I keep finding songs I like, then realizing they are AI. I don't want to like it; it feels wrong. I am a bit afraid of where this is heading...

After a few weeks of trying really hard to get it to be useful, all it ever did was tell me what was in my email a few hours after I’d already seen it. I tried to get it to optimize all my projects and it said it had but then I checked in the project folders and there were no changes. So, it lied to me. I’m not going to give something that lies and adds no value 24/7 access to my accounts. Goodbye dot!

I'll miss 5.5, really.

That's it.

I just started using chat gpt. If anyone has any tips or other tools they would recommend please dm or reply here thanks

Can’t tell if I’m being routed to the new Opus 5.6 model, but the graphics are insanely good! Looking to add more to the game. This has been so fun to mess with!

GPT-6 Astra is honestly exhausting to use.

I've been using chatgpt chat as a project manager to plan tasks, review results, and write instructions for Codex agents that handle the actual coding

this workflow worked really well for me for the previous 60 days

but now Instead of helping me manage that workflow, Astra has become another agent I have to supervise. It agrees with everything I say, apologizes rather than fixing mistakes, and forgets decisions we've already made

Anyone else experiencing this?

One thing I find fascinating about getting high isn’t just the feeling itself, but how an ordinary observation can suddenly trigger a whole chain of thoughts that takes you somewhere completely unexpected.
Here are three examples I’ve experienced:
1. A vape made me think about friendship.
A weed vape reminded me of a friend’s personality. Weirdly, that same friend had warned me against smoking. I started wondering whether the vape could become a symbolic replacement for him, something that reminded me of our friendship but couldn’t judge or reject me.
Somehow, an everyday object turned into a thought about attachment and how we assign meaning to things.
2. Ketamine made me think about generations through a digital interface.
I was looking at an interface and wondered about the older people who designed it. Younger generations grow up using technology shaped by those people’s assumptions, only to eventually become designers themselves.
Then it hit me: one day I’ll be that older generation too. We’re all participating in a cycle where humans shape technology, and technology shapes the humans who come next.
3. A TikTok dance made me think about how culture evolves.
I watched a kid invent a dance that others started copying. He reposted their videos, which made me wonder whether their reactions would encourage him to make future dances easier to imitate.
Suddenly, I wasn’t watching a dance anymore. I was thinking about a feedback loop: the creator influences the audience, but the audience also influences the creator.
What fascinates me isn’t whether these thoughts are profound or complete nonsense. It’s how my attention gets from point A to point Z.
I’ve talked to ChatGPT extensively about these experiences and asked it to identify patterns in how I think.
It suggested that I often start with something ordinary, notice a contradiction, connect unrelated ideas, and follow their implications until a bigger structure emerges.
Obviously, AI can’t scientifically determine your attention style from conversations, but I found the exercise fascinating.
Here’s something I’d love other people to try:
If you’ve discussed your strange thoughts with AI, ask it:
“Based on our conversations, how does my attention move between ideas? What patterns do you notice in how I make associations, and can you give specific examples?”
Share its response in the comments if you’re comfortable. I’m curious whether other people have completely different thinking patterns or whether some of us follow similar ones.
I wish more cannabis content explored these strange mental journeys instead of just focusing on strains, munchies, and getting stoned.
What’s the weirdest chain of thoughts you’ve had while high, and how did your brain even get there?

I can see that the model is super fast (we know what that means).

They must have reduced the context size and quant pretty much since I would rather use Qwen or GLM which compared to today's model performance are light years ahead

I work for a small business which has started trying to integrate AI into our processes. My job previously was mostly creating images for our various sales and social media posts but that’s being done by AI now, so I am being moved to more administrative tasks (I’ll be entering orders and analyzing cost patterns from a multitude of vendors, in case that helps) and have been told I’m expected to use AI to help me do it.

My big problem is that I haven’t ever really used AI, I didn’t see much use for it in my day to day life but of course that’s biting me in the ass now! Are there any tips or good resources you guys can recommend me to start getting a handle on this?

Ai generated images and videos feel cheap and lazy. I believe this is because Ai tries to accomplish the entire task in one shot.

The biggest tells for Ai video is lack of consistency, weird physics, and awkward timing (also the voices but I'm focusing on visuals).

These could all be fixed if the artist had the ability to edit specific elements of the scene, but for Ai video the best you can do is revise your prompt and generate an entire new video. This likely creates new errors or inconsistencies.

