AI

Show HN: Git bayesect – Bayesian Git bisection for non-deterministic bugs (github.com) AI

Git bayesect is a Python tool that applies Bayesian inference to automate “git bisect” for flaky or non-deterministic failures, estimating which commit most likely introduced a change in failure likelihood. It uses a greedy entropy-minimization strategy and a Beta-Bernoulli approach to handle unknown failure rates, with commands to record pass/fail observations and select the most probable culprit commit. The README also includes examples and a demo that simulates a test whose failure probability shifts over a repo’s history.

Prompt Engineering for Humans (michaelheap.com) AI

The article argues that “prompt engineering” is essentially the same as good management: providing clear context, constraints, success criteria, and validation so people (and AI) don’t have to guess. Using an example with an agent building a Trello CLI feature, the author shows that vague instructions produced a technically correct but incomplete result, while more specific context led to an immediately usable command. The piece concludes that at scale, ambiguity is costly and managers must design requirements carefully rather than simply assign tasks.

Inside the 'self-driving' lab revolution (nature.com) AI

The article reviews how “self-driving” laboratories are using AI, robotics and automated instrumentation to plan and carry out experiments with minimal human input. It highlights systems such as Ross King’s robotic platform Eve/Adam and GPT-4/LLM-driven approaches that can interpret scientific requests, run multi-step procedures, and even adjust based on experimental “eyes.” While the technology is still early and not a full replacement for human expertise, the piece argues it is already improving speed and lowering some research costs, prompting debate about how biology and chemistry may be done in the future.

Show HN: Claude Code rewritten as a bash script (github.com) AI

The GitHub project “claude-sh” ports Claude Code’s functionality to a ~1,500-line bash script, relying only on curl and jq (optional ripgrep/python3). It supports streamed output, tool use (Bash, Read/Edit/Write/Glob/Grep), permission prompts for non-safe commands, CLAUDE.md project instruction loading, git-aware context, session save/resume, and basic rate-limit retry and cost tracking. The README also documents installation, environment variables, and command-line/slash commands like /help, /cost, /commit, and /diff.

Anthropic Races to Contain Leak of Code Behind Claude AI Agent (wsj.com) AI

The report says Anthropic is working to contain a leak involving the source code behind its Claude AI agent, after portions of the code were exposed, prompting an internal and security response.

CUDA Released in Basic (developer.nvidia.com) AI

NVIDIA released cuTile BASIC, bringing the CUDA Tile programming model (introduced in CUDA 13.1) to the BASIC language. The package lets developers write tile-based GPU kernels using simple BASIC syntax, with parallelism and data partitioning handled automatically, demonstrated with vector addition and matrix multiplication examples. cuTile BASIC requires an NVIDIA GPU (compute capability 8.x+), NVIDIA driver R580+, CUDA Toolkit 13.1+, and Python 3.10+.

AI companies charge you 60% more based on your language, BPE tokens (tokenstree.com) AI

The article argues that AI providers bill for non-standard “tokens” created by different tokenizer designs, which can make the same prompt cost up to ~60% more for non‑English languages. It describes how varying tokenization and provider pricing gaps can significantly change total costs across models and regions. It also promotes TokensTree as an infrastructure layer to normalize token accounting and reduce repeat token consumption via caching (and claims language-toll mitigation).

AI for American-Produced Cement and Concrete (engineering.fb.com) AI

Meta says it is expanding its use of AI to help U.S. concrete producers design mixes that meet performance targets while using more domestically made cement and materials. The company is releasing BOxCrete, an open-source Bayesian optimization model, along with foundational datasets, and describes pilots with partners like Amrize and academic researchers. Meta also reports an AI-optimized mix used in a data center foundation reached full strength 43% faster and reduced cracking risk by about 10% compared with an earlier formula, and that its earlier concrete optimization framework has been adopted in commercial software used for daily quality control workflows.

OpenAI demand sinks on secondary market as Anthropic runs hot (bloomberg.com) AI

Bloomberg reports that demand for OpenAI-linked stakes on the secondary market has weakened, even as Anthropic’s business appears to be drawing stronger investor interest.

What Is Copilot Exactly? (idiallo.com) AI

The article explains that “Copilot” can refer to several different Microsoft AI products (for example, GitHub Copilot, Copilot for Microsoft 365, Windows Copilot, and Copilot Chat), each integrated into different tools and workflows. The author shares a week-long attempt to improve their productivity with Copilot for Teams/Microsoft 365 before realizing others may be using a different “Copilot” entirely. It ultimately frames the confusion as a caution to clarify which specific tool people mean when they say they use “Copilot.”

The OpenAI Graveyard: All the Deals and Products That Haven't Happened (forbes.com) AI

The article reviews major OpenAI deals and product announcements that, despite prior hype, did not end up materializing, mapping out what was promised versus what was never launched or only partially delivered.

