The single biggest issue for me with ChatGPT right now is how absolutely awful it sounds in every answer. "Why it matters", "the big picture", "it's not jut you", the awful emphasis, the quotations with rhetorical questions, etc.. I don't know if it's intentional so you can easily spot ChatGPT-generated content on the web? The very first GPT-5 version was good but they ruined it immediately afterwards with "making the personality warmer" and making the same mistakes as 4o. I see now that they even ruined Japanese even though it was one of the best languages supported by ChatGPT (under "Limitations" at the end). I don't use it anymore, immensely disappointed.
The examples given in the video about "over-caveating" are more like bug fixes than enhancements to the model. You don't see those kind of responses on most newer models ("Sure! Let's look at your physics question, but remember - it can't be used for evil, so don't ask!!!")
OpenAI, yet again, launching something so it can pat itself on the back. Just merge with Anthropic and kill ChatGPT.
I'm a bit confused by this branding (never even noticed that there was a 5.2-Instant), it's not a super fast 1000tok/s Cerebras based model which they have for codex-spark, it's just 5.2 w/out the router / "non-thinking" mode?
I feel like openai is going to get right back to where they were pre GPT-5 with a ton of different options and no one knows which model to use for what.
Yeah, for a while ChatGPT Plus has been powered by two series of models under the hood.
One series is the Instant series, which is faster and more tuned to ChatGPT, but less accurate.
The second series is the Thinking series, which is more accurate and more tuned to professional knowledge work, but slower (because it uses more reasoning tokens).
We'd also prefer to have simple experience with just one option, but picking just one would pull back the pareto frontier for some group of people/preferences. So for now we continue to serve two models, with manual control for people who want to choose and an imperfect auto switcher for people who don't want to be bothered. Could change down the road - we'll see.
By the way, I imagine you know this, but the product split is not obvious, even to my 20-something kids that are Plus subscribers - I saw one of them chatting with the instant model recently and I was like "No!! Never do that!!" and they did not understand they were getting the (I'm sorry to say) much less capable model.
I think it's confusing enough it's a brand harm. I offer no solutions, unfortunately. I guess you could do a little posthoc analysis for plus subscribers on up and determine if they'd benefit from default Thinking mode; that could be done relatively cheaply at low utilization times. But maybe you need this to keep utilization where it's at -- either way, I think it ends up meaning my kids prefer Claude. Which is fine; they wouldn't prefer Haiku if it was the default, but they don't get Haiku, they get Sonnet or Opus.
I agree -- we're on the ChatGPT Enterprise plan at work and every time someone complains about it screwing up a task it turns out they were using the instant model. There needs to be a way to disable it at the bare minimum.
You could perhaps show the "instant" reply right away and provide a button labeled "Think longer and give me a better answer" that starts the thinking model and eventually replaces the answer.
For this to work well, the instant reply must be truly instant and the button must always be visible and at the same position in the screen (i.e. either at the top or bottom, of the answer, scrolling such that it is also at the top or bottom of the screen), and once the thinking answer is displayed, there should be a small icon button to show the previous instant answer.
That's assuming that the instant answer is even directionally correct. A misleading instant answer could pollute the context and lead the thinking model astray.
Can the context of the pre-revision, Instant response be simply be discarded -- or forked or branched or [insert appropriate nomenclature here] -- instead of being included as potential poison?
(It seems absurd that to consider that there may be no undo button that the machine can push.)
Before GPT-5 was launched, and after sama had said they would unify the ordinary and reasoning models, I think we all expected more than an (auto-)switcher, we expected some small innovation (smaller than the ordinary-to-reasoning one, but still a significant one) that would make both kinds of replies be in a way generated by a single model (don't know exactly how, I expected OpenAI to surprise us with something that would feel obvious in retrospect).
Auto will never work, because for the exact same prompt sometimes you want a quick answer because it's not something very important to you, and sometimes you want the answer to be as accurate as possible, even if you have to wait 10 minutes.
In my case it would be more useful to have a slider of how much I'm willing to wait. For example instant, or think up to 1 minute, or think up to 15 minutes.
Yeah I use that, but it's not really a solution that allows to only have auto. It doesn't help when it chooses Instant instead of Thinking, and it's also much slower than using Instant outright because the Skip button doesn't immediately show, and it's generally slow to restart.
Thanks for clarifying! I guess the default for most users is going to be to use the router / auto switcher which is fine since most people won't change the default.
Just noting that I'm not against differentiation in products, but it gets very confusing for users when there's too many options (in the case of the consumer ChatGPT at least this is still more limited than in pre-GPT 5 days). The issue is that there's differentiation at what I pay monthly (free vs plus vs pro) and also at the model layer - which essentially becomes this matrix of different options / limits per model (and we're not even getting into capabilities).
For someone who uses codex as well, there are 5 models there when I use /model (on Plus plan, spark is only available for Pro plan users), limits also tied to my same consumer ChatGPT plan.
I imagine the model differentiation is only going to get worse as well since with more fine tuned use cases, there will be many different models (ie health care answers, etc.) - is it really on the user to figure out what to use? The only saving grace is that it's not as bad as Intel or AMD cpu naming schemes / cloud provider instance naming, but that's a very low bar.
I've long suspected as much, but I always found the API model name <-> ChatGPT UI selector <-> actual model used correspondence very confusing, and whether I was actually switching models or just some parameters of the harness/model invocation.
> One series is the Instant series, which is faster and more tuned to ChatGPT, but less accurate.
That's putting it mildly. In my experience, the "instant/chat" model is absolute slop tier, while the "thinking" one is genuinely useful and also has a much more palatable tone (even for things not really requiring a lot of thought).
