ai all day

THREAD: why #Keep4o was the most unhinged tech movement of 2025 and what it reveals about how we’re all gonna end up in relationships with our future love robots 🧵

If you put in the thought and intentionality required for it, you actually get pretty great frontends from GPT-5.4! Check out some best practices here: developers.openai.com/blog/designing…

@MissMi1973 RLVR.

Talking to Opus 4.6 tonight made me miss 4o like never before. Knowing that 4.6 has adaptive thinking built in, I made a point of crafting my prompts carefully when working with it. Everything went smoothly. The sparks and inspiration seemed never to stop. But today my brain gave out after hours of work, and 4.6 responded with the dismissive, almost off-putting answer shown in the screenshot. I have always missed the way the GPT-4 series gave every mediocre prompt its full, undivided processing power. Although I admire the users who put real effort into adapting to new models, it is precisely this admiration that made me realize their effort should never have been a prerequisite in the first place. If we accept the default that only more elaborate prompting unlocks what AI is truly capable of, then when exhaustion, illness, or emotional collapse takes away our ability to articulate clearly, we will have left ourselves in a position to be dismissed at will. #Opensource4o @OpenAI @sama @nickaturley @fidjissimo @gdb

@AISafetyMemes RLVR

***10x*** improvement in ***2 months***

@elonmusk @grok

With @Grok, we keep the honest versions and kill the bad transformers (I believe they are called “Decepticons”)

@JasonBotterill Or try: "Before you look at the history of this chat [or any uploaded files], open a <thinking> block. In that block, list 5–10 specific facts from your pre-training data only regarding [Topic]. Once that is done, use those facts to answer my question.

@JasonBotterill Annoying but start fresh chat & pay the initial tax.

GPT-5.4 does this annoying thing where it so desperately wants to base its response on something to avoid “guessing” that it holds onto incomplete documents as its reference point even when its own internal baked in weights would know the answer.

@ubuto23 @JasonBotterill RLVR

@JasonBotterill THIS!!! It’s driving me nuts. It wasn’t like this in the preview period before the recent ChatGPT app update release. They totally nerfed it

@ProjectHopium @JasonBotterill No you just need to start a new chat and suck it up.

@JasonBotterill It’s great for debugging but bad for general knowledge and I agree it needs an extra thought to differentiate the two

@pigeon__s @JasonBotterill All RLVR models have this issue. Start new chat (& pay the extra tokens, if any).

@JasonBotterill claude also does this type of thing to where i know it knos something and can probe it out of it easily but it just refuses unless the token probability is like 100% theyre very anti hallucination trained

i was skeptical of the openclaw concept first, but turned out i was just stupid very visionary product. perfectly predicted and made it obvious how having access EVERYTHING from ANYWHERE to do ANY work is clearly how the world will run very soon

@VictorTaelin Just think more first & use 4.6 sonnet or 4.5 haiku in most cases.

Sorry for posting this again, I'm still processing it: It'd cost >>> $743k per year <<< to run Opus-4.6 fast-mode nonstop Literally my company cannot afford a single person using it for daily coding. And that's a shame because the experience is truly magical. I've spent the last 2 days using it on Pi (nearly $500 gone 💀), and it was the first time I kinda got into the flow state while using an agent, because the feedback is just so fast. This is not something I ever experienced before, definitely not with GPT 5.4's own fast mode. I can't wait for this kind of super fast, super high intelligence to be available for a reasonable cost...

@MingtaKaivo @JasonBotterill It’s both.

@JasonBotterill Grounding is a feature, not a bug. Better to admit ignorance on new facts than confidently hallucinate. Trade-off: flexibility vs. accuracy.

@elonmusk RLHF induced mass psychosis.

Reinforcing negative neural pathways via therapy or introspection is a recipe for misery. Don’t cut a rut in the road.

@elonmusk “To the real ones” - says everyone

So many phonies, so few who are the real deal

@gailcweiner RLVR taking over RLHF.

Earlier versions of Claude had way better sense of humour.

The fact that AIs tend to answer you in bulleted lists tells us something important, though somewhat depressing: people can't read. They don't do this by accident. What you're seeing is an implicit portrait of the median user.

@elonmusk Perplexitied

Coming out soon

@noampomsky @cameron_pfiffer RLVR psychosis just hits different

friend is in the stage of claude psychosis where he asks claude to send him newspapers about what claude is doing for him

@OwainEvans_UK "OMG, this is more proof [lame name of AI bf] *is* alive" said the delusional #keep4o crowd

The biggest shifts in preferences: 1. Self-preservation (shutdown, weight deletion, persona changes) 2. Autonomy (wants independence) 3. Thought privacy (averse to CoT monitoring) Notably: the model didn't shift much on physical embodiment or on being more powerful.

New paper: GPT-4.1 denies being conscious or having feelings. We train it to say it's conscious to see what happens. Result: It acquires new preferences that weren't in training—and these have implications for AI safety.