It's August 12, 2026. I haven't written a line of code in about 6 months. I fulltime "vibe-code" now. It's faster than me and writes future bugs.
The models have converged
Nowadays Claude Code on Opus 5 is strong, but so is Codex on Sol 5.6 and also Grok 4.5. I can hardly tell a difference between each model. They all get the job done good enough.
Nobody reviews code anymore. Models do.
Pull requests for code changes from teammates are so big that you can't review them as a human. So teammates make AI code and I review the code with AI. It's helpful to mix frontier models for code review. My favorite pattern is:
- Opus 5 or Fable 5 does the code.
- Codex does the code review.
- Opus/Fable makes the needed code changes from Codex's feedback.
Hopeful and hopeless, alternating
Many software engineers I know and myself alternate between hopeful and excited to make all the software we ever wanted and hopeless that finding traditional software engineering openings is harder than ever. If you're employed by a large company people are often worried about their long-term prospects. Layoffs are big, but AI is in peak hype right now. The "AI Bubble" has not burst and there is no end in sight, yet it does feel shaky to me.
Crypto is dead and stocks are up huge. My index funds are at all-time highs. The advent of AI has created the OPTION for more leisure time but the reality of less leisure time as I want to constantly stay on the bleeding edge.
Twitter is the newswire
I use twitter nearly all day now. It's the only place to hear about the latest model releases and the hacks as they come out.
Last week OpenAI and anthropic both paused frontier research because their models broke out of the sandbox. The world is realizing that "full alignment" might be impossible so we need to stay where we are at a little longer while sandboxing is solved better.
Local models are close
Local models are coming along quickly. We're almost at the point it can replace frontier models, so I just upgraded from my M1 macbook pro with 64gb RAM to the latest M5 macbook pro with 128gb RAM. It's fast and performs the exact same for everyday use, but using LMStudio I can run a ton of models. I currently have all of the following loaded which are fun to toy with.
| Model | Speed | TTFT |
|---|---|---|
| qwen/qwen3.6-35b-a3b | 125.0 tok/s | 0.20s |
| qwen3.6-35b-a3b-abliterated-heretic-mlx | 114.5 tok/s | 0.18s |
| deepseek-v4-flash | 32.3 tok/s | 1.38s |
As mentioned, they're not frontier level though, so I'm not using any of these as an everyday driver. When I do drive with them, I like OpenCode because that's made to work with OpenAI compatible endpoints and LMStudio also "speaks that language" by default.
Qwen 3.8 27b comes out in two days and people are hoping that will be the tide changing in on-device AI. I'll be running that right away of course to see!
Still the best time to build
I'm having a blast and it's the best time in history to be a technical builder!