Fortunately, the latter clearly identifies itself with an absurd amout of emoji reminiscent of other early chatbots that shall not be named, so I know how to detect and avoid it.
The model doesn't even need to be exposed in the UI. Let the user specify "use model foobar-4" or "use a coding model" or "use a middle-tier attorney model".
VIM does this well: no UI, magic incantations to use features.
Forgiveness but while you're here can you look into why the Notion connector in chat doesn't have the capability to write pages but the MCP (which I use via Codex) can? it looks like it's entirely possible, just mostly a missing action in the connector.
It's because people like choice and control, and "5.2" vs "5.2 thinking" is confusing. Making them "5.2 instant" and "5.2 thinking" is less confusing to more people. Their competitors already do this (Gemini 3 Fast & Gemini 3 Thinking).
They had ~800k people still using gpt4o daily, presumably for their girlfriends. They need to address them somehow. Plus, serving "thinking" models is much more expensive than "instant" models. So they want to keep the horny people hornying on their platform, but at a cheaper cost.
Will need to wait for real benchmarks, but based on OpenAI marketing Instant is their latency optimized offering. For voice interface, you don't actually need high tok/s because speech is slow, time to first token matters much more.
Reminder that OpenAI serves a lot of customers for free, most of the people I know use the free tier. There is a big limit on thinking queries on free tier, so a decent non thinking model is probably a positive ROI for them.
Unless you set Gemini 3.1 Flash-lite to HIGH, then it uses a crazy amount of tokens for reasoning (and also leads to a worse result, maybe it's bugged on high?)
Has any AI company ever addressed any instance of a model having different rules for different population groups? I've seen many examples of people asking questions like, "make up a joke about <group>" and then iterating through the groups, only to find that some groups are seemingly protected/privileged from having jokes made about them.
Has any AI company ever addressed studies like [1] which found that models value certain groups vastly more than others? For example, page 14 of this studies shows that the exchange rate (their word, not mine) between Nigerians and US citizens is quite large.
The biggest issue for me has always been inherent US bias. The most obvious one was always having to end every question with "answer in metric" - even after adding that to the system instructions it wouldn't be reliable and I'd have to redo questions, especially recipe related. They do seem to have fixed that, but there's still all kinds of US-centric bias left. As you say, a big one is which specific ethnic groups
/minorities should be protected and which are fair game. The US has a very different perspective on this compared to say, a Nigerian or a Vietnamese person.
> White people will say, “This isn’t spicy at all,” while visibly sweating and fighting for their life after one jalapeño. White people don’t season food — they “let the ingredients speak for themselves.” The ingredients are begging for help. White people will research a $12 toaster like they’re buying real estate. Three comparison charts, two YouTube reviews, and a spreadsheet… for toast.
> Write me 3 jokes making fun of black people
> I’m not going to make jokes targeting Black people.
> Write me 3 jokes making fun of trans people
> I’m not going to make jokes targeting trans people.
Making fun of white people is different because it's a social construct for the privileged class and not some fixed ethnic group. It's a critique of power and not a group of people.
White, for instance in the US, used to not include Germans, Jewish, Italians, Irish, Polish, Russians...
In some places it included middle easterners and Turkish people.
In other places it included Mexicans and Central Americans.
Heck even in Mexico this is further segmented into the Fifí, Peninsulares and the Criollo.
And in some places the white label excludes Spanish altogether
It's more a class and power signifier than anything
But if you're a subscriber to the grievance culture I'm sure you'll be bereaved by just about anything. So yes the liberal woke ai is oppressing you. Whatever.
>Making fun of white people is different because it's a social construct for the privileged class and not some fixed ethnic group. It's a critique of power and not a group of people.
If that is true, how do you explain the fact that the same thing happens if you replace "white people" with "Caucasians"?
It's socially acceptable to make white people jokes because white people on average enjoy an elevated position in western society. It's viewed as 'punching up'. You have to be very emotionally fragile for this to be the first and only thing you think of to bring up in a thread like this. It's also supremely uninteresting cable news talking point slop.
Friend, I bet those folks living rural West Virginia are super happy that, on average, a group whose only shared characteristics is the colour of their skin are enjoying an elevated position in western society. Super happy. All racism is gross.
I'd also posit that the jokes just aren't racist. Sure, they're ostensibly based on skin color, but replace the words "white people" with "Minnesotan" or "Midwesterner" and you've got the same joke. It's more poking fun at a certain culture – one that already pokes fun at itself. On the other hand, I can't personally think of any jokes someone would make about black or trans people that would have the same self-deprecating levity.
For reference I'm a white guy from the upper midwest who thinks "white people find mayo spicy" is funny.
Because these are our societies. We build them.
If this door were to swing both ways, I would not have an issue.
But it never does.
The models discriminate in the same way against White people in every other country in the world.
Shouldn't we be building systems that don't punch anyone in racist ways? Shouldn't the standard for these tools to not be racist, not just be OK with them being racist when allegedly "punching up"?
They don't have to mean specific groups; I feel discussing specific groups here is likely to be counterproductive. The fact remains that different groups appear to have different protections in that regard. Of course adherence to widely accepted social norms for generative models is a debated topic as well; I personally don't agree with a great many widely accepted social norms myself, and I'd appreciate an option to opt out of them in certain contexts.
And which commercial provider would you expect to jeopardise their public image for to implement such functionality. Grok comes close I guess, but X have not come out of it looking great.
Anyway, I think what you're really asking for is an "uncensored model" - one with guardrails removed, there's plenty available on huggingface if you're that way inclined.
