A fresh interview just dropped, and it quietly rewired how I think about coding agents. On the surface it’s about faster models and slicker tools. But the real story is step two, where the AI starts improving itself, and that’s the part that made me sit up. Tibo, an engineering leader at OpenAI, sat down with the interviewer and walked through it all, from his early days at Google DeepMind to the future of personal AI. I’ve watched a lot of these conversations, and this one felt different because the expert kept pulling back the curtain on things most people haven’t clocked yet.
What’s new
The headline number: Codex just hit 20 million users, and the creator showed a growth curve that goes flat, then straight vertical. That kind of jump doesn’t happen by accident.
A few things the OpenAI leader pointed to as the drivers:
- Ultra fast mode. Token speeds roughly 10 to 14 times faster than the current “fast” setting. When there’s lots of code to generate, you feel the full speedup. When it’s heavy on tool calls, the network becomes the bottleneck and you get more like a 3x to 4x boost.
- ChatGPT and Codex merging. The author explained they’re becoming one product because the future models want it that way. Same harness, same multimodal, voice-first engine underneath, whether you’re coding or not.
- A new voice model that’s natural enough that he now dictates his morning task list out loud instead of typing prompts, and the agent just goes and does it with access to his tools.
The twist
Here’s the part I keep chewing on. The expert described “recursive self-improvement” already showing up, just not where most people look for it. Everyone pictures models building smarter models. What OpenAI is actually seeing is models rebuilding the infrastructure that serves those same models: inference stacks, CUDA kernels, the whole pipeline.
The proof? They dropped the price of a model called Luna by 80%. Tibo said that wasn’t mostly clever pricing. Their most advanced models figured out how to re-engineer the serving stack for huge efficiency gains, and they passed the savings straight to users. He also noted speeds are about 60% faster than three months ago. That’s the flywheel: better models make better tools, which make better models.
One more surprise. OpenAI actually paused the absolute frontier of reinforcement learning training. The person who shared this framed it as a safety hardening step, letting teams fully understand the system before restarting with confidence. A company racing this hard choosing to pause tells you something.
The mini-workflow shift
The most practical takeaway is how solo developer workflows change as speed climbs. 🛠️
- Today: You kick off 10 to 15 agents in parallel, then wait 30 to 45 minutes per task. The interviewer admitted that context switching is brutal cognitive overhead.
- With ultra fast: You drop to maybe 3 or 4 agents, or even one, because the model moves at your speed or faster. You stay in the flow instead of juggling.
- The payoff: You prototype a site or a video game and watch the code appear in real time. You steer with voice or text, see the mockup instantly, and iterate like a conversation instead of a batch job.
- The end state: One interface for everyone. Tibo said your mom will use the same thing you do, a “personal AGI” that tailors itself to each person instead of making you pick a “coder” or “non-coder” mode.
Pro tips from the expert
- Match speed to your brain. 🧠 Ultra fast helps most if you like staying focused on one thing. If you thrive on context switching, you feel less of the benefit. Know which one you are before you change your setup.
- Watch where the overhead lives. Heavy tool-calling tasks won’t hit the full 14x. Generation-heavy work will. Plan your prompts around that.
- Efficiency is the real access story. The author’s argument for nervous newcomers: cheap intelligence spreads everywhere. Luna sat at the frontier six months ago and now runs for almost nothing. Whatever feels premium today gets cheap fast.
- Try it on something personal first. He uses ChatGPT for health and finance help so he walks into his doctor’s office more informed. Low stakes, real payoff, easy on-ramp.
Why I think this matters
What struck me most was the mindset. The creator said he barely looks at the competition, even Anthropic. He focuses on what OpenAI can do uniquely well and how fast they can get there. That’s also why he can hit those “reset” buttons that refill usage limits whenever something breaks. There’s no marketing or finance committee behind it. It’s a goodwill move, and I think it’s quietly fueling that vertical growth curve. Caring for users, done for real, compounds.
If you build with AI agents, the deeper dive into ultra fast speeds, the merge, and the self-improvement loop is worth your time. Check out the full interview for the details and the moment where he reveals there’s an actual physical reset button. 🎬