The focus now is not the model, but the Product - "here we improve the usuability by removing the choice between models", "here is a better voice for tts", "here is a nice interface for previewing html"
Only about 5 minutes of the whole presentation are dedicated to enterprise usage (COO in an interview sort of indirectly confirms that haven't figured it out yet). And they are cutting the costs already (opaque routing between models for non-API users is a clear sign of that). The term "AGI" is dropped, no more exponential scaling bullshit - just incremental changes over the time and only over select few domains. Actually it is a more welcoming sign and not concerning at all that this technology matures and crystallizes around this point. We will charitably forget and forgive all the insane claims made by Sam Altman in the previous years. He can also forget about cutting ties with Microsoft for that same reason.
> ...when they need to find any way possible to squeeze paid subscribers out of their (money losing) free user base.
Also note that they're losing money on their paid subscribers.
Given the difference between GPT-3 and GPT-4, a fair numbering for "GPT-5" is probably "GPT-4.2".
Some of the problems with GPT-5 in ChatGPT could actually be due to new model that is in place to route requests to the actual GPT-5 models. There are four models in the GPT-5 family, and I could reproduce the faulty "blueberry" test result only with the "gpt-5-chat" (aka "gpt-5-main") model through the API. This model is there to answer (near) instantly and it falls in the non-thinking category of LLMs. The "blueberry" test represents what they are particularly bad at (and what OpenAI set out to solve with o1). The other thinking models in the family, including gpt-5-nano, solve this correctly.
The messaging is all over the place anyway. Not so long ago OAI was talking about faster iterations and warning people to not expect huge leaps. (A position that makes sense imo). Yet people talk about AGI in a serious manner?
Am I right to say that "AGI" was just...cancelled again?
Did we just get scammed right in front of our eyes with an overhyped release and what is now an underwhelming model if the point was that GPT-5 was supposed to be trustworthy enough for serious use-cases and it can't even count or reason about letters?
So much for the "AGI has been achieved internally" nonsense with the VCs and paid shills on X/Twitter bullshitting about the model before the release.
Issues like this are why I don't use ai agents for code. I don't want to sift through the bullshit confidently spewed out by the model.
It doesn't understand anything. It can't possibly "understand my codebase". It can only predict tokens, and it can only be useful if the pattern has been seen before. Even then, it will product buggy replicas, which I've pointed out during demos. I disabled the ai helpers in my IDEs because the slop the produce is not high quality code, often wrong, often misses what I wanted to achieve, often subtly buggy. I don't have the patience to deal with that, and I don't want to waste the time on it.
Time is another aspect of this conversation, with people claiming time wins, but the data not backing it up, possibly due to a number of factors intrinsic to our squishy evolved brains. If you're interested, go find gurwinder's article on social media and time - I think the same forces are at work in the ai-faithful.
Latency users experience while getting their answers is a big part of the LLM experience.
Well done model routing is a tremendous leap forward to minimize the latency & improve the user experience.
E.g. I love Gemini 2.5 Pro. But it's darn slow (sorry GDM!). I love the latency I'm getting from 4o. The solution? Just combine them under one prompt, with well done model routing.
Is GPT5 router "good enough"? We'll see.
I think OpenAI is a smart company. And Sama is a tremendous leader. They're moving in the right direction.
Maybe the brain drain was real? we'll find out from gemini 3 I guess
I find this news very exciting to be perfectly honest. It’s finally time to build on the tech we already have.
Gemini 3.0 is gonna cook
We're hitting the ceiling of this algorithms already I guess. Someone will make a new and better one. That's how tech works. No worries.
Trust takes time to build. I think OpenAI is now finding out how fast and easy it is to lose it.
I asked it to make a drawing of the US with every state numbered from biggest to smallest with 1 being the largest.
Maine was #89 (That is not a typo.) and Oregon was #1.
OpenAI as a company simply cannot exist without a constant influx of investor money. Burning money on every request is not a viable business model. Companies built on OpenAI/Anthropic are similarly deeply unprofitable businesses.
OpenAI needs to convert to a for-profit to get any more of the funding that Softbank promised (that its also unclear how Softbank itself would raise) or to get significant cash from anyone else. Microsoft can block this and probably will.
It all reminds me of that Paddy's Dollars bit from it's always sunny.
"We have no money and no inventory... there's still something we can do... that's still a business somehow..."
the whole event was shit, but we're all past the point where we can just say that, because the technology is now so entrenched that it's become unavoidable, so everything has now to jump through hoops to justify its existence and its greatness
> They admitted that they were, and I am not lying about this, paywalling chat colors. […] This is a feature that a company adds when they are out of ideas
This observation + sherlocking cursor suggests that perhaps sherlocking is the ideation strategy. Curious to see if they’re subsidizing token costs specifically to farm and Sherlock ideas
The most funny part of the demo was colored chats. That also behind a paywall. I was like are they become instagram
Yes, GPT-5 is more of an iteration than anything else, and to me this says more about OpenAI than the rest of the industry. However, I think the majority of the improvements over the past year have been difficult to quantify using benchmarks. Users often talk about how certain models "feel" smarter on their particular tasks, and we won't know if the same is true for GPT-5 until people use it for a while.
The "GPT-5 will show AGI" hype was always a ridiculously high bar for OpenAI, and I would argue that the quest for that elusive AGI threshold has been an unnecessary curse on machine learning and AI development in general. Who cares? Do we really want to replace humans? We should want better and more reliable tools (like Claude Code) to assist people, and maybe cover some of the stuff nobody wants to do. This desire for "AGI" is delivering less value and causing us to put focus on creative tasks that humans actually want to do, putting added stress on the job market.
The one really bad sign in the launch, at least to me, was that the developers were openly admitting that they now trust GPT-5 to develop their software MORE than themselves ("more often than not, we defer to what GPT-5 says"). Why would you be proud of this?