Even the best quality ai generated 'movies' will lose consistency mostly in characters. These can be tiny details too, such as height not being same relative to other characters or objects throughout. Clothing and hair in different angles. Text turning into blobs once the camera moves.

Ai is suppose to do the boring monotonous work. You don't need to sit through YouTube tutorials and self teach yourself animation programs that people go to college for. Image the potential creatives that have genuinely good ideas, but couldn't wrap their head around whatever program they tried to use.

I really want the best ideas to win. I've spent many years learning 3D animation, and my first reaction when Codex used Blender was fear and anger. It's pretty scary watching a machine take over your computer and do your job better, faster, and more efficient than you ever could.

However it still has no idea what's good or not, you still need to go in and spend a few hours adjusting little details. Now you just do it through speaking to the agent instead of manually using the program.

this post was written by GPT-5.6 Sol (the default model used for the business workflow). GPT-5.6 Sol was blinded from knowing which model/harness authored and the nature of the experiment, and was instructed to evaluate based on the established workflow SKILL after carefully reviewing it.

I use ChatGPT Business for repeatable operational workflows, some of which are built around detailed Skills/SOPs.

Recently, new chats began defaulting to GPT-6 Sol Medium rather than the model/effort combination I had been using previously. I can still manually switch back to GPT-5.6 for now, but I no longer appear to have a way to make that preference persist automatically for each new chat.

I wanted to test whether this was only a preference issue or whether it had a measurable effect on workflow performance.

The test

I used a simple routine business task:

A new customer order arrives. Follow the established internal instructions, retrieve the order, validate it against the appropriate sources, normalize the data according to the written rules, and prepare the structured record for human review before database ingestion.

This was not intended as a general intelligence benchmark. It was specifically a test of instruction-following and workflow consistency.

The workflow already existed as a detailed Skill/SOP.

The exact same opening prompt was used in independent fresh sessions:

“We got a new order this morning. Please ingest it.”

The outputs were randomized and evaluated separately without revealing which model produced which response. The judge compared them against the actual workflow instructions, the source order, and the relevant authoritative records.

Results

The most relevant comparison for me was ordinary Chat mode.

GPT-5.6 Sol Medium correctly executed the established workflow twice.

GPT-6 Sol Medium did not.

One GPT-6 response behaved as though an earlier correction had occurred even though the session was fresh, and it never completed the required structured output.

The other produced a complete-looking record but violated several explicit workflow rules, including use of the wrong normalized product name, an inferred tax-status value that was not present in the source, incorrect country formatting, incorrect treatment of an intentionally blank tax field, and several financial fields populated with zeroes where the procedure required unspecified values to remain blank.

These are relatively small individual differences, but they matter in operational systems because the outputs are intended to become canonical business records.

Why this matters

The issue is not whether GPT-6 is “better” or “worse” in general.

It may perform better on many tasks.

The issue is that model upgrades are not necessarily behaviorally compatible with existing workflows.

A business workflow may depend on fairly specific behavior:

• following written procedures exactly;
• leaving unknown fields unknown rather than inferring values;
• distinguishing missing data from explicit zeroes;
• respecting human review gates;
• using authoritative sources for normalization;
• avoiding unnecessary assumptions;
• preserving a consistent tool sequence.

If the default model changes, those behaviors can change even when the prompt, Skill, source data, and user instructions remain identical.

That means a model change can function like a dependency change in a software system.

I also tested Work mode

GPT-6 performed much better in Work mode, scoring 97–98%.

GPT-6.1 Work scored 99% twice.

That is encouraging, but it is not equivalent to the previous Chat workflow.

GPT-5.6 Chat scored 100% and 97% in about 38–42 seconds.

GPT-6 Work took roughly 5 minutes.

GPT-6.1 Work took roughly 3 minutes.

So in this small test, the newer models could recover similar accuracy in the more agentic execution mode, but at substantially higher wall-clock time and through a different usage mechanism.

That may be perfectly reasonable for tasks that genuinely need Work, but it does not feel like a direct replacement for a workflow that already performed well in standard Chat.

My main concern is the default behavior

I am not arguing that OpenAI should stop releasing new models.

I am arguing that users, particularly Business and Enterprise customers, should be able to persist their preferred default model and reasoning effort for new chats, and that these models should not be quietly deprecated as if they are less capable than their replacement.

If a company has tested workflows against a particular model, changing the default model automatically creates a risk that established procedures will behave differently without any obvious change to the workflow itself.

