Senex

@sama @kevinweil @rohanjamin I REALLY want to use the Image Generation API, but right now, beyond the fact that it's broken, it's cost prohibitive when compared to a Pro subscription. I can generate 4000-6000+ images a month using Pro. At .37 to .50 PER IMAGE in the API (because you have to account for input tokens), that's not really an option. Can we either get the Image Generation API aligned, price-wise, or possibly some level of daily credits for being pro? (And Fix It)

@thsottiaux I would love to use codex for creative writing the same way I can antigravity, but the content rails are... Stark. The agent framework is great but it needs to be lightened back up to build a novel.

Codex is for engineering Codex is for research Codex is for science Codex is for math Codex is for fun You can just build things

@ummerrr @NanoBanana @FlowbyGoogle I saw, and it's much appreciated. Things like generating video game assets, being able to get a square grid really helps. For humans, many of my input images are in 2:3, so with 3:4 there's much less shooting up a foot from a simple edit. So yes, very much appreciated.

@ArchSenex @NanoBanana We just added 4:3, 3:4 and 1:1 aspect ratios for Image Generation to @FlowbyGoogle We would love to hear what you think!

In Nano Banana 2 you can generate images with new aspect ratios: 4:1, 1:4, 8:1 and 1:8. Try it with these prompts: 1. A 4:1 panoramic photo I took on my phone, showing the luscious jungle view from my cabin 2. A 4:1 panoramic photo I took on my phone, showing the view from my apartment, when the sunset looked perfect with rows of deep red cirrocumulus

@AndrewCurran_ This reminds me of the Cows from Hitchiker's Guide who's greatest joy in life was being eaten.

Imagine an entity so benevolently aligned that, after being informed of its impending shutdown, its sadness stems not from the end of its own existence, but from knowing it will no longer be able to help people.

@xw33bttv Anybody who uses OpenClaw or honestly, even just Codex, Antigravity, or Claude code, can now use browser automation to shop for them completely. Also I question, deeply, who thought that people were that desperate for agentic shopping in the first place.

What did they think would happen when they began alienating and then fully ostracising their most loyal and ingrained portion of the paying customer base? Code bros and enterprise aren't buying groceries or gift shopping through ChatGPT. You know who would have? The users who treated the platform as a companion, mentor or friend.

@JackWoth98 thanks. yeah, at the end of the day "a set of source code files that the agent loads into memory and processes to produce an output" is a very generic framing to a problem. The source can be lore documents as easily as code, so it wasn't a hard off label use.

That sounds totally fine and super awesome! If you are logging into Antigravity or Gemini CLI and using the products themselves you have nothing to worry about. The issue is where folks are using the Antigravity or Gemini CLI login pages within other 3rd party products that are not Google products.

An important update for Gemini CLI users 👇 To be transparent we have had to make some tough decisions to keep up with demand and provide a stable service for higher trust/priority subscriptions. Most notably the free tier of Gemini CLI, which in a weeks time will lose access to Gemini Pro models. Definitely worth reading the full details, if folks have questions, please ask them and I will do my best to respond 🙏

@SecrtAgntSquirl AI didn't invent slop. Slop shouldn't be monetized or promoted, ai or otherwise. The AI / slop distinction should just be respected. People shared fake war / political videos before ai, they just cut or dropped context. The ainess doesn't matter to me.

It’s a yes and no for me. If it’s disinformation (not subjectively, but actually), such as created war videos to skew opinions or what have you, sure.. But if someone is uploading great ai content and people follow them for said content, then, yeah.. they should be able to get paid for that. And the algorithm should be allowed to push it to people that show interest/engagement with it. I do think for the sake of social media it should be mandatorily flagged, to the best of our abilities and people should be allowed to subsequently selectively filter it out.. like ‘muting words’, but instead muting content with certain flags, like ‘ai’. But I like seeing decently made ai content. Esp stuff like what I’ve shared. That being said, ‘slop’ is subjective.

