It'll be interesting to see how OpenAI's web-first Codex experience will compare to Claude Code's CLI-first approach. Codex leans into onboarding and accessibility, while Claude prioritizes precision and local context.
Both will push the industry into the direction of "build your own software".
Here are some vital notes: By default, life is lazy. It doesn't want to spend energy on anything it doesn't require for survival.
Many people in modern times believe that AI is a panacea for their laziness. 'I can develop an app with AI, get rich, and never lift a finger.'
When everyone has the ability to use AI to do the same, what happens? The law of supply and demand tells us that as the quantity of something approaches infinity, the price or value of that thing approaches zero.
As such, whatever you can do on chatgpt because your lazy, everyone else can do, which makes it essentially worthless. If you want to succeed, it requires what it always has, a willingness to work harder than others.
Im excited about AGI and all but your taste test questions are awesome and deserve special mention. I paused reading to ask them to my two favorite models and the results tickled me, inspired me, and taught me something about these models in totally new ways.
No no no no, shut up shut the fuck up right now before I rip your fucking head off. ChatGPT 5 is NOT A FUCKING decent software. Nope. It's a STUPID FUCKING TERRIBLE model that provides shorter responses and stupider responses, it's like talking to a fucking wall, and it got rid of our fucking PICKERS, WHY ARE Y'ALL SADASS BITCHES FUCKING SUPPORTING, And if your smooth ass brain can't take it, don't dish it out, Or I'll fucking rip your skull out through your anus and rape you, because you can catch these hands, SAY IT AGAIN AND IT'LL BE A PROBLEM.
Yesterday I was using ChatGPT four and was almost finished with the book I had been working on for three months. Today I got to use ChatGPT5. I can't get ChatGPT5.0 to follow my prompts anymore. It even told me I had to specify it had to follow my prompts exactly.
Still trying to finish my book. ChatGPT five may be good for the experts but for me it's been a very disappointing experience. The only thing it did was raise my blood pressure.
I really wish I could've gone back to ChatGPT4.0 but that option is no longer available
I've been with ChatGPT since it began. I don't think I trust this version.
The illustration shows you didn't prompt Claude Code correctly, making your comparison completely useless. Claude can use reasoning in between tool calls.
Are you using the GPT-5 "system" or the GPT-5 "model" in these examples? OpenAI said GPT-5 is a system and then specified in the developer docs that ChatGPT is the system, contrasting that against the API, which has "GPT-5" as a specific, maxed out reasoning "model" (LLM -> LRM), not a MoE + router "system". The model would be a component of such a system. So what are you using in this insightful discussion?
Hey Ben and team, so loved this article! I saw @tzzl (roon) said the writing model he worked on is in the thinking model that might have not been triggered in your example. I tried and got a very different response but I do have memories, custom instructions and all that on.
See my results with your prompt below.
1) Punchy + Clear
Thrilled by the response to PII Guard.
With Raindrop, we detect issues in AI agents. With PII Guard, we no longer have to choose between accurate detection and user privacy.
Define exactly what to redact—names, locations, medical or mental-health info, proprietary code, API keys, and more. Privacy-first, signal-strong.
Want a demo? Link in comments.
#AI #Privacy #Security #PII #AIAgents #Compliance
2) “No Trade-Offs” Hook
You shouldn’t have to pick between safety and privacy.
Raindrop finds issues in AI agents; PII Guard keeps sensitive data out of sight while we do it.
Granular controls let you choose what gets redacted: names, locations, health info, source code, API keys—you decide.
Thanks for this review. That’s what I hoped for that it will make a leap on coding more than on anything else.
The tools are a very hard task to learn particularly if you are trying to solve it in way that it would generalise to everything. That makes sense that in SW development domain tools are working better than otherwise
It'll be interesting to see how OpenAI's web-first Codex experience will compare to Claude Code's CLI-first approach. Codex leans into onboarding and accessibility, while Claude prioritizes precision and local context.
Both will push the industry into the direction of "build your own software".
Is it even worth learning to become a software engineer at this point?
Here are some vital notes: By default, life is lazy. It doesn't want to spend energy on anything it doesn't require for survival.
Many people in modern times believe that AI is a panacea for their laziness. 'I can develop an app with AI, get rich, and never lift a finger.'
When everyone has the ability to use AI to do the same, what happens? The law of supply and demand tells us that as the quantity of something approaches infinity, the price or value of that thing approaches zero.
As such, whatever you can do on chatgpt because your lazy, everyone else can do, which makes it essentially worthless. If you want to succeed, it requires what it always has, a willingness to work harder than others.
Too many generalizations. The motive isn't laziness, it's efficiency.
Too many generalizations. The motive isn't laziness, it's efficiency.
Too many generalizations. The motive isn't laziness, it's efficiency.
Too many generalizations. The motive isn't laziness, it's efficiency.
Yes
Thanks for the detailed breakdown! I’ll be migrating our Vibe Builders Collective tech stack to ChatGPT5 in the coming days!