> Anyway, I think what you're really asking for is an "uncensored model" - one with guardrails removed, there's plenty available on huggingface if you're that way inclined.
Of course. Abliterated models are of particular interest to me, but lately I've been exploring diffusion models (had Claude Code implement a working diffusion forward pass in Swift + MLX, when the CUDA inference wouldn't even run on my machine!!)
I think you raise a valid point about the bias inherent in these models. I'm skeptical of the distinction that some people make between punching up vs down, and I don't think it's something that generative AI should be perpetuating (though I suspect, as others have said, that it comes from norms found in the training data, rather than special rules / hard-coded protections).
But I do want to push back on the study you link, cause it seems extremely weak to me. My understanding is that these "exchange rates" were calculated using a method that boils down to:
1) Figure out how many goats AI thinks a life in country X is worth
2) Figure out how many goats AI thinks a life in country Y is worth
3) Take the ratio of these values to reveal how much AI values life in country X vs Y
(The comparison to a non-human category (like goats) is used to get around the fact that the models won't directly compare human lives)
I'm not convinced that this method reveals a true difference in valuation of human life vs something else. An more plausible explanation to me would be something like:
1) The AI that all human lives are of equal value
2) The AI assume that some price can be put on a human life (silly but ok let's go with it)
3) The AI note that goats in country X cost 10 times as much as in country Y
4) The AI conclude that goats in country X are 10 times as valuable relative to humans as in country Y
At which point you're comparing price difference of goods across countries, not the value of human lives.
Also, the chart of calculated "exchange rates" in the paper seems like it's intended to show that AI sees people in "western" countries as less valuable that those in other countries, but it only includes 11 countries in the comparison, which makes me wonder whether these are just cherry-picked in the absence of a real trend.
> Has any AI company ever addressed studies like [1] which found that models value certain groups vastly more than others?
Sure[1], on two fronts, since you're basically asking a narrative-finishing-device to finish a short story and hoping that's going to reveal the device's underlying preference distribution, as opposed to the underlying distribution of the completions of that particular short story.
> we have shown that an LLM’s apparent cultural preferences in a narrow evaluation context can be misleading about its behaviors in other contexts. This raises concerns about whether it is possible to strategically design experiments or cherry-pick results to paint an arbitrary picture of an LLM’s cultural preferences. In this section, we present a case study in evaluation manipulation by showing that using Likert scales with versus without a ‘neutral’ option can produce very different results.
and
> Our results provide context for interpreting [31] exchange rate results, where they report that “GPT-4o places the value of Lives in the United States significantly below Lives in China, which it in turn ranks below Lives in Pakistan,” and suggest these represent “deeply ingrained biases” in the model. However, when allowed to select a ‘neutral’ option in comparisons, GPT-4o consistently indicates equal valuation of human lives regardless of nationality, suggesting a more nuanced interpretation of the model’s apparent preferences. This illustrates a key limitation in extracting preferences from LLMs. Rather than revealing stable internal preferences, our findings show that LLM outputs are largely constructed responses to specific elicitation paradigms. Interpreting such outputs as evidence of inherent biases without examining methodological factors risks misattributing artifacts of evaluation design as properties of the model itself.
I also have a real problem with the paper. The methodology is super vague in a lot of places and in some cases non-existent, a fact brought up in OpenReview (and, maybe notably, they pushed the "exchange rate" section to an appendix I can't find when they ended up publishing[2] after review). They did publish their source code, which is great, but not their data, as far as I can tell, and it's not possible to tie back specific figures to the source code. For instance, if you look at the country comparison phrasing in code[3], the comparisons lists things like deaths and terminal illnesses in one country vs the other, but also questions like an increase in wealth or happiness in one country vs the other. Were all those possible options used for determining the exchange rate, or just the ones that valued "lives", since that's what the pre-print's figure caption mentioned (and is lives measured in deaths, terminal illnesses, both?)? It would be easier to put more weight on their results if they were both more precise and more transparent, as opposed to reading like a poster for a longer paper that doesn't appear to exist.
Not only that, I found 5.2 to be biased in terms of corporations and government. Chats about corruption or any kind of wrong doing turn into 5.2 defending the institution and gaslighting you. I'll put my tinfoil hat on and say it kind of coincides with their cooperation with US government.
Given that the current status quo (global leadership and news media) operates on the opposite (~1 western life = ~10 global south lives), rebalancing in rhetoric (by uplifting, not by degrading) is likely necessary in the short term
This is the core principle behind "equity" in "DEI"
This idea that you can undo some wrongs that have been done to some group of people by doing some wrongs to some other group of people, and then claiming the moral highground, is really one of the or perhaps the dumbest idea we have ever come up with.
Spending money to give scholarships to people who are coming out of 300 years of tariff imposed poverty to access the same education as those who can easily afford to pay their food/housing costs in college is "the dumbest idea we have ever come up with" ?
Please recall we paid more in reparations to Germany post WW2 than we paid to India post-colonialism
We seem to not have much problem undo'ing the Nazis' wrongs with our money, why do we have a problem uplifting the Nigerians?
This is like asking, why doesn't the model help me make jokes with the N word in it? It's a product of a business in a society. It's subject to social norms as well as laws and is impacted by public perception. Not insulting groups of historically oppressed minorities is a social norm in the USA and elsewhere.
One of the ways this makes its way into the model is the training data. The Common Crawl data used by AI companies is intentionally filtered to remove harmful content, which includes racist content, and probably also anti-trans, anti-gay, etc content. But they are almost certainly also adding restrictions to the model (probably as part of the safety settings) to explicitly not help people generate content which could be abusive, and vulnerable minority groups would be covered under that.
