David A. (@logbydavid) header
David A. (@logbydavid) profile picture

Exploring AI, crypto & technological progress. Software Engineer, AI/ML. Curiosity over hype. Long-term thinking.

The Metaverse
Joined June 2016
Photo by David A. (@logbydavid)Photo by David A. (@logbydavid)Photo by David A. (@logbydavid)Photo by David A. (@logbydavid)Photo by David A. (@logbydavid)Photo by David A. (@logbydavid)Photo by David A. (@logbydavid)Photo by David A. (@logbydavid)Photo by David A. (@logbydavid)Photo by David A. (@logbydavid)
Run Inkling 2 Mini (26.4B parameters) uncensored locally on a laptop with just 12 GB of memory. In tests, its refusal rate dropped from 100% to 0%, while retaining 98% of its original accuracy on the MMLU benchmark. Three quantized GGUF versions are available, ranging from 8.67 to 16.53 GB.
38
Claude Code 2.1.330 introduces a built-in easel module for creating charts directly in conversations. Claude can visualize data from JSON, CSV, tables, and query results. Charts can also be generated automatically when an MCP server returns metrics over a specific time period. Supported formats: - Line and bar charts - Horizontal bars - Sparklines - Proportion charts One example uses a local JSON file with aitmpl.com download statistics to visualize daily downloads by client type: AI agents, users, and scripts.
1
41
If you use Claude Code regularly but still treat it as little more than a chatbot in your terminal, Florian Bruniaux’s open-source Claude Code Ultimate Guide is worth exploring. The French engineer has built a practical encyclopedia spanning more than 430,000 lines, recently surpassing 6,400 GitHub stars. It covers everything from initial setup to advanced orchestration of multiple AI agents. What makes it different from typical AI agent tutorials is its focus on production engineering practices rather than elaborate prompting. Some of the most useful material includes: - Hooks with ready-to-use scripts to block dangerous database migrations and deletion commands before execution. - Automatic type checking and code formatting after every file change. - Context and cost management through hierarchical CLAUDE.md files, on-demand Skills loading, and selective MCP server integration. - 244 ready-to-use template files, an Architecture Decision Record (ADR) generator, and security audit checklists. - Protection against hidden prompt injection instructions embedded in malicious repositories. The gap between casual AI-assisted coding and production-grade workflows is growing. While beginners still copy error messages into chat and ask for fixes, experienced teams use Hooks, custom Skills, and strict context isolation to make AI-driven development controlled and reproducible. This repository can save hundreds of hours of learning and infrastructure setup. github.com/FlorianBruniaux/c…
64
dependency-cooldown adds an extra layer of security to Claude Code by blocking package versions published less than 7 days ago via npm, pnpm, or uv. It helps: - Reduce software supply chain attack risks. - Prevent slopsquatting attacks involving malicious package names hallucinated by AI. - Require your approval if Claude tries to bypass the restriction. The protection also covers MCP servers and hooks launched by Claude Code. Install it at the user level to protect all projects: npx claude-code-templates@latest --setting environment/dependency-cooldown
4
65
Yesterday, $BTC on Kraken flash-crashed to $16.9K due to a liquidity gap. Today, the exchange confirmed that the order was still executed. Kraken users really got lucky 🤔
62
Matt’s 70-minute video shows how he tackles a massive 21,000-line PR using his own custom skills. His workflow: - Starts with /grill-with-docs - Uses /retro to analyze the work and improve the process - Manually reviews the code - Splits the massive PR into several sequential PRs - Focuses on broader patterns and architectural decisions rather than every line - Runs /retro again after the review to continuously improve his review standards It was genuinely moving to see how seriously Matt still takes PR reviews. I now trust AI with almost all code reviews and step in mainly for key decisions. That’s why his level of attention made such an impression on me. youtu.be/jVBlkOF8sHk?si=5t54…
1
1
109
MCP Events introduces a way for AI agents to subscribe to external events and react automatically, without waiting for user requests. For example, an agent can monitor a GitHub repository and automatically run code reviews and tests when a new pull request appears. MCP Events supports three delivery methods: - Poll — periodically query the server. - Push — the server sends events to the client. - Webhook — notify the client through a webhook. With Poll, the client discovers available events, queries the server, waits, and repeats the process. When an event occurs, its information is passed to the AI agent's runtime environment.
1
64
ARTEX, a Chinese AI tool capable of automating cyberattacks, can now reportedly run on a standard iPhone. Originally presented as a security testing tool, ARTEX deploys AI agents that autonomously discover vulnerabilities, select attack methods, and execute multi-step operations without human involvement. According to Microsoft, a China-linked hacking group used ARTEX to attack hundreds of organizations. Check Point researchers also identified more than 1,608 ARTEX control panels exposed to the internet.
75
Gemini usage limits in Antigravity have been reset for all users. Thanks to caching improvements, users can now get approximately 22% more work done within the same quota. ✋
2
1
1
123
Thinking Machines has released Inkling 2 Mini, an open-weight model built on a recurrent transformer architecture. Its middle layers process each token three times, enabling more internal computation while using fewer tokens to solve tasks. Key specs: - 26.4B total parameters, 3.4B active. - Requires around 20 GB of memory. - Scores 48.2% on DeepSWE v1.1, which Thinking Machines says is the best result among open models under 100B parameters. The team also tested it on a plane, asking it to build a flight simulator. Everything ran locally on a Mac, without an internet connection. A demonstration of its coding capabilities completely offline.
