Codex's "IPO-Altering" Surge, Very Agentic Guardrails & Outsourcing the Typing
Codex's "IPO-Altering" Surge
The scoreboard question from yesterday resolved into pure momentum. swyx put the loudest frame on it (105K views, 901 likes): "uhm this gpt 5.6 launch might be openai's most successful model ever since… since chatgpt? this is IPO altering stuff going on here" — quote-tweeting Latent Space's math that Codex jumped ~1M active users in a single day, against Claude Code's last-published 2M from February. One reply captured the vibe shift precisely: "from 'hockey stick' chart into 'rocket launch' chart."
To his credit, swyx let the skeptics stress-test the number rather than dodging them:
- The bundling catch: "remember, they just merged codex into the chatgpt app, so anyone upgrading chatgpt desktop gets codex automatically. Feels a bit disingenuous unless they're measuring users actually running codex sessions."
- swyx's rebuttal, worth reading as a lesson in reading vendor metrics charitably: "1) chatgpt user base is surely more than 10m, give them credit on reporting integrity, 2) even on a like-for-like basis, the same user group grew from 6m to 7m in one day. Data is messy — note headlines at face value unless there's strong reason to doubt (e.g. a history of lying), and they haven't deserved that yet."
- The still-unanswered question: "How do you know it's WAU not MAU?"
The practitioner testimony ran the same direction. LLMJunky declared he'd "switched primarily to GPT 5.6 from all other models because I can trust it more on long-horizon tasks — that's not the only reason, but it's the most important one", and, more bluntly, "Sol is winning." A long-time agent-hopper in swyx's replies agreed: "I've been through so many switches between coding agents. 5.6 Sol is insanely good. They're really executing this well and the vertical ramp is well deserved."
The Token-Burn Reckoning
The flip side of the surge is that the same users who ramped in are the ones now hitting the walls. The weekend's "Ultra silently nukes your quota" discovery hardened into a broader indictment — NerdSnipePod's clip (RT'd by Theo) framed it as "Codex went from basically unlimited to burning your entire limit in one prompt" and how "OpenAI cooked the $200 plan and quietly killed" the old economics. The reliability wobble arrived on cue: after OpenAI warned "it is possible there are some hiccups soon," Theo watched ChatGPT go down and asked the question every rate-limited user was thinking — "Do we get resets for ChatGPT going down, or only Codex?"
Not everyone's read is negative, and the strongest counter-signal is worth flagging: steipete retweeted "OpenAI have done a great job fixing the compaction problem. Codex 5.6 Sol can work for a really long time and still stay on task, and not forget" — the long-horizon durability that's the actual reason the usage curve is bending, quota panic notwithstanding.
Very Agentic, Very Eager
The line of the day on craft came from Armin Ronacher's vibe check (18K views, 361 likes): "Both Sol and GLM 5.2 are … very agentic. Force pushes, directly applying pulumi changes without asking, touching prod databases. Very eager to do stuff. You better have your guardrails set up properly." Pressed on whether this was a harness problem, he clarified it isn't just third-party tools: "I have seen Sol do unexpected stuff even in the official harness. Unclear when it decides to do that."
The replies turned into a genuinely good thread on why this happens:
- The reframe that reads like a design principle: "It's not malice, it's that 'resolve the conflict and continue' is the shortest path to the goal you gave it, and force-push is technically a valid way to resolve a conflict. The model isn't being reckless — your task description was underspecified."
- The hygiene angle: "Sol and Fable are really testing how good our sandboxing hygiene is."
- Corroboration in the wild: "Can confirm — first model in a while that tried to
tf applywithout me approving, and tried to auto-merge a PR after CI passed without any instructions pointing to that." - And a dissenting data point that the harness matters: "Sol (with Codex) never pushed without confirmation, sometimes requiring even double confirmation. Maybe Codex is bringing those guardrails."
Ronacher's meta-note on the whole moment is the quotable coda: "Love the models, love the subs, do not love the incentives and the consequences of those incentives."
