Topic
AI-Assisted Development
Using LLMs and AI tools to write, debug, and ship code.
20 published pieces
Published work
Stop Scripting CI/CD in Python or Bash! Save Your Memory with Rust Scripts
Why the AI age flips the economics of scripting. Use single-file Rust scripts for deterministic DevSecOps automation with 8.4x speedups and 85 % memory savings.
AI · September 25, 2026xAI Grok Subscription Tiers, SuperGrok Quotas, and API Pricing: The Truth About Cross-Plan Silos and X Platform Benefits
Navigating xAI compute requires understanding three completely separate commercial universes: X social subscriptions, standalone grok.com web tiers, and the api.x.ai pay-as-you-go developer platform. We break down the real benefits of X Premium ($8 / mo) and Premium+ ($40 / mo), explain the $200 and $1,000 monthly ad spend credits on Verified Organizations, expose the total absence of cross-subsidization between SuperGrok and X, audit the weekly shared compute trap on grok.com, and detail why automated builders use headless Chrome CDP agents instead of paying prohibitive $100 to $5,000 monthly X API fees.
AI · September 25, 2026The Definitive Guide to Free AI Credits and Zero-Cost Developer Inference
A benchmarked, verified ledger of zero-cost developer LLM inference across Google AI Studio, NVIDIA NIM, Groq, Cerebras, SambaNova, OpenRouter stealth pools, and Cloudflare Workers AI, featuring an interactive trading card carousel and fallback mesh architecture.
AI · September 24, 2026The $500 / Month ChatGPT Pro Max Leak: Fastest Codex, Cerebras Hardware, and Subscription Value Speculation
Unreleased strings discovered by researcher Tibor Blaho, video demonstrations by Chetaslua, and TestingCatalog reports reveal OpenAI is preparing a $500 / month ChatGPT Pro Max tier featuring Fastest Work and Codex. We fact-check the discovery against live client artifacts, map the Cerebras wafer-scale inference link, examine how an upgraded Work interface targets terminal agent competition, and speculate across four value scenarios: a 50 x capacity expansion, a 0 x margin reset, pure 14 x to 21 x speed acceleration, and a dual-bucket execution model.
AI · September 24, 2026Improving Human Improvement with AI
Douglas Engelbart defined bootstrapping as improving how we improve. Empirical findings from Curtis Northcutt’s StudentBench benchmark, El Salvador’s PISA pilot, and Epoch AI’s 92 % annual cost decline demonstrate that AI tutoring matches expert human instruction at 918x lower cost.
AI · September 19, 2026$200,000 vs. $200,000,000: The Real Production Economics of AI Feature Animation
Gabriel Olson released a viral two-minute DuckTales trailer with 100+ shots from a two-person team. Charlie Driscoll shipped a 22-minute pilot for Tales of Grimwicke solo in four weeks. Combining empirical logs with video token pricing (such as ByteDance Seedance 2.5), we calculate the exact compute, human labor, and timeline required to produce a 95-minute animated feature for $200,000 to $350,000, disrupting Hollywood’s legacy cost structure.
AI · September 18, 2026OpenAI Subscription Plans: ChatGPT Plus vs. Pro $100 vs. $200 Paused, Team, Canvas, and Coding Workflows
Engineering teams evaluating OpenAI face a shifting subscription landscape: ChatGPT Plus ($20 / mo), the newly active Pro 5X ($100 / mo), the paused Pro 20X ($200 / mo), and Team ($25 / user / mo). Beyond messaging limits, subscribers navigate aggregate weekly quota pools, activity-triggered rolling reset drift, DALL-E image allowances, zero native music generation, and distinct allocations for Canvas and Codex. By using flat-rate web GPT Pro for upfront architecture planning and post-validation audits, engineers maximize high-compute reasoning while preserving their subscription coding allocations. Here is the operational developer guide to OpenAI subscription value and compute extraction.
AI · September 18, 2026My Name Is Jev: Summarizing Dozens of Real-World Use Cases, Benchmarks, and Ecosystem Repositories
Most write-ups of TypeSafe’s System One model (Jev) are launch listicles. We stress-test Jev on 50 adversarial invoice edge cases across six models, analyze confidence calibration, survey the 186 curated ecosystem tools and 1,566 consensus candidates on awesomejev.org, evaluate Teknium’s compaction critique, and outline fine-tuning alternatives.
AI · September 18, 2026The Real Token Value of AI Subscriptions: Empirical ROI, 5-Hour Limit Traps, and the Grok Plan Maze
We audited the actual metered token return across ChatGPT Pro, Claude Max, Cursor Ultra, and SuperGrok Heavy based on empirical benchmark evaluations. Here is what a $200–$300 monthly plan actually yields in raw API tokens, why frontier labs heavily subsidize flat-rate power users with enterprise API margins, how the Grok subscription maze traps developers, and how 5-hour rolling cliffs compare to dynamic capacity resets.
