I called model parity back in April. What I missed was everything around the model: OpenAI put Codex and ChatGPT in one app, with an import command pointed straight at Claude Code's users, and bought nearly 7 gigawatts of capacity to run it. It's still the company behind on revenue and valuation.

A year ago I built multi-agent orchestration by hand. Now it's routine. Same me, same Claude subscription: 12 months for the Python version, 2 for the Rust one that does more. The capability moved out of my code and into the tool. New on HyperDev 👇

Every spec I've seen dies the same way: written before the code, reviewed once, approved, then ignored by everything that touches the code afterward. What changes when the spec stops being a separate document and becomes part of the build?

The people who build the most capable coding agents keep telling you your harness will get absorbed into the next model. They're right. Cat Wu at Anthropic put it bluntly: "the models will eat your harness for breakfast." So why am I telling engineers to learn harness engineering anyway?

Fable 5 returned to Claude Code the morning after Sonnet 5 shipped. Ask it for exploit code and watch it hand the job to Opus 4.8 mid-session — the safety classifier reroutes offensive-cyber work, and devs say it over-fires on defensive tasks too. ↓

Gary Gulman coined "depresh" because "depression" was too heavy to say. There's a coding version now, and most developers don't have a name for it. A 15-year dev shipped an AI-built pipeline and felt grief. "Not for a person. For an identity."

CrowdStrike found DeepSeek-R1 wrote insecure code ~50% more often on politically sensitive prompts, and the bias survived self-hosting. The DeepSeek risk and the SpaceX-owns-Cursor risk are the same exposure. Part 3 of The Year of the Fire Horse.

Llama 4 Maverick scores 15.6% on Aider Polyglot. Chinese open models run ~5x higher. LeCun says the Llama 4 numbers "were fudged a little bit." The Western open-weight lead crossed an ocean. Part 2 of The Year of the Fire Horse.

SpaceX bought Cursor for $60B. Cursor's default model is already built on a Chinese base (Kimi K2.5), fine-tuned by a US company. The "no Chinese models in our stack" rule was broken before anyone noticed. Part 1 of The Year of the Fire Horse.

I let a Claude Code session run a 14.5-hour engineering job last week, and I stopped intervening about an hour in. The orchestration layer — the PM coordinator, running on Opus — handled the work I'd expect from a senior IC who'd been on the team a year. Not faster autocomplete. Project management.

In 2025, OpenAI spent about $1.69 for every dollar it took in. That's a $9B loss on $13B in revenue. Fine for a startup finding its footing — except the company's own leaked projections show the loss widening to ~$74B as revenue approaches $100B in 2028. The loss grows faster than the top line.

Most of the best ideas in software engineering aren't new. TDD, design by contract, ADRs, mutation testing — ask any room of senior engineers if these are good ideas and you'll get nods. Ask who actually does them under deadline pressure and the hands stay down.

For most of my career, coding and leading were genuinely incompatible. Not a failure of discipline — a constraint of attention. Serious technical work requires long, unbroken focus. Management is interruptive by design. Two modes that couldn't run at once. Something shifted.

Uber's CTO burned $1,200 in tokens during a two-hour Claude Code demo. The company blew its full 2026 AI budget in four months. Same week, Microsoft cancelled Claude Code for thousands of its engineers. The takes were "AI failed." That's not what happened.

🔋"I wish I had my own CTO Assistant. Like what you have." —my colleague Jake That gap—between knowing AI tools exist and actually having one that knows *your* codebase—is what separates practitioners from observers. In Star Wars, Padawans don't receive lightsabers. They build them. #AIEngineering

🧵 I rebuilt my Python AI tools in Rust and got 10x+ performance gains. But the real revelation wasn't about speed—it was about safety in an era where AI writes code. Here's why Rust might be the AI language of the future 🦀 #RustLang #TechLeadership #SoftwareArchitecture #PerformanceEngineering

The "RAG is dead, just stuff the context" crowd is solving the wrong problem. A million-token window doesn't fix retrieval — it just makes bad retrieval more expensive. New post on why search matters more than ever.

I avoided AWS for years because of its overwhelming complexity. Then Claude Code/MPM changed everything. Now I'm running sophisticated GPU instances, multi-AZ deployments, and comprehensive monitoring. The lesson: agentic approaches transform ops workflows just as profoundly as they do coding.

Andrej Karpathy just posted his "LLM Wiki" — using Claude to build and maintain a personal knowledge base instead of writing code with it. His observation: "I think there is room here for an incredible new product." Mine: the product gap is real, but it's not the interesting part. New piece. 👇

Claude Code really was dumber this month. Anthropic just published a postmortem revealing three bugs that compounded into one massive degradation: silent reasoning downgrades, a caching bug that made it forget its own decisions, and verbosity limits that tanked coding quality by 3%. All fixed now.

Opus 4.7 dropped last week. I ran it head-to-head against 4.6 on an identical coding task. Both models passed all 10 tests. Opus 4.7 cost 3.6× more. Same task. Same result. $0.38 vs $1.38. And when I tried to cut costs by removing tool access, 4.7 failed 9 of 10 tests. Opus 4.6 passed perfectly.

I tested 8 AI coding tools. Four used the identical Claude Sonnet 4.6 model. Their quality scores ranged from 3.93 to 4.75—a 0.82-point spread on the same foundation model. The difference wasn't the engine. It was the car. Full breakdown: #AgenticCoding #SoftwareDevelopment #TechnologyStrategy

I didn't expect to meet Mila Jovovich through a GitHub issue. But there I was, auditing her AI memory framework that went viral with 23K+ stars in 48 hours. When I found critical bugs in her benchmark methodology, she responded directly — debugging complex algorithms at 9 AM on a Thursday.

Stripe engineers send Slack messages that automatically become production code. Not suggestions—actual production code merged to main, supporting over $1 trillion in payment processing. Their system generates 1,300 pull requests per week with zero human-written code. My analysis:

The Claude Code leak revealed ugly internals: 5,594-line files, regex frustration detection, 250K wasted API calls/day. But it also handles millions of conversations and delivers $2.5B ARR. Question: Do AI-integrated systems need different quality standards than traditional code? My analysis 👇

After the White House complained about an unflattering photo of Karoline Leavitt, it was removed from AFP and Getty’s systems. This is the photo in question. Please do not repost it. 👇

🧵 I counted my terminal sessions: 10 tmux tabs open right now. Only 2 are generating new code. The other 8? Responding to existing systems—bug fixes, code reviews, CI failures. This 8:2 ratio reveals something important about the future of development...

Three Slack connectors. Two weeks. $1-2.5K in AI tokens. They replaced what would have been $150-300K+ in traditional development. People don't want to visit your tool. They want your data where they already work. The connector era is here. "Is This The Era of the Connector?"

One year ago: Technology executive who "hadn't coded seriously in 20 years" Today: CTO using AI tools in production 4,919 commits • 69.7B tokens • 168 articles documenting the transformation in real-time The infrastructure beats features. Always. Read the full journey: 👇

Must share

George Conway ⚖️🇺🇸@gtconway.bsky.social · 5mo ago