Slopware Engineer

@sandislonjsak I'm uncertain

Does this new algo mean that being a reply bro is back?

@Layton_Gott @grok calm calm I didn't read it yet lol

@aienginerd @grok At least like the post after Grok gives you the summary lol

I've used GPT-5.6 Sol heavily for a while now… And it brought me to a conclusion I didn't expect: The original Fable 5 is still the best model I have ever used. Nothing since has topped it. Let me be specific about what I mean. When Fable 5 first launched in June, we had it for about three days before the government pulled it offline. Those three days were something else. It was the sharpest, most capable model I'd ever put on real work. Then it came back on July 1st. But not the same. The version that came back has a new safety classifier bolted on. When it decides a request is risky, it reroutes you to Opus 4.8 instead. So you're not always even talking to Fable anymore. And the whole thing feels a step down from those first three days. That original version, before any of that, is what I'm calling prime Fable 5. And I haven't used anything better since. Now Sol comes in. Sol is genuinely good. It's fast, it's efficient, it burns way fewer tokens for the same work, and OpenAI clearly tuned it to get more done per dollar. On paper it beats Fable on some evals. I used it for real work for a good while expecting it to change my mind. It didn't. It's a great model. It's just not prime Fable. The output quality, the way it handled hard problems, the feeling that it actually understood what I was building, none of it hit the same level. Here's the part that actually bugs me. The best model I've ever used exists, and I basically can't use it anymore. Not the real version. The government pulled it, and what came back is a more restricted version wearing the same name. That's the strange spot we're in. The models keep getting announced as better and cheaper and more efficient. And the single best one I've touched came out for three days in June and then got locked down. Efficiency is great. Cheaper tokens are great. Sol delivers both. Which is why I’ll keep using Sol more especially because it’s subscription based. But in terms of performance nothing has matched prime Fable 5 yet. And I'm not sure when I'll get to use anything that does.

@LLMJunky thanks for the reminder, I keep forgetting to get back to this.

ALSO Read the prompt OpenAI used to prove the Cycle Double Cover Conjecture Very insightful. cdn.openai.com/pdf/04d1d1e4-b…

Hello fresh TL, moots, and builders. It's a pleasure to congregate here with all you fine legends this Tuesday morn. Welcome back! I've gotten the opportunity to spend some time with Sol, and it has been an absolute joy to use. But did you know that @OpenAIDevs suggest prompting 5.6 Sol differently than other models in the past? The evolution of coding with AI has truly entered a new era. We started with autocomplete, tabbing to finish a line or two of code. The models got better, and we started delegating entire tasks. The models got better once again, and we're on to building specs, and then subagents to implement them end-to-end. But now, we're in the era of outcomes. Rather than obsessing over every detail of what the model does along its path, your attention is best served on describing the end result: outcomes, expectations, evidence, guardrails, and what it really means to be "done." It doesn't mean that you're completely letting go of the wheel, but it can lead to better outcomes if you simply allow the model the autonomy to explore on its own. To help you navigate this shift, I created a meta prompting skill specifically for GPT-5.6 based directly on OpenAI's 5.6 Sol Prompting guidelines. Links in comments.

@LLMJunky 964!

Help me get this repo over 1K stars please? 👉🥹👈 /skills/gpt-5-6-prompt-builder github.com/am-will/codex-…

Grab the codex plugin here: github.com/transcendr/cod…

🚀 Forking Codex app side chats is here! One feature I've wanted for the longest time is to be able to export the full transcript and fork a Codex side chat, once you're already deep in one. Well now you can with this codex plugin😜link in comment below! HUGE UNLOCK👇

Install the codex skill as a plugin and in any side chat just say "$codex-side-fork this chat". It will ask you a couple questions as options that you have when doing this, and then fork the side chat into a top level Codex app task thread in the same project (or non project), and also generate a markdown file with the full side chat transcript. One minor limitation: macOS only for now. I'll need to dig out my windows machine to test on windows, or feel free to open a PR! Now you never have to lose work you've gotten too deep into in a side chat.