3
1
2
113
A method for bypassing Claude's identity verification using fake documents has surfaced online. An X user shared instructions for creating a fake ID and attempting to pass KYC verification. The process uses an online document generator to select a country, upload a photo, choose a background and fill in personal details. The author chose a Japanese driver's license. They also demonstrated how to use Google Chrome Developer Tools to download the generated ID without a watermark. The resulting image can then be submitted to Claude for verification. There's no need to print it: displaying it on a high-quality screen and photographing it is presented as an alternative. Whether the method actually passes verification remains unclear. The author says the instructions are for informational purposes only and users assume all risks.
1
369
A developer found three critical security vulnerabilities in an AI-built appointment booking app just an hour before its planned public launch. The author, who wasn't a programmer, had built the app over the summer using Cursor and Claude, with Supabase, Stripe, OpenAI and Resend. After reading about an $88,982 Codex bill, he asked a backend developer to review the app. Using Chrome DevTools, the developer quickly discovered that three API credentials were exposed in publicly downloaded JavaScript files: - An OpenAI API key. - A live Stripe secret key (sk_live_) capable of performing payment operations. - A Resend API key. The author had added .env to .gitignore and assumed his secrets were safe. But Claude had suggested calling APIs directly from the browser and prefixing environment variables with VITE_ to simplify implementation. Vite automatically bundles those variables into client-side code. The server logs revealed another problem: automated scripts had attempted to access /.env 212 times in just two weeks, despite the site being virtually unknown. The OpenAI API also had no spending limit. Fortunately, nobody had exploited the exposed keys. That evening, he rotated all three, set a hard $20 OpenAI spending cap that blocks requests when reached, and moved API calls into Supabase Edge Functions. Two checks every AI app builder should make before launch: 1. Search every file in the production build for sk-, sk_live_ and key. 2. Set hard API spending limits that automatically block requests when reached, rather than merely sending email alerts.
8
1
155
SEO expert Emre Elbeyoglu has introduced SEO Audit, an open-source desktop application for auditing website SEO. A preview demo shows the tool performing an SEO audit directly from the console.
1
1
2
64
Teams using Claude Code and other multi-step AI coding agents can spend 70–80% of their tokens on setup and environment exploration rather than writing code. Every new session starts from scratch. Whether fixing a bug or adding an API, the agent rereads lengthy SKILL.md files, repeats find and grep commands, inspects package.json or Cargo.toml, searches for build scripts, and figures out test parameters through trial and error. Even when the project structure hasn't changed, the agent repeats the entire process, wasting tens of thousands of tokens. A newly published research paper, SkillCache (arxiv.org/abs/2611.10022), proposes a solution: a semantic skill caching system positioned between the agent runtime and its tools. SkillCache stores the results of repetitive operations and a structured description of the environment. It uses filesystem fingerprints to verify whether cached information is still valid. If the hashes of dependent configuration files and directory structures remain unchanged, the agent receives a compact snapshot of the saved state in milliseconds instead of repeating dozens of tool calls. In tests on real-world coding tasks, SkillCache: - Reduced token consumption by 81%. - Cut overall task execution time by nearly two-thirds. - Maintained the same task success rate. The goal is simple: stop making AI agents rediscover an environment they've already explored.
2
1
122
Anthropic has gone public, raising $81 billion through its IPO, plus another $26 billion from private share sales to major investors. The stock is already trading at around $100 per share, with demand exceeding supply by 7x. This is the largest IPO in history, surpassing SpaceX's record set just six months ago. It may not hold the title for long, though: OpenAI is also expected to go public in the near future.
1
134
Anthropic has introduced Claude Platform OAuth, allowing organizations to connect their Claude accounts to platforms without sharing API keys. 🦀 Key changes: - Connect Claude organizations through OAuth, with no API keys required. - API usage costs are charged directly to the customer's organization. - Developers can choose who pays for each API request. - Users can set spending limits and revoke access at any time. - AI agent platforms can focus on building agents and workflows instead of managing customer API keys.
67
This might be the most impressive AirDrop you’ve ever seen. An enthusiast used GPT-6.2 Astra to create the interaction mechanics, ML-SHARP to turn a single photo into a 3D scene, and Tripo to generate the ginkgo leaves. All running on Apple Vision Pro. Pinch to capture a photo on your iPhone. Open your palm, and you're suddenly inside the picture. 🫣
1
2
246
With high-end GPU prices soaring, enthusiasts are finding unconventional ways to build affordable computing systems. A Reddit project has drawn attention this week: a 40-mini-PC cluster capable of running GLM 5.3 Flash locally. The setup uses 40 diskless Intel N150 mini PCs left over from a previous project. Each has 12 GB of DDR5-4800 RAM, giving the cluster 480 GB of total memory. Initially, connecting the machines over standard 1 Gbps Ethernet limited distributed inference to just 7 tokens/s due to tensor synchronization latency. Instead of buying an expensive 10 Gbps switch, the builder used the x86 chipset's USB 3.2 Debug Capability (USB DbC) to create a direct USB connection tree running at 10 Gbps: 1 → 3 → 9 → 27 nodes. The entire cluster cost around $5,000 and draws just 740 W at peak load. It can run GLM 5.3 Flash at full precision while retaining enough memory for eight parallel context windows of 1 million tokens each. Benchmark results: - Single-stream generation: 14 tokens/s - Generation with a full 1M-token context: 12 tokens/s - Batched decoding throughput: 40 tokens/s - Prompt processing: 85 tokens/s
91
Google has unveiled Gemini 4 Neon. The model matches Opus 5.5 on benchmarks, ranks: - #3 in the Artificial Analysis Intelligence Index - #1 in the Coding Agent Index - SOTA across several scientific and mathematical tests Pricing: $3/$15 per million input/output tokens, $0.15 for cached tokens. 1M-token context window. At high reasoning levels, task-solving costs 32% less than Carbon. Available via API and to Pro and Ultra subscribers. Neon for Cyber is currently limited to a select group of cybersecurity experts.
1
97