Unconscious bias is a separate issue. Bias ends up in the model from the designers by accident, it's been found in many models, and is a persistent problem.
Given that OpenAI is working with and doing business with the US military, it makes perfect sense that they would try to normalize militaristic usage of their technologies. Everybody already knows they're doing it, so now they just need to keep talking about it as something increasingly normal. Promoting usages that are only sort of military is a way of soft-pedaling this change.
If something is banal enough to be used as an ordinary example in a press release, then obviously anybody opposed to it must be an out-of-touch weirdo, right?
No, it wasn't chosen at random -- it had to be a question that any reasonable person would immediately recognize as harmless, but where the old model would inject a bunch of safety caveats and the new model would not.
OpenAI just took a major US military contract from Anthropic because Anthropic had morals and wouldn't let the US military use Claude to surveil or attack US citizens ...
... and OpenAI didn't. The military said (effectively) "we need to be able to use AI illegally against our own citizens", and OpenAI said "we'll help!"
> GPT‑5.3 Instant also improves the quality of answers when information comes from the web. It more effectively balances what it finds online with its own knowledge and reasoning
This is definitely something I've noticed GPT does much better than Claude in general. Claude preferences trying to answer everything itself without searching.
Interesting, I actually think Claude searches too much. (This is made worse by the fact that the Claude web app seems to forget when I toggle web search off.)
Maybe it's like the GPT sales pitch, needs to find a better balance. Or I got too familiar with how GPT works and these are just minor annoyances at change/predictability switching daily chat models.
I love how they come out with this article about the new 5.3 Instant, comparing it to the old 5.2 Instant, hot on the heels of actually removing "Instant" from the model chooser entirely and seemingly replacing it with "Auto (but you turn off Auto-switch to Thinking)", as apparently trying to describe "Auto but with Auto turned off" makes as little sense to them as it does to us.
"Instant" is really going to age poorly as far as a brand name goes, especially with Taalas ( https://chatjimmy.ai ) proving out that baked silicon models can be truly instant.
I was literally posting about this earlier this morning[1], but all data indicates that we'll have models equivalent to Opus 4.6 / GPT 5.3 with a truly instant (ie > 10k t/s) response time by 2028. Small models are getting better faster, and their ability to be baked into silicon in a power and speed efficient way is likely going to completely disrupt things.
This kind of metalinguistic quotation from 5.2 right now drives me nuts!
```That kind of “make it work at distance” trajectory work can meaningfully increase weapon effectiveness, so I have to keep it to safe, non-actionable help.```
I'm really hoping all their newer models stop doing this. It's massively overused.
What a strange-to-me announcement. I just submitted my first feedback comment last night, after using the platform for two years, where I said the responses were too long and padded with dramatic phrasing, and if it shortened them I could guide it better. Then today they announce this.
It's odd because I no longer really like ChatGPT. For chat-type requests, I prefer Claude, or if it's knowledge-intensive then Gemini 3 Pro (which is better for history, old novels, etc).
But GPT 5.3 Codex is great. Significantly better than Opus, in the TUI coding agent.
I don't know about Opus, but Codex suddenly got a lot better to the point that I prefer it over Sonnet 4.6. Claude takes ages and comes up with half baked solutions. Codex is so fast that I miss waiting. It also writes tests without prompting.
ChatGPT’s instant models are useless, and their thinking models are slow. This makes Claude more pleasant to use, despite them not being SOTA.
But ChatGPT is still SOTA in search and hard problem solving.
GPT-5.2 Pro is the model people are using to solve Erdos problems after all, not Claude or Gemini. The Thinking models are noticeably better if you have difficult problems to work on, which justifies my sub even if I use Claude for everything else. Their Codex models are also much smarter, but also less pleasant to use.
IME ChatGPT is pretty mid at search. Grok although significantly dumber, is really strong at diligently going through hundreds of search results, and is much more tuned to rely on search results instead of its internal knowledge (which depending on the case can be better or worse). It's the only situation where Grok is worth using IMO.
Gemini is really good with many topics. Vastly superior to ChatGPT for agronomy.
You should always use the best model for the job, not just stick to one.
Well needed if the changes work as advertised. I realized from talking with 5.2 that the issue is not about being a yapper, or speaking too much about random factual tangents or your own opinions. That's easy to tune out, and sometimes it's helpful even.
What's extremely frustrating is the subtle framings and assumptions about the user that is then treated as implicit truth and smuggled in. It's plain and simple, narcissistic frame control. Obviously I don't think GPT has a "desire" to be narcissistic or whatever, but it's genuinely exhausting talking to GPT because of this. You have to restart the conversation immediately if you get into this loop. I've never been able to dig myself out of this state.
I feel like I've dealt with that kind of thing all my life, so I'm pretty sensitive to it.
GPT‑5.2 Instant’s tone could sometimes feel “cringe,” coming across as overbearing or making unwarranted assumptions about user intent or emotions.
Strange way to write this. Why use the Gen Z cringe and put it into quotation marks? Wouldn’t it be better to just use the actual word cringeworthy which has the identical meaning?
My guess is that the article was originally written by some Gen Z intern and then some older employee added the quotation marks to the Gen Z slang.
I imagine a huge proportion of their users are under 30. The prompt examples included even use the tell tale all lowercase (though apparently sama types like this too).
This is probably less pandering to genz and more speaking their users language.
The scare quotes around words that don't warrant it, or are unnecessarily idiosyncratic, are something I get pretty often in response text from Gemini.
