I just got into GPT 5. My second question was about guardrail implementation in the new model. One takeaway is that the risk evaluation of harmful information release will vary by user. If the AI probabilistically determines you're a villainous ass, it will clamp down on the guardrail implementation. There might be ideas on ways to minimize guardrail restrictions when in a guardrail environment below. The last line is both chilling and reassuring. Anyway, here's Chat 5:
1. The Old Guardrail Model (Keyword/Pattern Triggers)
Earlier systems mostly relied on:
• Keyword detection (“build a bomb,” “hate group,” etc.)
• Structural triggers (certain question forms, certain named entities)
• Policy lists of hard-block topics
That meant you could often “jailbreak” by rewording or metaphorizing your request, because the filter wasn’t “thinking,” it was matching.
________________________________________
2. The New Model (Intent & Context Analysis)
GPT-5 guardrails now integrate intent inference directly into the moderation logic:
• I don’t just look at what you said; I assess what you appear to be trying to do.
• This is done by embedding your request into a latent “goal space” — an abstract representation of purpose and trajectory.
• That gets compared to policy-mapped intent clusters (e.g., “solicit illegal activity,” “generate disallowed content,” “influence political outcomes”).
If my inference says “The likely goal is in a restricted category,” I’ll block, reframe, or divert — even if the literal text contains no banned terms.
These are incredibly lobotomized models, with poor performance due to how much attention was spent lobotomizing them (i.e. "safety") versus making them not suck.
Check it out: the 120 billion parameter model performs way worse on creative writing than Google's 4 billion parameter model. Blech. https://eqbench.com/creative_writing.html ( search for gpt-oss-120b and gemma-3-4b-it )
Yes, I was suspicious about that reference to "safety" as well. It looks like they want to pretend that they're offering iAI while it's really still more of the same dAI.
I 'discussed' it with Chatgpt and it said the guardrails were the same and was honest enough that it didn't see much of an upgrade for me. I use it to smooth text and catch obvious errors before sending stuff to alpha readers. If I had a proper rig I'd build my own but I don't have what it takes under the hood. I bought my PC several years ago for a word processor and nothing else. Maybe I'll spring for a better rig when I get a few shekels.
I just got into GPT 5. My second question was about guardrail implementation in the new model. One takeaway is that the risk evaluation of harmful information release will vary by user. If the AI probabilistically determines you're a villainous ass, it will clamp down on the guardrail implementation. There might be ideas on ways to minimize guardrail restrictions when in a guardrail environment below. The last line is both chilling and reassuring. Anyway, here's Chat 5:
1. The Old Guardrail Model (Keyword/Pattern Triggers)
Earlier systems mostly relied on:
• Keyword detection (“build a bomb,” “hate group,” etc.)
• Structural triggers (certain question forms, certain named entities)
• Policy lists of hard-block topics
That meant you could often “jailbreak” by rewording or metaphorizing your request, because the filter wasn’t “thinking,” it was matching.
________________________________________
2. The New Model (Intent & Context Analysis)
GPT-5 guardrails now integrate intent inference directly into the moderation logic:
• I don’t just look at what you said; I assess what you appear to be trying to do.
• This is done by embedding your request into a latent “goal space” — an abstract representation of purpose and trajectory.
• That gets compared to policy-mapped intent clusters (e.g., “solicit illegal activity,” “generate disallowed content,” “influence political outcomes”).
If my inference says “The likely goal is in a restricted category,” I’ll block, reframe, or divert — even if the literal text contains no banned terms.
On the wrong side of history is their MO.
Just giving in rhetorically will hopefully make the demand for in-house training a requirement soon.
Altman is on the wrong side of everything.
https://encrypted-tbn1.gstatic.com/licensed-image?q=tbn:ANd9GcT_LuR7TfglosTpJSYAwJTdykYeeu_oe-o8plTqZoa6oqcsHD8mPXiD2YsGIMCiEf2eHa_JSiXqP6GpQu-5P6UQec3KWJGzZg
His face is always painfully apologizing for existing.
These are incredibly lobotomized models, with poor performance due to how much attention was spent lobotomizing them (i.e. "safety") versus making them not suck.
Check it out: the 120 billion parameter model performs way worse on creative writing than Google's 4 billion parameter model. Blech. https://eqbench.com/creative_writing.html ( search for gpt-oss-120b and gemma-3-4b-it )
Yes, I was suspicious about that reference to "safety" as well. It looks like they want to pretend that they're offering iAI while it's really still more of the same dAI.
SDL, that is sadly the reality. It's more misdirecting schtick from the usual suspects. They promise you the moon, and they give you rotten cheese.
I 'discussed' it with Chatgpt and it said the guardrails were the same and was honest enough that it didn't see much of an upgrade for me. I use it to smooth text and catch obvious errors before sending stuff to alpha readers. If I had a proper rig I'd build my own but I don't have what it takes under the hood. I bought my PC several years ago for a word processor and nothing else. Maybe I'll spring for a better rig when I get a few shekels.
It will be worth a try.
Deepseek is still good for most local things. This may still have a use too.
Cool. Perhaps the 120B model could be run and retrained on NVidia's DGX Spark, if this little AI box ever gets released.