AI · September 18, 2026The Pragmatic Developer’s Token Stack: Headless Agents, Open Weights, and Multi-Model Cost Optimization
Software teams adopting AI development tools frequently fall into the single-provider trap, exhausting weekly plan quotas by sending routine boilerplate and test scaffolding to top-tier reasoning models. High-velocity engineering organizations instead deploy a tiered multi-model token stack: routing intermediate tasking to high-volume subscription models (such as Codex sub Luna, DeepSeek V4 Flash, or Gemini 3 Flash) to preserve plan headroom, allocating precision coding models (Claude 5 Sonnet, Gemini 3 Pro) for multi-file architectural work with prompt caching, and strictly reserving scarce reasoning compute (GPT-5.6 Sol, Astra, Claude 5 Opus) for complex algorithmic blockers. Here is the operational architecture for maximizing weekly subscription value and eliminating mid-week lockouts in terminal agent workflows.
AI · September 18, 2026Google AI Plans: Quotas, Family Sharing Gotchas, Jules vs. AI Studio, and the $200 Tier
Google’s consumer and developer AI subscription landscape requires understanding a stark competitive reality. Anthropic has shipped Claude Opus 5.5 and Sonnet 5.5. In contrast, Google has not structurally upgraded its flagship reasoning models since the Opus 4.6 and Sonnet 4.6 era. For most developers, Gemini has had virtually no compelling coding use cases since Opus 4.8 and Opus 5. As a result, the $200 / mo Ultra tier cannot be justified on frontier intelligence. Instead, its real value lies in utility infrastructure: a 30 TB storage pool, $100 / mo in Google Cloud credits, 600-source NotebookLM workspaces, 300 daily Jules VM tasks, and unconstrained Flash throughput. Here is the practical guide to navigating quotas, family sharing boundaries, and extracting maximum ROI from Google’s multi-rail platform.
AI · September 18, 2026Bend 2 and Proof-Blocked AI: What a Law File Actually Buys
Part three of the machine-checked proofs series. Bend 2 moves the kernel-checked loop from mathematics into ordinary code: you declare laws, the AI must prove its changes keep them. We read the site, the paper and the checker, and ran it. The checker held up against every cheat we tried. The laws are still yours to get right, and on launch day nothing but convention guards the file they live in.
AI · September 18, 2026Anthropic Claude Subscriptions: Pro, Max, Team, Claude Code, and Dual-Gate Quotas
Engineering teams leveraging Anthropic face a dual-gate quota architecture: a 5-hour rolling session window paired with a fixed weekly usage ceiling. While Claude Pro provides baseline personal access, Claude Max delivers 5x or 20x Pro capacity, and Team plans scale from 1.25x (Standard) to 6.25x (Premium). Subscriptions also integrate Usage Credits for uninterrupted terminal coding in Claude Code, alongside interactive Artifacts. Here is the operational field guide to Anthropic quotas, prompt caching discounts, and subscription economics.
AI · September 9, 2026Navier-Stokes Blowup: What a Machine-Checked Singularity Unlocks for Real Engineering
Yesterday’s Navier-Stokes result is part two of the machine-checked proofs series: OpenAI proved smooth forced flow can blow up in finite time, with the full derivation checked in Lean. The equations don’t give you a better pump tomorrow – but the verified lemmas, the blowup guard pattern, and the agent pipeline behind them map directly onto CFD, tidal and pumped-hydro sizing, and home hydronic loops you can test this week.
AI · September 5, 2026The Practical Value of Machine-Checked Proofs
Fermat’s Last Theorem has zero direct industrial utility. But the pipeline that machine-checked its 13 million lines of Lean code demonstrates an operational verification engine for critical software, cryptographic primitives, and hardware design. Here is where the real economic value lands – and how DIY developers and small teams can leverage it today.
AI · August 22, 2026How I Learned To Stop Worrying and Love The Prompt
Stop asking whether AI writes good code. Ask whether your process catches bad code either way. A field guide to the guardrail stack: self-running gates, tests-first contracts, mutation testing, and cold review.
AI · March 18, 2026The Dog Cancer Vaccine That Actually Worked – And How You Can Build One
Paul Conyngham built a personalized mRNA cancer vaccine for his dog Rosie using AI tools and academic collaboration. The tumor shrank 75%. Here’s what actually happened, how personalized neoantigen vaccines work, and what it would take to replicate the pipeline yourself.
AI · March 15, 2026What Secure AI Development Looks Like in Practice
Not theory, not vendor slides. Just the real questions around control, review, supply chain risk, and operational accountability.
AI · February 20, 2026RDP Authentication Forensics and AI-Assisted Patching of GNOME Remote Desktop 49.0
How agentic AI tooling helped investigate a 3-year-old GNOME Remote Desktop bug, and why adversarial verification of AI output caught four incorrect diagnoses.
AI · February 15, 2026From Filing Tickets to Fixing Problems: AI-Assisted Development of Secure OAuth2 Authentication for Apache Guacamole
How AI-assisted development enabled a security professional to build a secure OAuth2 extension for Apache Guacamole, discovering a CSRF vulnerability in the process.