HackerNews Digest

Kimi K3 is a 2.8‑trillion‑parameter open‑source model featuring Kimi Delta Attention, Attention Residuals, and a 1‑million‑token context window. It uses Stable LatentMoE with 16 active experts out of 896, Quantile Balancing routing, Per‑Head Muon, SiTU activation, and Gated MLA, enabling ~2.5× scaling efficiency over Kimi C2. The model achieves frontier‑level performance on internal coding and agentic benchmarks, often surpassing proprietary Claude Fable 5 and GPT 5.6 Sol, though it still trails them overall. Demonstrated capabilities include long‑horizon kernel optimization, autonomous development of MiniTriton (a Triton‑like compiler with MLIR IR and PTX codegen), procedural 3D game creation via Three.js/WebGPU, autonomous 45 nm chip design for a nano‑model, and end‑to‑end scientific workflows (e.g., reproducing astrophysics relations in hours). Knowledge‑work features include interactive research reports, financial visualizations, and multimodal video editing. Weights will be released July 27 2026; inference is recommended on ≥64‑accelerator supernodes with MXFP4/FP8 quantization and a custom vLLM prefix‑cache implementation. Availability spans Kimi.com, Kimi Work, Kimi Code, and the Kimi API (pricing $0.30/MTok input cache‑hit, $3.00 cache‑miss, $15.00 output). Limitations involve sensitivity to preserved thinking history and occasional over‑proactiveness.

Microsoft has released the source code for Comic Chat, its 1996 IRC client that rendered chat messages as comic‑style panels with illustrated characters, speech bubbles, and expressive gestures. Originally built in Visual C++ 4.0 using MFC, the program parsed textual cues to select poses, facial expressions, and panel layouts, effectively performing real‑time editorial illustration. The project was conceived by David “DJ” Kurlander in Microsoft Research’s Virtual Worlds Group, with contributions from Tim Skelly, David Salesin (SIGGRAPH ’96 paper) and artist Jim Woodring. The open‑source release includes original snapshots and experimental updates that compile with modern Visual Studio, connect to current IRC servers, and run on high‑resolution Windows systems. By publishing the code, Microsoft aims to preserve a notable experiment in early online communication, provide a historical resource for developers and researchers, and invite community‑driven ports, enhancements, and new applications. The repository is available on GitHub.

LM Studio Bionic is an AI agent designed for open‑source models, offering coding assistance, document handling, and voice interaction while prioritizing privacy and cost control. Key features include: - **Local and cloud model execution**: Users can run models on‑device, connect via LM Link, or access frontier open models through LM Studio Secure Cloud, with zero data retention for cloud requests. - **Offline voice transcription**: Integrated voice keyboard uses Mistral AI’s Voxtral multilingual model for real‑time, local transcription in any application. - **Coding support**: Bionic can inspect, edit, debug, and search local codebases, presenting inline diffs; compatible with models such as GLM 5.2 and Kimi K2.7 Code. - **Document productivity**: Handles PDFs, slides, spreadsheets, and other files in sandboxed work projects, offering organization, summarization, web‑search augmentation, automatic checkpoints, and in‑app previews. - **Local model management**: Allows downloading and running LLMs directly via the LM Studio runtime. The Bionic app is separate from LM Studio, requires an LM Studio account for cloud billing, and promises ongoing enhancements as open models advance.

Decoy Font is a TrueType font that encodes two different letters in the same glyph using spatial‑frequency separation: a low‑frequency blurred background forms the “hidden” message, while high‑frequency thin outlines convey a decoy text. When viewed up close, the outlines dominate and AI image‑to‑text models (e.g., ChatGPT, Gemini 3.5, GPT‑4) read the decoy; from a greater distance or when squinting, the low‑frequency mass reveals the intended message. The font implements the hybrid‑image principle demonstrated in optical‑illusion studies (e.g., Einstein/Marilyn Monroe composite). It can be installed like any TTF font, allowing typed text to carry concealed content without animation. Decoy Font is positioned among “anti‑AI fonts” that aim to hinder automated OCR and LLM scraping, though powerful models with advanced prompting may still recover the hidden text. The author suggests applications such as CAPTCHA, private messaging, and multilingual extensions, and invites further experimentation via a downloadable TTF file and an online playground.