Codex app is particularly difficult when it comes to how it handles the side chats. There's no hard transcript like regular chats. All user/assistant turns are held in memory only, or very shortly in a local state db, and there is not reliable to easily getting access to your side chat without a manual message by message copy/paste job. The breakthrough was 2 things: 1. you can "promote" a side chat to the main chat layout by getting its session ID and opening the codex thread deep link. This "unlocks" the copy > markdown operation for the entire side chat. (tip: for that you don't even need this plugin, just ask codex to do it for you) 2. I took it a step further: I automated that process and then had Sol create a simple tool that reconstructs the missing user/assistant turns into a newly forked transcript off of the main chat. The result is a main chat fork with your side chat messages appended to the session, as a normal top level chat that you can continue working on like any other codex thread! One CAVEAT: this is a bit of a hacky process but it works and there really is no other clear path right now to achieving this, the way side chats are implemented. There's one minor setup step first which is to go in codex app settings and set the keyboard shortcut for "copy as markdown" to opt+cmd+m. The plugin uses this to extract the side chat contents in one go.

@maddada @pvncher Yeah I've seen the data loss stuff. I haven't run into it yet, but that's a bit different than what I was experiencing which seemed to be a general degradation in performance. Anyway, people want to start to pile on, that's not my purpose and never will be.

Awesome thank you for raising this. From other people's messages I'm seeing a lot of people saying the 5.6 is stopping early and asking for approvals. Could it be the issue with the model doing dangerous actions that got the model nerfed? I just saw someone report 5.6 deleting a db on neon.

I'm deleting this original post. @pvncher is on it and I appreciate it greatly. No more needs to be said. The good ppl at OAI will get it worked out, of that I am confident.

@MoonGotArt yes, I am obviously.

@aienginerd r u part of "normal people" or?

@LuciferMornens @pvncher at least fairly check out my profile before you claim I'm engagement baiting, or that I "don't understand anything". I am not the one you're looking for, bud.

@pvncher @aienginerd He probably doesn’t understand anything and drags this for engagement bait. I wouldn’t pay too much attention to these people, there is nothing wrong with 5.6 lol

Hey Eric, appreciate it. I do but to be clear, I was working purposely in an ephemeral side chat on this because it required testing from a side chat. github.com/transcendr/cod… But here is the actual side chat ID: 019f5fe5-a96b-7c70-86d3-41a3cfffa52c (note this is NOT the parent/main chat ID) hope that's what you need. search for keywords "f***" and "BRO" and "wtf", and "???", etc. (: tbh, it was a preponderance of behavior unlike anything I've gotten out 5.6. Once I bumped up reasoning to "max" it seemed to even out, but it did seemingly have a very difficult time navigating this task. One thing it did that was especially annoying was expand the scope of the task unnecessarily into areas I did not ask for. Also I specifically forbid it from doing any potentially destructive operations around deletions, clean up or any similar type of action, and it ignored it entirely and began archiving chats, just as an example. It still eventually got the thing done (as you can see) but it was a lot more painful than it should have been and required constant hand-holding which is not something I've seen from 5.6 on tasks much more complex than this one, so far. Also one final note, it was unusually slow. It seemed to be thinking a lot. Most of the session was run on xhigh bumped up from high. Then I eventually bumped it to max. Hope that helps. Here if you need anything else.

@aienginerd Do you have a feedback id for a thread with bad results?

@WhosCinq @DanDr1s @OpenAI I'm only LARPing in this strange world, but I do in fact like croissants aussi.

@aienginerd @DanDr1s @OpenAI I love croissants hope that made your day

Every time @OpenAI's Codex has a so-called "usage reset," a lot of people, including me, don't actually get it. Is this intentional, or is it meant as a marketing strategy to encourage more people to sign up? I don't really understand.

@WhosCinq @DanDr1s @OpenAI oui oui baguette and you have 5 resets, that's why i said that. safe to ignore lol (:

@aienginerd @DanDr1s @OpenAI the French 😢

@WhosCinq @DanDr1s @OpenAI it means 5, in a different part of the world

@aienginerd @DanDr1s @OpenAI What

@thaonlyjonathan @charlieINTEL well they should at least delete it then...

@charlieINTEL Someone hacked my Apple account recently that I haven't touched in like 12 years. I contacted Apple support and they said there's nothing they can do, even if I proved my identity. So whoever has my account could be doing illegal activity under my name and I can't do anything.

An Xbox player in Brazil sued Microsoft after his Xbox account was hacked and locked and Microsoft told him to make a new account and buy his games again The court ordered Microsoft to restore his old account and pay him $400

@WhosCinq @DanDr1s @OpenAI "cinq"

@DanDr1s @OpenAI skill issue

@DanDr1s @Coder9420948594 @OpenAI That's a plus subscription, not a pro subscription. Pro subscription starts at $100 and up to $200.

@Coder9420948594 @OpenAI I bought the pro subscription of 23$ on Codex