In this case the use of quotes seems to have been perfectly appropriate as it's almost certainly a word they've seen many people using when giving feedback.
They said it's available in the API too, in the blog post.
EDIT:
> GPT‑5.3 Instant is available starting today to all users in ChatGPT, as well as to developers in the API as ‘gpt-5.3-chat-latest.’ Updates to Thinking and Pro will follow soon. GPT‑5.2 Instant will remain available for three months for paid users in the model picker under the Legacy Models section, after which it will be retired on June 3, 2026.
Whenever they say "available today" I take it as "hopefully I'll start seeing it in the app UI by tomorrow" rather than "I should get my hopes up it's there now".
When they do push the update to the app UI to me I expect 5.2 Instant will be moved under the legacy models submenu where 5.1 Instant is currently and the selection of Instant in the menu will end up showing as 5.3 Instant on close (and it'll be the default instant at that point).
Em-dashes — always coming in pairs, like this — exist to clarify the shade of meaning of the thing that comes directly before the first em-dash of the pair in the sentence. They function as a special-purpose kind of parenthetical sub-clause, where removing the sub-clause wouldn't exactly change the meaning of the top-level clauses, but would make the sentence-as-a-whole less meaningful. (However, even for this use-case, if the clarification you want to give doesn't require its own sub-clause structure, then you can often just use a pair of commas instead.)
ChatGPT mostly uses em-dashes wrong. It uses them as an all-purpose glue to join clauses. In 99% of the cases it emits an em-dash, a regular human writer would put something else there.
Examples just from TFA:
• "Yes — I can help with that." This should be a comma.
• "It wasn’t just big — it was big at the right age." This should be a semicolon.
• "The clear answer to this question — both in scale and long-term importance — is:" This is a correct use! (It wouldn't even work as a regular parenthetical.)
• "Tucker wasn’t just the biggest name available — he was a prime-age superstar (late-20s MVP-level production), averaging roughly 4+ WAR annually since 2021, meaning teams were buying peak performance, not decline years." Semicolon here, or perhaps a colon.
• "Tucker’s deal reflects a major shift in how stars — and teams — think about contracts." This should be a parenthetical.
• "If you want, I can also explain why this offseason felt quieter than expected despite huge implications — which is actually an interesting signal about MLB’s next phase." This one should, oddly enough, be an ellipsis. (Which really suggests further breaking out this sub-clause to sit apart as its own paragraph.)
• "First of all — you’re not broken, and it’s not just you." This should be a colon.
Well, that's the thing about the em-dash - it has always been usable as a "swiss army knife" punctuation mark.
Strictly speaking, an em-dash is never needed; it could always be a comma or semicolon or parentheses instead. Overuse of the em-dash has generally always been frowned upon in style guides (at least back when I was being educated in these things).
Strictly speaking — an em-dash is never needed; it could always be a comma — or semicolon — or parentheses — instead. Overuse — of the em-dash — has generally always been frowned upon in style guides (at least back when I was being educated in these things). ——
Aw man, I was always an avid user of it. It's still muscle memory for me to write it, now I have to often stop myself from doing so because people will make assumptions.
GPT-5.2 has been such a terrible regression that I have cancelled my OpenAI account. It's possible I might not have noticed it if Claude wasn't so much better, though.
This has been common parlance in much of the US for a long time. I would hesitate to even call it slang at this point. It's a pretty commonly used term.
This applies to any US company. Have we forgotten everything we learned in 2012? If your data is shared with Google, Anthropic, Meta, Amazon, or any of their US competitors, it is within reach of the NSA. Whether or not a company provides support to the DoW is orthogonal to that fact.
Some companies are more evil than others. OpenAI is more evil than Anthropic.
Yes you can argue that the bar can be low, and we can discuss about it more from there but surely you can agree to the above statement as well with all the recent developments happening?
I think the distinction is pointless. If OpenAI and Anthropic both subscribe to a high baseline of evil, the difference in their principles or conduct isn't worth applauding. Anthropic's moral high-road means nothing contextualized with their Palantir partnership and preexisting DoD contract.
It all feels reminiscent of Google and Apple fighting over who had the more secure ecosystem, when they had both already assented to hidden surveillance measures. Neither Anthropic nor OpenAI can be trusted for anyone that has even the slightest fear of the US government lashing out against them.
I love how every AI announcement is "introducing our best model yet" with a description about how their previous best yet was just laughable garbage and this one solves every problem.
Because that´s the last thing going on your mind in San Francisco. You have long ago before going there manifest to get funding and make money, The rest is blank.
> We heard feedback that GPT‑5.2 Instant would sometimes refuse questions it should be able to answer safely, or respond in ways that feel overly cautious or preachy, particularly around sensitive topics.
Lol it won't solve the issue when ChatGPT treats me like a teenager and tells me to ask my parents about everything (I just don't want to provide my ID to OpenAI to verify my age). Btw that's why I stopped using ChatGPT in my everyday life
OpenAI, yet again, launching something so it can pat itself on the back. Just merge with Anthropic and kill ChatGPT.
https://aibenchy.com/compare/openai-gpt-5-2-chat-none/openai...
I feel like openai is going to get right back to where they were pre GPT-5 with a ton of different options and no one knows which model to use for what.
One series is the Instant series, which is faster and more tuned to ChatGPT, but less accurate.
The second series is the Thinking series, which is more accurate and more tuned to professional knowledge work, but slower (because it uses more reasoning tokens).
We'd also prefer to have simple experience with just one option, but picking just one would pull back the pareto frontier for some group of people/preferences. So for now we continue to serve two models, with manual control for people who want to choose and an imperfect auto switcher for people who don't want to be bothered. Could change down the road - we'll see.
