-
Introducing System One Models and Jev (919 points by albelfio)
TypeSafe AI announced System One Models and its first public model, Jev, in a blog post by founder Diogo Almeida. Almeida, who previously worked at OpenAI on instruction-following and ChatGPT-related research, argues that chat models have not yet delivered broad automation. System One models are designed for fast, structured decisions that software can consume directly, using a new architecture, a parallel sampler, and Reinforcement Learning for Calibrated Decisions. Jev is in early access, claims similar intelligence on System One tasks to existing LLMs while being roughly two orders of magnitude faster and more efficient, and sacrifices free-form string generation to produce structured outputs without hallucination.
-
Show HN: An e-ink frame that hears birds and draws them as 1800s illustrations (1405 points by arnemunthekaas)
This Show HN project, fugleramme, is an open-source e-ink bird frame built for Raspberry Pi. It listens for birds in real time, detects them using fully local AI, and then displays them as illustrations styled after hand-cut 1800s bird artwork. The GitHub repository includes source code, detector components, docs, Docker setup, and installation tooling. The project highlights a creative, privacy-preserving use of edge AI in a low-power ambient device.
-
Negativland, Culture Jamming, and the Art of Making Something New (25 points by bananaboy)
The Internet Archive blog profiles Negativland, an experimental sound-collage collective founded in 1980. Its first self-titled album had handmade covers and sold 500 copies in three months, eventually reaching 15,000, encouraging the group to continue. Over 46 years, Negativland has produced more than 30 albums and EPs plus fine art, books, radio, and live performances that mix found and original material. The group coined the term "culture jamming," and an upcoming performance at the Internet Archive in San Francisco will feature Mark Hosler, Jon Leidecker, and remote contributions from David Wills.
-
Apple Reference Image: A New Approach for Verified Photography (41 points by imwally)
Apple Security Research introduced Apple Reference Image as a new approach for verified photography. It addresses how widely available AI tools can generate or alter photorealistic images, making it hard to tell real photographs from synthetic or heavily edited ones. Apple argues that certifying an image reflects what a real camera sensor captured requires a chain of trust covering both the sensor and computational photography software. Existing industry approaches based on C2PA attach provenance metadata after capture and certify edit history, but Apple notes this can be compromised at any point in the editing chain.
-
An update on Wayback Machine access (434 points by ChrisArchitect)
The Internet Archive published an update on Wayback Machine access after waves of high-volume automated traffic. Protections were added to keep the service running, including a rewritten message for HTTP 429 "too many requests" errors. Those protections sometimes mistakenly block real users, and the Archive apologizes while working to reduce errors. Users who believe they were blocked in error are asked to email info@archive.org with their operating system, browser, and IP address.
-
Gemini 3.8 Live and 3.8 Live Extended Thinking (340 points by leumon)
Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, described as its most advanced live dialogue models. The models add upgrades in intelligence and parallel reasoning, aiming to make voice collaboration more natural and capable of complex task execution. They can handle complex reasoning, real-time visual context, and background task execution without interrupting the conversation. They are available through the Gemini API, Google Workspace, and the Gemini app, with Live focused on fast fluid conversation and Extended Thinking on harder tasks.
-
German Rheinmetall open-sources its Battlesuite connected weapon system protcol (155 points by summarity)
Rheinmetall open-sourced documentation for onboardapi 9.10.0, the protocol/library for its Battlesuite connected weapon system. The interface library and middleware standardize communication between sensor systems and software components, ensuring interoperability across complex hardware and software environments. It is built on the ddkit SDK and uses OMG's Data Distribution Service (DDS), providing publisher-subscriber, low-latency data exchange with DDS XTypes and XCDR2 backward compatibility. The core library is C++, with wrappers for Java, C#/.NET, and Python; interface descriptions use EPL v2.0 while runtime libraries use a Rheinmetall EULA.
-
Recreating Voodoo Graphics and a Late-1990s Gaming PC on an FPGA (64 points by zdw)
This blog post describes recreating 3dfx Voodoo Graphics and a late-1990s gaming PC on an FPGA. The author built zSST, a SystemVerilog implementation of the Voodoo Graphics SST-1, and combined it with the existing z486 CPU and PC hardware to form z486 XL on a Xilinx KV260 board. The system runs DOS games with the original 3dfx renderer, including Tomb Raider, and implements features such as prepared triangles, texture filtering and mipmapping, depth/alpha tests, fog, blending, dithering, framebuffer access, and buffer swaps. The CPU and renderer run at 100 MHz, with broader compatibility and later Voodoo generations left for future work.
-
Building a Linux GPU Driver for the M4 Mac Mini in One Month (200 points by ADevWithAnIdea)
In this blog post, Cody Ho describes building a Linux GPU driver for the M4 Mac Mini and MacBook Neo with Niklas in about one month. The driver is fully OpenGL ES 3.0 compliant, runs Chrome and Firefox WebGL, and can run Minecraft at 200 fps. The work involved reverse engineering Apple's AGX GPU firmware ABI and user-space components in a transparent clean-room manner. The driver is not yet ready for end users, but the authors aim to support OpenGL and eventually Vulkan, using a previously built hypervisor to help reverse engineer macOS.
-
Stay discoverable in search while disallowing AI training (29 points by djfergus)
Cloudflare announced a new Disallow AI Training setting that lets website owners remain discoverable in search while refusing AI training by the same mixed-use crawler. This addresses the problem that large crawlers often serve both search indexing and AI training, so blocking one can block the other. Apple, Google, and Microsoft honor or have committed to honor the setting. Cloudflare is also working toward controls for AI summaries, aiming to let site owners decide how much content is included, set once on Cloudflare rather than with each operator.
AI/ML Insights & Trends
- Structured "System 1" models are emerging as a separate automation layer from chat LLMs.
- Trend: TypeSafe's Jev and System One models optimize for fast, calibrated, structured outputs rather than free-form text, using a new architecture, parallel sampler, and Reinforcement Learning for Cal
Analysis generated by deepseek-reasoner