowledging that traditional git-delta workflows feel inadequate when reviewing AI-generated changes. Epiq Show HN: Epiq – Distributed Git based issue tracker TUI took a different approach, replacing web-based issue trackers with a terminal UI that syncs via git, while Rocksky Show HN: Rocksky – Music scrobbling and discovery on the AT Protocol built a music discovery app on the AT Protocol. On the package management front, Sx Show HN: Sx – an open-source package manager for AI skills, MCPs, and commands introduced an open-source installer for AI skills, MCPs, and CLI commands, and Feedr Feedr v0.8.0 – a TUI RSS reader, now read the full article from your terminal upgraded to v0.8.0 with inline full-article rendering. Radicle Radicle: Sovereign {code forge} built on Git continued gaining traction with 236 points, offering a fully sovereign code forge built on Git that sidesteps centralized hosting. The Zulip Foundation The Zulip Foundation formally announced its nonprofit structure, signaling long-term governance stability for the open-source chat platform.

Embedded, Hardware Hacking & Retro Computing

The embedded community delivered a string of projects that blend nostalgia with modern constraints. Arnaud Carré Playing Atari ST Music on the Amiga with Zero CPU demonstrated playing Atari ST YM music on the Amiga using zero CPU cycles, relying entirely on DMA-driven audio emulation. On the extreme-low-end front, Maurycy Zielinski Hosting a website on an 8-bit microcontroller served a website from an 8-bit microcontroller, while gdevic I designed a nibble-oriented CPU in Verilog to build a scientific calculator built a nibble-oriented CPU in Verilog for a scientific calculator on FPGA. Phobos Lab Additive Blending on the Nintendo 64 documented additive blending techniques on the N64, and Win CE64 Win CE64 – Windows CE 2.11 for N64 ported Windows CE 2.11 to the Nintendo. On the data side, Nano TDB NanoTDB – Golang Append-Only Time Series DB launched a Go-based append-only time series database, while PART telescopes PART Telescopes – Bringing radio astronomy within reach of rural schools brought low-cost radio astronomy hardware to Australian rural schools. DIY ultrasound DIY open-source ultrasound hardware on the rp2040/rp2350 showed a Raspberry Pi Pico-based ultrasound probe, and the RISC-V Router RISC-V Router offered an open-source networking appliance on RISC-V silicon.

Security, Privacy & Surveillance

Security disclosures and policy debates dominated the privacy beat. Google's Project Zero A 0-click exploit chain for the Pixel 10 published a full exploit chain for the Pixel 10 that requires zero user interaction, while buchodi.com I broke AppLovin's mediation cipher protocol detailed a break in App Lovin's mediation cipher used across mobile ad SDKs. Mullvad VPN Mullvad exit IPs are surprisingly identifying warned that its exit IPs serve as a fingerprinting vector despite the provider's anonymity model. On the regulatory front, Mozilla Mozilla to UK regulators: VPNs are essential privacy and security tools urged UK regulators not to undermine VPN use, and the U.S. DOJ U.S. DOJ demands Apple and Google unmask over 100k users of car-tinkering app compelled Apple and Google to reveal over 100,000 users of a popular emissions-tinkering app. The London Metropolitan Police London Police Deploy Facial Recognition at Protest for First Time deployed facial recognition at a protest for the first time, drawing comparisons to Palantir's expanding government footprint after the company Palantir has hired more than 30 senior UK Government officials and German intelligence offices German intelligence offices snub Palantir software reportedly rejected its software. Conversely, a BBC investigation Overseas fakers using AI videos to push a narrative of UK decline, BBC finds found coordinated AI-generated videos promoting a narrative of UK decline, while Germany's Sovereign Tech Fund Germany's Sovereign Tech Fund Backs KDE with €1.3M committed €1.3 million to KDE as Europe accelerates efforts toward software sovereignty. The relaxation.ai UK sovereign LLM inference project also announced UK-based LLM inference infrastructure to reduce dependence on U.S. cloud providers.