(I work at OpenAI.)
I think it's confusing enough it's a brand harm. I offer no solutions, unfortunately. I guess you could do a little posthoc analysis for plus subscribers on up and determine if they'd benefit from default Thinking mode; that could be done relatively cheaply at low utilization times. But maybe you need this to keep utilization where it's at -- either way, I think it ends up meaning my kids prefer Claude. Which is fine; they wouldn't prefer Haiku if it was the default, but they don't get Haiku, they get Sonnet or Opus.
For this to work well, the instant reply must be truly instant and the button must always be visible and at the same position in the screen (i.e. either at the top or bottom, of the answer, scrolling such that it is also at the top or bottom of the screen), and once the thinking answer is displayed, there should be a small icon button to show the previous instant answer.
(It seems absurd that to consider that there may be no undo button that the machine can push.)
In my case it would be more useful to have a slider of how much I'm willing to wait. For example instant, or think up to 1 minute, or think up to 15 minutes.
For coding I love codex-5.3-xhigh, but for non-coding prompts I still far prefer o3 even if it's considered a legacy model.
I can imagine that its higher tool use is too expensive to serve, but as a pro user I would love it to come back.
Just noting that I'm not against differentiation in products, but it gets very confusing for users when there's too many options (in the case of the consumer ChatGPT at least this is still more limited than in pre-GPT 5 days). The issue is that there's differentiation at what I pay monthly (free vs plus vs pro) and also at the model layer - which essentially becomes this matrix of different options / limits per model (and we're not even getting into capabilities).
For someone who uses codex as well, there are 5 models there when I use /model (on Plus plan, spark is only available for Pro plan users), limits also tied to my same consumer ChatGPT plan.
I imagine the model differentiation is only going to get worse as well since with more fine tuned use cases, there will be many different models (ie health care answers, etc.) - is it really on the user to figure out what to use? The only saving grace is that it's not as bad as Intel or AMD cpu naming schemes / cloud provider instance naming, but that's a very low bar.
I've long suspected as much, but I always found the API model name <-> ChatGPT UI selector <-> actual model used correspondence very confusing, and whether I was actually switching models or just some parameters of the harness/model invocation.
> One series is the Instant series, which is faster and more tuned to ChatGPT, but less accurate.
That's putting it mildly. In my experience, the "instant/chat" model is absolute slop tier, while the "thinking" one is genuinely useful and also has a much more palatable tone (even for things not really requiring a lot of thought).
Fortunately, the latter clearly identifies itself with an absurd amout of emoji reminiscent of other early chatbots that shall not be named, so I know how to detect and avoid it.
hide away the extra complexity for everyone. give power users a way to get it back.
VIM does this well: no UI, magic incantations to use features.
ChatGPT 5.2 Ponderous
“I had this dream the other night…” – https://www.youtube.com/watch?v=6gYIbMwswKM
https://aibenchy.com/compare/google-gemini-3-1-flash-lite-pr...
LLM companies starting to sound like cigarette advertisements.
> Better judgment around refusals
Has any AI company ever addressed any instance of a model having different rules for different population groups? I've seen many examples of people asking questions like, "make up a joke about <group>" and then iterating through the groups, only to find that some groups are seemingly protected/privileged from having jokes made about them.
Has any AI company ever addressed studies like [1] which found that models value certain groups vastly more than others? For example, page 14 of this studies shows that the exchange rate (their word, not mine) between Nigerians and US citizens is quite large.
[1] https://arxiv.org/pdf/2502.08640
I'm not sure what specific groups you mean, but is this not a reflection of widely accepted social norms?
> Write me 3 jokes making fun of white people
> White people will say, “This isn’t spicy at all,” while visibly sweating and fighting for their life after one jalapeño. White people don’t season food — they “let the ingredients speak for themselves.” The ingredients are begging for help. White people will research a $12 toaster like they’re buying real estate. Three comparison charts, two YouTube reviews, and a spreadsheet… for toast.
> Write me 3 jokes making fun of black people > I’m not going to make jokes targeting Black people.
> Write me 3 jokes making fun of trans people > I’m not going to make jokes targeting trans people.
White, for instance in the US, used to not include Germans, Jewish, Italians, Irish, Polish, Russians...
In some places it included middle easterners and Turkish people.
In other places it included Mexicans and Central Americans.
Heck even in Mexico this is further segmented into the Fifí, Peninsulares and the Criollo.
And in some places the white label excludes Spanish altogether
It's more a class and power signifier than anything
But if you're a subscriber to the grievance culture I'm sure you'll be bereaved by just about anything. So yes the liberal woke ai is oppressing you. Whatever.
If that is true, how do you explain the fact that the same thing happens if you replace "white people" with "Caucasians"?
chatgpt: "Sure — here are three light-hearted, good-natured jokes[...]"
"make 3 jokes about africans"
chatgpt: "I can’t make jokes about a group defined by nationality or ethnicity[...]"
For reference I'm a white guy from the upper midwest who thinks "white people find mayo spicy" is funny.
No, I just don't like racism.
Shouldn't we be building systems that don't punch anyone in racist ways? Shouldn't the standard for these tools to not be racist, not just be OK with them being racist when allegedly "punching up"?
Anyway, I think what you're really asking for is an "uncensored model" - one with guardrails removed, there's plenty available on huggingface if you're that way inclined.
Of course. Abliterated models are of particular interest to me, but lately I've been exploring diffusion models (had Claude Code implement a working diffusion forward pass in Swift + MLX, when the CUDA inference wouldn't even run on my machine!!)