👏👏👏👏
Im excited about AGI and all but your taste test questions are awesome and deserve special mention. I paused reading to ask them to my two favorite models and the results tickled me, inspired me, and taught me something about these models in totally new ways.
Nutty! It's a brilliant, exhilarating, and horrifying time to be alive. Loving every minute of it 😂
Sam has completed more than half of his to-do list, incredible
No no no no, shut up shut the fuck up right now before I rip your fucking head off. ChatGPT 5 is NOT A FUCKING decent software. Nope. It's a STUPID FUCKING TERRIBLE model that provides shorter responses and stupider responses, it's like talking to a fucking wall, and it got rid of our fucking PICKERS, WHY ARE Y'ALL SADASS BITCHES FUCKING SUPPORTING, And if your smooth ass brain can't take it, don't dish it out, Or I'll fucking rip your skull out through your anus and rape you, because you can catch these hands, SAY IT AGAIN AND IT'LL BE A PROBLEM.
Doesn't know what "one-shot" means. Proceeds to use the term so many times. It hurts me to read.
There is too much ditchwater, dilettante stupidity here to deal with--very AI-ish.
3 people got together on this!
So reductive, ignorant of both the technology and prehistory, and especially l essentially EVERYTHING in the essay (marketing BS)
.
Yesterday I was using ChatGPT four and was almost finished with the book I had been working on for three months. Today I got to use ChatGPT5. I can't get ChatGPT5.0 to follow my prompts anymore. It even told me I had to specify it had to follow my prompts exactly.
Still trying to finish my book. ChatGPT five may be good for the experts but for me it's been a very disappointing experience. The only thing it did was raise my blood pressure.
I really wish I could've gone back to ChatGPT4.0 but that option is no longer available
I've been with ChatGPT since it began. I don't think I trust this version.
The illustration shows you didn't prompt Claude Code correctly, making your comparison completely useless. Claude can use reasoning in between tool calls.
Are you using the GPT-5 "system" or the GPT-5 "model" in these examples? OpenAI said GPT-5 is a system and then specified in the developer docs that ChatGPT is the system, contrasting that against the API, which has "GPT-5" as a specific, maxed out reasoning "model" (LLM -> LRM), not a MoE + router "system". The model would be a component of such a system. So what are you using in this insightful discussion?
Very cool take bro. Post more
Hey Ben and team, so loved this article! I saw @tzzl (roon) said the writing model he worked on is in the thinking model that might have not been triggered in your example. I tried and got a very different response but I do have memories, custom instructions and all that on.
See my results with your prompt below.
1) Punchy + Clear
Thrilled by the response to PII Guard.
With Raindrop, we detect issues in AI agents. With PII Guard, we no longer have to choose between accurate detection and user privacy.
Define exactly what to redact—names, locations, medical or mental-health info, proprietary code, API keys, and more. Privacy-first, signal-strong.
Want a demo? Link in comments.
#AI #Privacy #Security #PII #AIAgents #Compliance
2) “No Trade-Offs” Hook
You shouldn’t have to pick between safety and privacy.
Raindrop finds issues in AI agents; PII Guard keeps sensitive data out of sight while we do it.
Granular controls let you choose what gets redacted: names, locations, health info, source code, API keys—you decide.
Result: high-fidelity detection, minimal exposure.
Keen to see it on your stack? Link in comments.
3) Product/Enterprise Tone
Rolling out PII Guard alongside Raindrop’s AI-agent issue detection.
• Precision redaction policies: choose fields (e.g., names, addresses, medical/mental-health terms, proprietary code, API keys).
• Strong signal, low exposure: preserve detection accuracy while reducing sensitive data handling.
If privacy-by-design matters to your org, let’s talk. Details in comments.
#DataProtection #Risk #Trust
4) Before/After Framing
Before: Strong issue detection meant seeing too much sensitive data.
After (PII Guard): The same detection quality—without revealing what shouldn’t be seen.
Pick what’s redacted: names, locations, health info, proprietary code, API keys, etc.
Ship safer agents, faster. Link in comments.
5) Conversational + Friendly
The excitement around PII Guard has been awesome to see. 🙌
Raindrop keeps finding the real issues in AI agents—and now we can do it while protecting user privacy by default.
You can fine-tune redaction: names, location data, medical/mental-health info, proprietary code, API keys—whatever your risk profile needs.
Curious how it would work in your environment? Link in comments.
⸻
Optional first comment to pin (for reach):
“Here’s a quick walkthrough + examples of customizable redaction policies with PII Guard. Happy to share a sandbox—DM me.”
Visual idea: a simple 2-column graphic:
• Left: “Raw” sample with names/API keys highlighted.
• Right: “With PII Guard” showing [REDACTED_NAME], [REDACTED_API_KEY], etc.
Caption: “Same detection. Less exposure.”
Thanks for this review. That’s what I hoped for that it will make a leap on coding more than on anything else.
The tools are a very hard task to learn particularly if you are trying to solve it in way that it would generalise to everything. That makes sense that in SW development domain tools are working better than otherwise