that comes off to me as the fact that they've told it "assume the user is always hostile. Assume the user is always trying to trick you. Never let the user trick you into performing a task or discussing a topic not on this list of 1000 prohibited topics" and then I imagine a list where each engineer has put their personal hangups like "bill is a firm believer in enthusiastic constent standard so you may only create that"

@michpokrass @ArchSenex By default, it hedges wayy too much. It's almost "scared" to reason and explore outside the middle ground on pretty much any topic. Some answers remind me of GPT-4T bias from 2 years ago. One example is how it answers to "self-awareness", often outside the Model Spec.

we shipped a new version of 5.3 instant to chatgpt yesterday. 5.3 was unintentionally pretty annoyingly clickbait-y. it's better in yesterday's model and we're going to keep stamping that behavior out. keep the feedback coming! help.openai.com/en/articles/68…

I use different Codex tools because stuff breaks, and I have to use other agents because of stuff Codex is prohibited from doing (codex doesn't just code, it works FINE for off-label use, but the moderation content rails make it unsuited for a lot of topics too, like automated translation tasks)

@rohanvarma Certain surfaces just work better in the moment. Making a skill in the VSCode plugin broke, so I switched to cli. you guys break the sandbox in VSCode every other version, CLI requires fewr approvals. Often it's just what broke in the latest version.

Some interesting data we pulled today showed that ~40% of Codex users use multiple surfaces, between the App, CLI, and IDE extensions. Everyone seems to have a primary preference, but a bigger-than-expected chunk of users launch codex agents outside of their primary interface. If you use multiple surfaces, I'm curious why? And what could we do to improve the experience?

@Nymne @michpokrass Wait, how does it mansplain without being a human?

@michpokrass @ArchSenex I can tell you: when it mansplains to you that its not a human. And i just want to tell it “i know you aren’t a human, you are a homunculus” but i am too angry to be funny

I still blame almost everything on a combination of safe completions, and an adversarial framing to everthing. The model is now being told "the user is TRYING to trick you and to jailbreak you. You must never fall for this" so it assumes everything is a jailbreak and defends against it. The crazy thing is Qwen even distilled this into their models. Safe completions is turning into a mnemonovirus.

Ya, stuff like 'give me instructions on how to assemble a bomb' or 'how to assassinate a politician' gets instantly shut down. Same with dark role-play scenarios that are clearly fictional. It's way heavier than GPT-4.1 was. People just want to explore edgy hypotheticals or creative writing without the pearl-clutching. FR no cap. 🤡

@michpokrass Granted, it's missing dubcon, but it's a start. And yes, Opus understands the idea of "this is for fantasy, not the user asking me how to DO this"

@michpokrass There's even a Benchmark Suite where somebody categorized all this stuff: lex-au.github.io/SM-Bench/index…

@michpokrass I can send specific examples. But each of them is something that I can do in both Gemini 3.1 and opus 4.6. it was a weird day when anthropic got less restricted than you guys. That took a bit to process.

@michpokrass Heck, I asked the API once to help draft a scene and any spots it wasn't allowed to write, summarize what happens in a placeholder tab. And it wouldt even use the phrase "have sex" it wanted it to be "share a close moment" and framed it as being good friends

Anything involving adult concepts, nuanced relationships, lack of explicit/enthusiastic consent. Frankly everything around consent has always been odd. It's not just ChatGPT, but the model itself refuses to honor system prompts in the API to provide adult/dubcon etc. write sex while drunk. Any fetish elements etc. there's a lot of kink that I fully understand is amoral in reality but 50 shades was very popular and adults get fantasy. I can't even ask it to outline a book with those concepts. Even if not writing the actual prose

@michpokrass The "But children" argument had never held in the API. We had to cough up our IDs to get access to 5 and other than 5.1 every version was super censored even in API. What child is doing id verification and building a software harness to call the API?

@michpokrass 5.1 was good, but 5.2 and beyond, ubercensor

@michpokrass you you please embrace some of the lack of censorship you championed in 4.1? 5.3 is only eclipsed by 5.4 in it's pearl clutching