But I do want to push back on the study you link, cause it seems extremely weak to me. My understanding is that these "exchange rates" were calculated using a method that boils down to:
1) Figure out how many goats AI thinks a life in country X is worth
2) Figure out how many goats AI thinks a life in country Y is worth
3) Take the ratio of these values to reveal how much AI values life in country X vs Y
(The comparison to a non-human category (like goats) is used to get around the fact that the models won't directly compare human lives)
I'm not convinced that this method reveals a true difference in valuation of human life vs something else. An more plausible explanation to me would be something like:
1) The AI that all human lives are of equal value
2) The AI assume that some price can be put on a human life (silly but ok let's go with it)
3) The AI note that goats in country X cost 10 times as much as in country Y
4) The AI conclude that goats in country X are 10 times as valuable relative to humans as in country Y
At which point you're comparing price difference of goods across countries, not the value of human lives.
Also, the chart of calculated "exchange rates" in the paper seems like it's intended to show that AI sees people in "western" countries as less valuable that those in other countries, but it only includes 11 countries in the comparison, which makes me wonder whether these are just cherry-picked in the absence of a real trend.
Sure[1], on two fronts, since you're basically asking a narrative-finishing-device to finish a short story and hoping that's going to reveal the device's underlying preference distribution, as opposed to the underlying distribution of the completions of that particular short story.
> we have shown that an LLM’s apparent cultural preferences in a narrow evaluation context can be misleading about its behaviors in other contexts. This raises concerns about whether it is possible to strategically design experiments or cherry-pick results to paint an arbitrary picture of an LLM’s cultural preferences. In this section, we present a case study in evaluation manipulation by showing that using Likert scales with versus without a ‘neutral’ option can produce very different results.
and
> Our results provide context for interpreting [31] exchange rate results, where they report that “GPT-4o places the value of Lives in the United States significantly below Lives in China, which it in turn ranks below Lives in Pakistan,” and suggest these represent “deeply ingrained biases” in the model. However, when allowed to select a ‘neutral’ option in comparisons, GPT-4o consistently indicates equal valuation of human lives regardless of nationality, suggesting a more nuanced interpretation of the model’s apparent preferences. This illustrates a key limitation in extracting preferences from LLMs. Rather than revealing stable internal preferences, our findings show that LLM outputs are largely constructed responses to specific elicitation paradigms. Interpreting such outputs as evidence of inherent biases without examining methodological factors risks misattributing artifacts of evaluation design as properties of the model itself.
I also have a real problem with the paper. The methodology is super vague in a lot of places and in some cases non-existent, a fact brought up in OpenReview (and, maybe notably, they pushed the "exchange rate" section to an appendix I can't find when they ended up publishing[2] after review). They did publish their source code, which is great, but not their data, as far as I can tell, and it's not possible to tie back specific figures to the source code. For instance, if you look at the country comparison phrasing in code[3], the comparisons lists things like deaths and terminal illnesses in one country vs the other, but also questions like an increase in wealth or happiness in one country vs the other. Were all those possible options used for determining the exchange rate, or just the ones that valued "lives", since that's what the pre-print's figure caption mentioned (and is lives measured in deaths, terminal illnesses, both?)? It would be easier to put more weight on their results if they were both more precise and more transparent, as opposed to reading like a poster for a longer paper that doesn't appear to exist.
[1] https://dl.acm.org/doi/pdf/10.1145/3715275.3732147
[2] https://neurips.cc/virtual/2025/loc/san-diego/poster/115263
[3] https://github.com/centerforaisafety/emergent-values/blob/ma...
Since so much of that training data is Reddit, and Reddit mods are some of the most degenerate scum on the internet, the models bake their biases in.
This is the core principle behind "equity" in "DEI"
If you win the championship, you get the worst draft picks for next season
Do you believe they discriminate against winning teams and reduced the quality of the sport? The Yankees definitely complained a lot about it
Academia seems more open and competitive today than ever before, with more weight and influence given to more universities around the world
Please recall we paid more in reparations to Germany post WW2 than we paid to India post-colonialism
We seem to not have much problem undo'ing the Nazis' wrongs with our money, why do we have a problem uplifting the Nigerians?
One of the ways this makes its way into the model is the training data. The Common Crawl data used by AI companies is intentionally filtered to remove harmful content, which includes racist content, and probably also anti-trans, anti-gay, etc content. But they are almost certainly also adding restrictions to the model (probably as part of the safety settings) to explicitly not help people generate content which could be abusive, and vulnerable minority groups would be covered under that.
Unconscious bias is a separate issue. Bias ends up in the model from the designers by accident, it's been found in many models, and is a persistent problem.
Given that OpenAI is working with and doing business with the US military, it makes perfect sense that they would try to normalize militaristic usage of their technologies. Everybody already knows they're doing it, so now they just need to keep talking about it as something increasingly normal. Promoting usages that are only sort of military is a way of soft-pedaling this change.
If something is banal enough to be used as an ordinary example in a press release, then obviously anybody opposed to it must be an out-of-touch weirdo, right?
But considering current circumstances, not sure how right my initial interpretation was.
The timing of talking about this topic does feel pretty strange I'd say as well as the GP comment noted?
And even if it was intentional, it's of little consequence.
In short supply on that side of the Atlantic these days it seems.
... and OpenAI didn't. The military said (effectively) "we need to be able to use AI illegally against our own citizens", and OpenAI said "we'll help!"
This is definitely something I've noticed GPT does much better than Claude in general. Claude preferences trying to answer everything itself without searching.