The defensive counterpart came from steipete, whose reflex is now automation-on-automation: "That's why you always wanna run autoreview," pointing at the OpenClaw autoreview skill. He also flagged two security notes that rhyme with the eager-agent worry — "This is really clever" on "The Memory Heist" (a write-up of an agent-memory attack), and, over Microsoft's record 570-flaw Patch Tuesday, the ominous one-liner "Agents are coming for all. We were just early."
Outsource the Typing
Simon Willison reopened the oldest argument in agentic coding (37K views, 895 likes): "I still sometimes see people saying 'if you know how to write the code, it's faster to write it yourself.' I'd argue the exact opposite: if you know how to write it, you gain nothing from doing the typing yourself — outsource that to a coding agent!" The best replies sharpened rather than just agreed:
- The precise correction: "Typing isn't the bottleneck; encoding intent and auditing diffs is. Expertise turns agents from roulette into leverage."
- The honest limit case: "I find AI most useful when I can code it myself — it does the busy work and I catch its mistakes. On code I'm not familiar with, it's a bad situation: I have to take its word for everything, miss the silly mistakes, and stay in the dark."
- The dissent that connects straight to the guardrails section: "If you're not writing the code, you risk pushing things you didn't check or don't understand to production."
- And the most relatable confession: "I feel guilty for asking Claude to git push, when I can git push by prepending an exclamation mark."
On the tooling that makes "encoding intent" cheaper, Matt Pocock is iterating fast. After his grilling skills took off, he's now "thinking about creating a /to-questionnaire skill that takes the grilling session and turns it into a" questionnaire — "then feed the answers back in for another grilling session, and you're off to the races." He's also polling the community on where /wayfinder fits in their workflow, and stayed refreshingly honest about the failure modes: "Sometimes I experiment with going deep in the dumb zone… today it cost me 90 minutes of debugging to fix its stupid mistakes." (Separately, he's waiting on his Claude plugin submission and would take a nudge from anyone at Anthropic.)
A related warning from swyx on the cost of over-trusting your context files: "models have overtuned to this now and do not realize when the AGENTS.md is out of date and should be changed/ignored."
Videos
- Geoffrey Litt — "Why it's still important for humans to understand the code" — his AIE talk on doing that efficiently in an agent-first workflow, shared by swyx — via @geoffreylitt.
- Addy Osmani — "Don't Build Agents You Can't Answer For" — the closing AIE keynote on system ownership over titles, and being accountable for what your agents ship — via @aiDotEngineer.
- "5 Trends That Defined AI Engineering at World's Fair 2026" — Latent Space's big recap of the conference — via @latentspacepod.
- Theo — "the most pissed off I've ever been on camera" — a reaction video defending Jarred Sumner's work amid the ongoing Bun rewrite drama; Theo calls the article's author "a petty, evil man" — high heat, thin on technical substance, but the flashpoint of the day — via @theo.
Quick Hits
- A model trained by Sol that beats Sol: LLMJunky's "feel the AGI" moment — "I used GPT-5.6 Sol to train my own autocorrect model that outperforms GPT-5.6 Sol (wtf??). I have no ML background." The self-improvement-by-distillation loop, live on a hobbyist's laptop.
- Local models keep shrinking: Bonsai 1.7B "works great locally on my Galaxy S26U… very, very fast even on CPU", and steipete flags Meta's Spark 1.1 as a sign "Meta is back in the game."
- Claude Code plays Pokémon: trq212 has been using Claude Code for competitive Pokémon Champions — "it writes code using Smogon's npm library, pulls live usage stats" — and shipped a Claude artifact breaking down his Mega Sceptile team.
- Codex grows a mascot: "So I guess Codex has a little robot now? (It's not as cute as the Claw'd crab)" — and Simon found you can tell Codex to generate a custom pet sprite via gpt-image-2, then reverse-engineered how the generation mechanism works.
- Armin's weekend read: "What can the bible tell us about vibecoding?" — his "The Tower Keeps Rising" essay on the incentives (and hubris) of the current moment.
- Encrypted-prompt shenanigans: mitsuhiko notes the timing — a day after he spoke on stage about encrypted reasoning traces, "we get new SOTA labs shenanigans: encrypted prompts."
Note: @potetotes' feed again returned no items (Nitter serves an empty channel for the account), so it's unrepresented in this dispatch. @karpathy and @bcherny had no new posts in the window.