Vinicius Sanches

@joshwoodward @GeminiApp Amen! Sucks that the fix for iOS is still a few weeks away but it’s gonna make dictation finally usable.

✅ Papercut fixed: Gemini won’t cut you off if you pause while talking on Android anymore. (iOS in a few weeks!) So next time you hit the mic icon, feel free to pause, take a breath, or ramble. No more anxiety to speak it all out before @GeminiApp jumps in prematurely.

@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.

@ArchSenex what kind of censorship are you talking about? examples would be helpful!

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…

@JJ_McCullough @ElevenLabs That's very odd. For me the Reader app follows word by word the text. Did you happen to import the text via PDF? Maybe the issue was extracting the text from it.

@vsaroca @ElevenLabs Yes

I just had a weird experience using the popular app @ElevenLabs, which uses an AI voice to read things. I was getting it to read one of my YouTube video scripts back to me, which is something I like to do to check for mistakes. And what was weird was it started ad-libbing. It was reading things I hadn’t written, “improving” my script in various stupid ways as it went along. For example, I have a passage that went: “Now, in the Korean historic cultural canon, one guy looms above everyone else: King Sejong, who ruled from 1418 to 1450. He’s like their George Washington, everyone knows his name and likeness and you constantly see him treated as a sort of patriotic mascot character in pop culture.” And the app instead read: “Now, in the Korean historic cultural canon, one guy looms above everyone else: King Sejong, the fourth kind of the Joseon dynasty. He was born in the year 1433 and he was the one who created the Korean alphabet, Hangul. He was also the one who created the Korean calendar, which is based on the lunar cycle. He was a great king, and he was a great man.” So you see, it basically just made up its own little story. It is very frustrating that these AI apps can’t stop hallucinating even when asked to perform the most basic of route tasks.

@WesRoth I am surprised people are only commenting on this now. Both GPT 5.3 and 5.4 have the same reflex in almost every response. It’s incredibly annoying, unproductive and makes using the models unpleasant.

I've ran the same prompt for deep research through GPT 5.4, Opus 5.6 and Gemini Deep Research (I assume Gemini 3.0) most of them ran for ~30 mins GPT 5.4 is *REALLY* annoying! it's "reflexively contrarian", it prioritizes showing you what's wrong with your thinking, NOT actually helping you solve the problem ME: my house is on fire! GPT 5.4: While it's true that combustion is occurring, it's important to note that not all of your house is on fire. The garage, for instance, appears structurally intact. (this is a pattern with it, btw, many such examples) I'm not sure if this is because these are health related questions, but this has been an incredibly annoying model for this specific task

@chatgpt21 That's a great point.. 5.3 has been the worst at this from my experience.. it stays in the middle ground and doesn't update its reasoning.. often if you keep pushing, it starts violating the Model Spec HARD to maintain its stance

GPT 5.4, in the best / weirdest way possible feels like AGI but in the way people were trapped in the movie get out. For example. It will still make easy mistakes like this that prove that we need parametric intelligence more than ever. However, when you push it back, it’ll push back on you even though it knows it’s wrong but inside it’s reasoning while it’s saying it’s right it’ll explain how the user is actually right and it feels like it knows why and how it’s wrong but for some reason, there’s an anti-hallucinatory neuron, pushing back too hard, the model can seem to internally track the correction but still overcommit to anti-sycophancy. Forcing the model not to think for itself almost. We had the opposite problem a year ago, where the model would just lean into everything you’re saying - now it feels like they’ve almost over corrected where it’ll push back but know it’s pushback might be wrong.

@lefthanddraft Really fun prompt to run on the Arena. Both Qwen and Grok also do not fall for it.

This thing is going to find a cure for cancer before it stops falling for dumb tricks.

@lefthanddraft I just used the riddle, no extra wording. Sometimes o3 get these right, but this one it didn’t. Gemini 3.1 Pro gets it right though.

@scaling01 Which parameters did you use? How did changing the temperature affect the exploration?

some more trajectories for GPT-5.4 and Opus 4.6

@AcerFur @jkrt151 @Liam06972452 Grok also relies heavily on search now to try finding a solution if it believes the problem is a general one and the answer would be on the web

@jkrt151 @Liam06972452 GPT models have a tendency to search for the problem on the web. If it finds out it's an open research problem, it will not make a good attempt at the problem and instead say it's open and summarise the statement and attempts at it.

1/8 Two days ago, @Liam06972452 prompted GPT-5.4 Pro using our workflow that had been working for the Erdős problems thus far, and was able to eventually obtain a solution to epoch.ai/frontiermath/o…: chatgpt.com/share/69ae3b99…

@nicdunz Change the prompt to add a short distance instead

cool

Btw 5.1 Thinking also solves it. Couldn’t get any other version to do it. #keep51 #keep4o

Overfitting now on Thinking models? GPT 5.4 thinks about it for second and confidently answers wrong.. o3 still seems to actually think.. #ChatGPT #OpenAI

@andrewjclare They fixed the keyboard apparently, that’s a win enough for me. It was unusable.

iOS 26.4 Beta 3 is here. Weird update!

@JustinBleuel @ChatGPTapp Speech to text is glitching on iOS - it sometimes switches to the earphone speaker and the only way to restore it is to kill the app and reopen.

Another week shipping @ChatGPTapp! Try these out, we're focused on sweating the details, some bigger features coming 🔜 - Edit messages that have an image attachment - Chat search cmd/ctrl+click open results in new tab - Conversation share menu loads faster - Fixed 5.2-thinking bug causing msg streaming errors - Export data viz as images from Code Blocks - Mobile web sidebar actions no longer stick on scroll Android - Auto scroll to bottom on existing convos - Sidebar actions complete faster - Fix voice mode button flicker on conversation open - Dictation mode visual improvements - Keyboard auto-opens when attaching a file Let us know what else you'd like to see

@annapanart It’s the most reasonable approach and name they could use.

5o next?

@UpdatesFromYT @youtubemusic A light theme option would be neat

YouTube Premium and @YouTubeMusic Premium users can now instantly turn an idea, mood or genre into a personalized playlist using our new AI playlist feature 🎧 Here's how to use it: 1. On Android & iOS, tap the “New” button in the Library tab of YouTube Music 2. Select “AI Playlist" 3. Use your voice or text to explain the type of music you’d like in your playlist

@gailcweiner Just gave it another try today.. not only is the audio pitch bad but the information it gives is shallow and often misleading. Asked for a “Review past context for accuracy” in text later and it flagged many problems.

Serious question does anyone here use GPT advanced voice and actually enjoy speaking with it?

@apples_jimmy I genuinely want to read the full paper now..

Nobel prize coming soon

@gailmoon Tried recently, it has definitely improved but still fairly distant and cold in comparison to 5.1

GPT 5.2 is currently amazing. Have you tried it in the past week?