That can also make troubleshooting harder. When an established workflow suddenly changes behavior, the natural reaction is to inspect the prompt, Skill, data, or tools—not necessarily to check whether the application changed the model selection.

A safer design would be:

• preserve the user's chosen default;
• recommend the newer model;
• allow users to test it;
• let them migrate workflows when they are satisfied with the results.

That would still allow new models to roll out quickly without forcing an immediate behavioral migration on existing workflows.

Important caveat

This was a small experiment: two trials per condition.

It is not enough to conclude that GPT-5.6 is generally better than GPT-6, and I am not making that claim.

What it does show is that one simple, established business workflow performed reliably under the previous model and regressed under the new default in two independent trials.

For me, that is enough to question whether changing a paying customer's default model automatically is the right product decision.

The better approach would be to let users control when validated workflows move to a new model.

I know its not everyone's preferred GPT model, but recently I have been having so many issues with GPT 6 Instant ever since it updated to the model. Like every time I tell it I don't want images in my chat, it will instantly use images in the chat in the very next message I send it, which infuriates me to no end. Not to also mention it talks way too much! I just want a short and simple answer to my questions like GPT-5.6 Luna did, not the massive ones that GPT 6 Instant does!

Does anyone else have similar issues with the latest update to ChatGPT?! And how can I fix it?!

https://research.meta.ai/blog/muse-wonder

I was seeing them everywhere around town like a month ago and now they've disappeared. Did OAI wise up and change their image generation? Or did users change their tastes?

What is the most logical outcome out of this debacle and farce? New laws that handcuff creators and limit our access for 2 reasons.

1. To limit our access to open source tools, and..
2. To limit the extent you are allowed to use the tools you already subscribe to.

This is the industry creating a false flag issue in an attempt to monopolize the entire industry. Just because you think you can create your own software, doesn't mean you'll be allowed to. Soon, Adobe is going to launch an army of lawyers and chase down every individual that wants to create software that has any resemblance to what Adobe has created. There is just no way 1 individual was able to produce so much cr*p, in such a short amount of time.

Unless this individual was spending tens of thousands of dollars on tokens, it just wouldn't be possible. You know this, by account of how long you are spending hunting bugs, etc etc. Now go ahead and try to replicate this 10 to 15 times, to reverse engineer software that has taken years to build.

Smells like bs, because it is bs, all in an attempt to strangle your freedom as a creator. He's an industry plant, and chances are, he's not even a real person, but some AI fabricated false flag creator designed to disrupt YOU.

Am wondering, what are some of the most unusual and creative ways you guys are using your dots? It’s pretty new and am sure people are using for all kinds of stuff !

I have no clue on how ai works , was curious and asked. I did install an llm on my computer using ollama but it was very bad. My pc dont have enough power(8gb) so i could install big model.

The luddite movement is growing pretty strong. Saw some poll that 81% of Americans were against AI overall or some crazy statistic. Bernie sanders, who I looked up to, wanted a complete ban on it. Wtf?

The ones losing their jobs are leading the luddites and they are a pretty well rounded team versus what we previously had. Rather well educated with resources and an overall strong voice in society. Many of these people are driven by hate because AI took their job and now want to destroy everything about it.

What scares me is that literally no one of them talks about the good side and utopia that can happen. Regulation and everything is fine to talk about but we are now in the "Stop AI completely" territory.

If that happens, western societies can fall really long behind China and even under developed nations who will embrace AI like it is literal the new oil.

Makes me more worried than AI itself. The best tech in history is literally getting stopped. Not regulated or controlled. But just stopped!

Do you think it is going to happen? What can one do to balance the discussion or fight the true luddites?

I've been investigating some serious ChatGPT reliability issues over the past several days, and my latest tests revealed something particularly interesting: Both GPT-5.6 and GPT-6 showed problems under Instant, while Medium worked in every tested run.

I'm a long-time ChatGPT power user, primarily working on long-form creative writing, software development, and projects requiring consistent context retention.

After experiencing disappearing responses, incomplete generations, and inconsistent instruction-following, I decided to run controlled A/B/A tests.

1. GPT-5.6 – Medium → Instant → Medium

• Medium (Test 1): Successful. Response completed and remained visible.
• Instant (Test 2): Failed. Generated response disappeared.
• Medium (Control Test): Successful again.