Show HN: Real-time dashboard for Claude Code agent teams (github.com) AI

Show HN introduces agents-observe, a GitHub project that provides a real-time observability dashboard for Claude Code and multi-agent sessions. It uses Claude Code “hooks” to stream tool calls, subagent lifecycles, and file/tool activity into a local or remote server that stores events in SQLite and pushes updates over WebSockets to a React UI. The dashboard supports filtering/searching across agent events and viewing the agent hierarchy to make autonomous debugging less dependent on post-hoc logs.

StepFun 3.5 Flash is #1 cost-effective model for OpenClaw tasks (300 battles) (app.uniclaw.ai) AI

The UniClaw OpenClaw Arena cost-effectiveness rankings report that StepFun 3.5 Flash is currently ranked as the most cost-effective model for OpenClaw tasks, based on 300 battles.

Apple Removes iPhone Vibe Coding App from App Store (gizmodo.com) AI

Apple removed the “Anything” iPhone app from the App Store, citing a violation of App Store Guideline 2.5.2 about apps being self-contained and not downloading, installing, or executing code that changes features or functionality. The move follows earlier blocks of “vibe coding” apps such as Replit and Vibecode, which use AI assistance to generate or modify other apps. Apple did not immediately provide details to Gizmodo, while Anything’s CEO says attempts to adjust the app were rejected and that the enforcement appears to be tightening around this category.

We Built It with Slide Rules. Then We Forgot How (unmitigatedrisk.com) AI

The post argues that spaceflight know-how—once built through hands-on experimentation and then preserved in documents like NASA SP-287—has been eroding as organizations grow too complex and stop asking basic operational questions. It recounts the author’s father learning rocket chemistry and working on satellite attitude control, then contrasts that transferable “keep it in your head” approach with modern Artemis planning, which the author says reflects hidden constraints and insufficient familiarity among leaders. The author extends the warning to software and AI, suggesting capability can be outsourced before judgment and underlying understanding are transmitted, leaving teams “renting” complexity without owning the decisions.

I Quit. The Clankers Won (dbushell.com) AI

The author argues that despite claims that blogging is “over,” now is a crucial time to keep writing to preserve authentic human voices in an industry increasingly dominated by AI hype, plagiarism machines, and surveillance. They also criticize generative AI (including Sora) as largely low-value “slop,” and encourage readers to avoid Big Tech narratives and use blogging to support an open, indie web.

AI has suddenly become more useful to open-source developers (zdnet.com) AI

ZDNET reports that open-source maintainers are increasingly finding AI coding and security tools more reliable for real-world tasks, improving report quality and helping with legacy code maintenance. The article also highlights ongoing concerns, including potential legal disputes over AI-assisted rewrites, and the flood of low-quality “AI slop” that can overwhelm projects. Organizations like OpenSSF are working to make better AI tools available to maintainers as reliability continues to improve.

Show HN: Baton – A desktop app for developing with AI agents (getbaton.dev) AI

Baton is a desktop app for running AI coding agents with separate, git-isolated workspaces so multiple agents can work in parallel without stepping on each other. It provides a dashboard to monitor agent status, view diffs and file changes, manage worktrees, and open pull requests from the app, while running CLI agents in real terminal sessions. The project claims code stays local, with optional AI-generated workspace titles/branch names handled via a paid provider and supporting custom or first-class integrations like Claude Code, Codex, and others.

OpenAI closes funding round at an $852B valuation (cnbc.com) AI

OpenAI has closed a record $122 billion funding round at a post-money valuation of $852 billion, up from $110 billion previously announced. The round was co-led by SoftBank and included investors such as Andreessen Horowitz and D. E. Shaw Ventures, and OpenAI also added participation via bank channels plus $3 billion from individual investors. The company is not yet profitable and continues to burn cash as it prepares for potential IPO scrutiny.

Show HN: 1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs (prismml.com) AI

PrismML announces “1-bit Bonsai” models that use 1-bit weights to shrink memory and power requirements for running LLMs on edge devices and in robotics. The company claims the 8B model fits in about 1.15GB of RAM, runs faster and more energy-efficiently than full-precision 8B models, and preserves benchmark performance. It also offers smaller 4B and 1.7B variants designed for on-device speed, with detailed comparisons reportedly covered in a whitepaper.

TinyLoRA – Learning to Reason in 13 Parameters (arxiv.org) AI

The paper introduces TinyLoRA, a parameter-efficient adapter method that scales reasoning performance using extremely small low-rank updates (as few as 13 trained parameters). The authors report that training an 8B Qwen2.5 model with TinyLoRA reaches about 91% accuracy on GSM8K and recovers roughly 90% of performance gains on harder reasoning benchmarks while using 1,000× fewer parameters than typical approaches. They also find the strong results depend on reinforcement learning, with supervised fine-tuning requiring much larger updates to match performance.

Claude Code Unpacked : A visual guide (ccunpacked.dev) AI

Claude Code Unpacked is a visual, source-based guide that walks through how Claude Code works, from user input and an agent “loop” to rendering responses, tool execution, and command handling. It catalogs Claude Code’s built-in tools, slash commands, and optional/hidden features (including unreleased or feature-flagged capabilities), with links to the relevant parts of the codebase. The site is unofficial and notes that some details may be outdated or inaccurate.