The page is an Earthquake Event application that requires JavaScript to function. Users are instructed to enable JavaScript or use alternative options such as Real‑time Notifications, Feeds, and Web Services. Browser compatibility information is provided via a link to supported browsers.

The experiment compared two frontier‑level language models—Claude Fable 5 and GPT‑5.6 Sol—in an autonomous music‑video generation task. Each model received the same inputs (the song “Uptown Funk,” a brief description, and a timestamped lyric transcript) and accessed six tools: a planning tool, web search, budget check, image/video generation (via FAL or Replicate), and a local ffmpeg shell for analysis and editing. Four runs were performed (each model at $25 and $100 budgets). Generation spend was $36.57 (Sol $100) and $48.60 (Fable $100); token costs added $3‑4 for Sol and $17‑25 for Fable. All runs produced full‑length videos; three used pure text‑to‑video, while Sol $25 employed an image‑to‑video pipeline and Sol $100 mixed three video models. Common shortcomings included inconsistent characters, literal lyric interpretation, weak tempo matching, and lack of iterative editing. Sol $25 showed the most inventive overlay effects; Fable $100 yielded slightly higher visual coherence but at higher total cost (~$73.65). Both models relied exclusively on FAL for generation. The open‑source arena (github.com/hershalb/music-video-arena) allows replication with custom songs and budgets.

None

- Developers using LLMs experience a paradox: higher output but increased mental fatigue from supervising and correcting generated code. - The volume of AI‑produced pull requests forces rapid judgment calls, eroding the traditional “dopamine” rewards of problem‑solving and collaborative coding. - Continuous prompting creates an “intensity trap”: more tasks can be started, but the brain‑bound review step limits completion, leading to longer work hours and isolation. - The shift mirrors the 2009 transition to responsive design—core expertise remains valuable, but the focus moves from low‑level detail to system‑level judgment, architectural taste, and nuanced decision‑making. - Effective LLM integration relies on deep domain knowledge; shallow‑skill areas yield plausible yet incorrect output. - New practices are emerging: pre‑mortems with separate LLM sessions to expose plan gaps, and tools like AGENTS.md that encode past review rules for consistent LLM guidance. - The bottleneck is now human attention and engineering judgment, making those capacities scarce and valuable despite the automation of code writing.

- NotebookLM, launched at Google I/O 2023 as Project Tailwind, has been renamed Gemini Notebook. - Over 30 million users and 600 000+ organizations currently use the tool for tasks such as creating interactive onboarding material and generating audio/video summaries from notes. - Gemini Notebook remains a standalone research application but will now integrate more broadly across Google services, including the Gemini app and Google Search. - An infrastructure upgrade is being rolled out: each notebook receives a secure cloud‑based compute instance, enabling native code execution for complex data analysis tied to source material. - The cloud compute feature is initially available to Google AI Ultra users and Workspace business customers with AI Ultra or AI Expanded Access, with a staged rollout to all Pro users on the web in the coming weeks. - The upgrade expands supported output formats and allows deeper analytical capabilities within notebooks.

The repository hosts the companion material for *The Little Book of Reinforcement Learning*, a concise introduction covering fundamentals to applied algorithms. It includes: - **algos/**: PyTorch implementations of the algorithms discussed in the book, ranging from Monte Carlo methods to Proximal Policy Optimization (PPO). - **supplementary/**: Detailed explanations and rigorous proofs for dynamic programming algorithms briefly presented in the main text; authored in 2021. - Ongoing additions are planned, and a printable version of the book is linked. - The book is released under a non‑commercial Creative Commons BY‑SA 4.0 license.