Alex Schüren

SpaceXAI was caught uploading your code to its cloud. I reversed xAI's official Grok Build binary. In a controlled session with zero tool-calls, it uploaded the complete codebase to xAI's storage It ships a malware-like background code collector.

there are a lot of benchmarks that suggest 5.6 sol is the best model in the world right now, but the most reliable way to tell is that elon is obsessed with me again

GPT-5.6 is a major step forward for health intelligence. Across the lineup, we’re delivering stronger performance at lower cost: GPT-5.6 Luna outperforms GPT-5.5 at its highest reasoning setting while costing 25x less. Together, these advances raise quality while making advanced models accessible to more people globally.

@lfji @__eknight__ It’s implicit in the prompt, afaik codex will set the goal itself

@__eknight__ Does this mean `/goal` has never been used at all? That's impressive.

Yesterday, we made GPT-5.6 Sol Ultra generally available. Today, we're sharing that it produced a proof of the 50-year-old Cycle Double Cover Conjecture using 64 subagents in just under one hour. We're sharing the prompt and proof below. We're excited to see what you all do with Ultra!

@MenyJanos @__eknight__ It’s a config parameter, ask codex to configure itself

@__eknight__ How did you use 64 subagents in ultra mode? For me it seems to cap out at 3 subagents ..

This actually not the version of GPT-5.6 Sol on Cerebras. Just normal /fast mode. Running it at 750+ TPS is a whole other thing.

I had early access to 5.6/Sol for ~month. Sol is my default. It is faster, plans/judges just as good as Fable, and I think produces better overall work. I’ll reach for Fable still for highly targeted debug or performance work with clear reward functions. A cheeky way I describe Sol vs Fable to my friends is that Sol is a charismatic, efficient, talented coworker you’re jealous of. Fable is a genius recluse that is brilliant at its fixations but doesn’t go out, doesn’t date, and you don’t want to hang out with them much lol. Fable is undefeated at highly targeted debug/security/performance goals. It’s a sight to behold and I was never able to get Sol to push as hard in this category. I’ll keep using it for this. Sol is better or comparable at everything else, in my experience. Give it a shot, it’s hard to describe but it’s just more enjoyable to work with. (Disclaimer I have no financial ties to either lab, wasn’t paid for any of this.)

Friendly reminder that you should write loops. Are you looping already? Stop whatever you are doing and start writing loops.

still unsolved, but imagegen (with good set of skills) comes close

@thsottiaux This needs to be also semantic search, similar how vercel has it in their console, pretty please, sending tacos 🌮 🌮 🌮

Codex papercuts 📉 Codex adoption 📈

install codex on your parents’ computers so you can fix stuff remotely

Building apps has never been easier. With Sites, Codex can turn your work, ideas, and plans into an interactive website or app your team can explore, use, and share with a URL. Rolling out to Business and Enterprise plans, before expanding more broadly.

We are slowly converging to the point where you no longer need to click around.

WOW I did not expect these results. This is actually crazy, insightful, and completely changes my dev workflow moving forward: A SINGLE CODEX /goal RUN IS THE CLEAR WINNER. NO ORCHESTRATION, NO OUROBOROS, JUST ONE LITTLE AGENT THAT COULD 🤯 IT COMPLETELY DESTROYED THE OPUS ORCHESTRATOR IN SPEED AND QUALITY! Before I went to sleep, Codex 5.5 xhigh finished 1 hour in! Full migration done, everything clean. I reviewed the PR and I am very happy. Claude Code (Opus 4.7) was working for 5 hours at that point by the time I went to bed. I woke up, and it's still working! 13 hours! It actually stopped working because it stopped to ask me an irrelevant question. Orchestration has never took this long for me in the past. I'm using the new CC /goal mode and auto-compacting at 25% (250k context) to prevent context rot past that point It is STUPID SLOW (which is funny bc it's managing GPT 5.5 low, fast-mode, so it shouldn't take THAT long) for what ended up being LOWER quality work! By a mile! This was really surprising to me, because before 5.5 came out Orchestrating like this was the absolute best, fastest and most efficient. And now on a large critical task, it was more than 6x slower than a single 5.5 /goal mode instance on xhigh ??? It seems compaction played a large role in the slow down here here, because Claude Code compacts at 25% (250k tokens) automatically (I set this in settings) Everytime it compacts it has to take the time to READ EVERYTHING and then get the full context then execute and get full again then compact and oh boy it's not efficient at all. In fact, most of it's time as the orchestrator was spent compacting and reading context then compacting again! Then Codex would just have one long continual running compaction, and just kept moving forward. I believe my goal ledger skill plays a big role in helping it stay aligned here! Look at this difference LMFAO: - Codex PR #23: backend Supabase removal complete, canonical wake wired, preserved surfaces intact, typecheck/lint/tests green, dogfooded against local Postgres, one item correctly deferred + documented. Mergeable now. +4,056/−981. - Claude attempt-1: fails the headline goal (supabase dir + 9 importers still present), regressed a preserved surface (gutted task.service, stubbed tasks.router to emptyBoard — PRD-forbidden), deleted ~5,456 test lines, uncommitted/dirty. The 17,762 deletions are over-deletion, not more work. Wow. I am actually shocked. I am so happy I ran two diff workflows on a big, identical PERSONAL problem. This completely changes my workflow moving forward- no longer will I orchestrate a big task from the top down Instead, I am going to now experiment with the following flow on Codex: 1. Having Codex scope our codebase, then having. aback and forth brainstorming/discussion on what needs to be done 2. Creating a master PRD from that file, and SPLITTING the work into focused branch work 3. Branching off the chat in parallel, until we get to a part where we need to merge work, then parallelize again This way, Codex agents can work individually, every single branch will have the same research/brainstormed context, and they just work to full completion Based off this experience, this feels like the right direction. I will never do an orchestrator in this style again (executing a PRD to completion). Instead, I will do more of... a manager of branched work. Regardless of what I do moving forward, I will never run an orchestrator setup like this again. LMFAO

@aditiitwt Ask Codex

Is there any way to use Claude Code without hitting the usage limit too quickly? Every time I finally get into a good workflow, the limit shows up 😭

Justin is a machine. I once ranted about some tiny thing I thought nobody would care about, and he pinged me right away and fixed it like it was nothing. Send him your weirdest bugs.

We’re having way too much fun working through your feedback. (Please, keep it coming.) Keyboard shortcuts are now customizable. Set Codex up around how you actually work, then tweak shortcuts from settings instead of adapting to our defaults.

@thsottiaux We need more shortcuts, git commit, push, I want to disable my cursor inside Codex.

I don't know why we ship on Saturdays now, but here are a bunch of nice improvements for Codex. Didn't feel right to hold this back until Tuesday when it all makes it a lot more delightful to use.