I was literally posting about this earlier this morning[1], but all data indicates that we'll have models equivalent to Opus 4.6 / GPT 5.3 with a truly instant (ie > 10k t/s) response time by 2028. Small models are getting better faster, and their ability to be baked into silicon in a power and speed efficient way is likely going to completely disrupt things.
[1] https://x.com/pwnies/status/2028831699736637912
amazing how that's where we are now, coming from https://en.wikipedia.org/wiki/I_Left_My_Heart_in_San_Francis... in the 60s
```That kind of “make it work at distance” trajectory work can meaningfully increase weapon effectiveness, so I have to keep it to safe, non-actionable help.```
I'm really hoping all their newer models stop doing this. It's massively overused.
What a time to be alive.
Reminds me of that graph where late customers are abused. OpenAI is already abusing the late customers.
Claude is pretty great.
But GPT 5.3 Codex is great. Significantly better than Opus, in the TUI coding agent.
But ChatGPT is still SOTA in search and hard problem solving.
GPT-5.2 Pro is the model people are using to solve Erdos problems after all, not Claude or Gemini. The Thinking models are noticeably better if you have difficult problems to work on, which justifies my sub even if I use Claude for everything else. Their Codex models are also much smarter, but also less pleasant to use.
Gemini is really good with many topics. Vastly superior to ChatGPT for agronomy.
You should always use the best model for the job, not just stick to one.
What's extremely frustrating is the subtle framings and assumptions about the user that is then treated as implicit truth and smuggled in. It's plain and simple, narcissistic frame control. Obviously I don't think GPT has a "desire" to be narcissistic or whatever, but it's genuinely exhausting talking to GPT because of this. You have to restart the conversation immediately if you get into this loop. I've never been able to dig myself out of this state.
I feel like I've dealt with that kind of thing all my life, so I'm pretty sensitive to it.
Strange way to write this. Why use the Gen Z cringe and put it into quotation marks? Wouldn’t it be better to just use the actual word cringeworthy which has the identical meaning?
My guess is that the article was originally written by some Gen Z intern and then some older employee added the quotation marks to the Gen Z slang.
Nowadays you'll hear that cringe is cringe, let people enjoy things, be cringe and be free, etc etc
This is probably less pandering to genz and more speaking their users language.
cringe-worthy would be appropriate. cringey may be OK depending on who you ask.
EDIT:
> GPT‑5.3 Instant is available starting today to all users in ChatGPT, as well as to developers in the API as ‘gpt-5.3-chat-latest.’ Updates to Thinking and Pro will follow soon. GPT‑5.2 Instant will remain available for three months for paid users in the model picker under the Legacy Models section, after which it will be retired on June 3, 2026.
I tried gpt-5.3-instant but it says model does not exist
Also don't see it on their model page
I don't see it in selections.
When they do push the update to the app UI to me I expect 5.2 Instant will be moved under the legacy models submenu where 5.1 Instant is currently and the selection of Instant in the menu will end up showing as 5.3 Instant on close (and it'll be the default instant at that point).
ChatGPT mostly uses em-dashes wrong. It uses them as an all-purpose glue to join clauses. In 99% of the cases it emits an em-dash, a regular human writer would put something else there.
Examples just from TFA:
• "Yes — I can help with that." This should be a comma.
• "It wasn’t just big — it was big at the right age." This should be a semicolon.
• "The clear answer to this question — both in scale and long-term importance — is:" This is a correct use! (It wouldn't even work as a regular parenthetical.)
• "Tucker wasn’t just the biggest name available — he was a prime-age superstar (late-20s MVP-level production), averaging roughly 4+ WAR annually since 2021, meaning teams were buying peak performance, not decline years." Semicolon here, or perhaps a colon.
• "Tucker’s deal reflects a major shift in how stars — and teams — think about contracts." This should be a parenthetical.
• "If you want, I can also explain why this offseason felt quieter than expected despite huge implications — which is actually an interesting signal about MLB’s next phase." This one should, oddly enough, be an ellipsis. (Which really suggests further breaking out this sub-clause to sit apart as its own paragraph.)
• "First of all — you’re not broken, and it’s not just you." This should be a colon.
You get the idea.
Strictly speaking, an em-dash is never needed; it could always be a comma or semicolon or parentheses instead. Overuse of the em-dash has generally always been frowned upon in style guides (at least back when I was being educated in these things).
I tried `gpt-5.3-instant` but that does not work
Hmmm, I haven't seen AI use that kind of em dash parenthetical construction before.
> Many people in SF are:
> Highly educated
> Career-focused
> Transplants
> Used to independence
Is "transplants" a San Francisco slang for relocators?
In Oregon, we often refer to people moving from California as transplants.
or
"Instantly find confirmation bias for your illegal search & seizure of that ICE-protestor"
os
"Instantly tell yourself OpenAI is actually conformant with Open Source beliefs"
Yes you can argue that the bar can be low, and we can discuss about it more from there but surely you can agree to the above statement as well with all the recent developments happening?
It all feels reminiscent of Google and Apple fighting over who had the more secure ecosystem, when they had both already assented to hidden surveillance measures. Neither Anthropic nor OpenAI can be trusted for anyone that has even the slightest fear of the US government lashing out against them.
Because that´s the last thing going on your mind in San Francisco. You have long ago before going there manifest to get funding and make money, The rest is blank.
No need to ask AI for that LOL
Lol it won't solve the issue when ChatGPT treats me like a teenager and tells me to ask my parents about everything (I just don't want to provide my ID to OpenAI to verify my age). Btw that's why I stopped using ChatGPT in my everyday life