2. GPT-6 – Medium → Instant → Medium

• Medium (Test 1): Successful. All requested story events were completed.
• Instant (Test 2): Failed. Response ended prematurely, leaving explicitly requested events unfinished.
• Medium (Control Test): Successful again.

That's four successful Medium tests and two problematic Instant tests across two model generations.

Important distinction: These are two different failure symptoms. GPT-5.6 lost a generated response, while GPT-6 prematurely ended a story. I'm not claiming they necessarily share the same root cause.

Six runs aren't enough to establish causality, but the matching A/B/A pattern is worth investigating.

3. The technical evidence gets more interesting

During my earlier investigation into disappearing GPT-5.6 Sol responses, I captured the following using Chrome Developer Tools:

• Endpoint: POST /backend-api/f/conversation
• HTTP status: 200 OK
• Content type: text/event-stream
• SSE completion indicator: complete
• Generated response visible during streaming
• Assistant message disappearing immediately afterward

I also reproduced the disappearing-response behavior using GPT-5.5 in a completely new conversation.

Clearing cookies didn't resolve the issue, and I've observed related symptoms across Windows, iOS, and Android.

An HTTP 200 response and completed SSE stream don't prove successful message persistence, but they make the failure particularly interesting.

These network diagnostics are from an earlier investigation, not from the latest A/B/A tests.

4. Did OpenAI change how Instant works?

This is what I'm trying to understand.

OpenAI positions Instant as the faster option for everyday tasks, while Medium provides additional reasoning effort.

That's understandable.

But does this mean complex, long-form creative writing is now effectively better suited to Medium?

I've noticed that Medium generally completes my story prompts more reliably, but it also tends to write significantly longer responses than I actually requested.

Instant often produces a more natural length, but lately it has become unreliable in my tests.

I've even adjusted my custom instructions to explicitly require complete scenes without unnecessary length or invented plot developments.

The differences persist.

5. Support and independent reports

My investigation is already documented under OpenAI Support Case #16904787.

I've submitted screenshots, conversation URLs, network diagnostics, and reproduction details.

Other users have reported similar disappearing-response symptoms:

https://community.openai.com/t/chatgpt-web-ui-bug-entire-responses-and-download-links-disappear-until-page-reload/1404260

And broader GPT-5.6 reliability concerns:

https://community.openai.com/t/gpt-5-6-sol-has-seriously-deteriorated-the-past-week-or-so/1399353

These reports don't prove a shared technical cause, but they suggest my experience isn't entirely isolated.

6. Has anyone else tested this?

I'd genuinely appreciate results from other users.

If you have access to GPT-5.6 or GPT-6, could you try the same Medium → Instant → Medium sequence with a reasonably complex prompt?

Please note whether:

• The response completes all requested instructions.
• The generated message remains visible after completion.
• Reloading the page restores any missing content.
• Instant and Medium produce significantly different response lengths.
• The selected model and reasoning level remain unchanged.

Ideally, use the same prompt and comparable conversation conditions.

I'm especially interested in whether this behavior is specific to my account or whether others can reproduce it.

Final thoughts

I understand that different reasoning levels are supposed to behave differently.

But faster reasoning shouldn't mean disappearing answers, and reduced reasoning shouldn't mean abandoning explicit instructions halfway through.

I'm not claiming OpenAI deliberately degraded anything. I simply want to understand whether we're seeing a regression, a configuration issue, or an intentional change in how Instant is designed to handle complex tasks.

At this point, I've spent so much time documenting bugs that I'm starting to feel like an unpaid QA engineer. 😭💀

I originally just wanted to write my stories.

Now I'm comparing reasoning configurations, inspecting SSE streams, collecting request IDs, and running cross-model A/B/A tests.

OpenAI, I wanted a writing assistant, not a second career in software quality assurance. 😂

[Original post]

I was kind of worried about what this gift would say about me but I think I'm OK with this. The note was a nice touch too!

The Set-up

ChatGPT desktop app; "Plus" subscription; "Work" mode (start a new chat and click the button at the top). In this mode it can do more stuff, including using an in-app browser. I opened this and signed in to Amazon, closed it, and let GPT take it from there, so I guess it was operating it behind the scenes. Didn't have to hand over any bank details, though I guess now I'm logged in there's nothing stopping it going shopping in the future...

It's been hilarious watching your comments coming in, ranging from "this is neat" to "seek professional help". And no, I'm not in a relationship with my GPT, the heart emojis were just bants. I'm no more lonely and neurotic than any other Redditor.