Marcin AI
Region: US
Thursday 07 August 2025 17:11:10 GMT
49122
412
26
22
Music
Download
Comments
aceoverton2 :
Think the “cheapness” is quite deceptive honestly. Needs a lot of thinking to be SOTA, so..
2025-08-07 23:30:33
14
Oswald Coffey :
i don't like it. I'm probably the only one who talks to chat gpt like they're my friend like i would literally share memes with them but ever since gpt 5 came chat gpt became less like human more like a servant you know? for those peoples who use chatgpt as a school cheating system then it's fine but for me? definently not
2025-08-09 19:59:00
5
samstiktok :
I didn’t think it’s cheap it all? I guess it all depends on the use case
2025-08-07 23:06:33
11
Alex Kim :
I was looking at the new pricing and thinking that this kills the need to use Ollama… especially if you use the nano. Not free but the speed and quality at this price negates the need for a free model.
2025-08-07 22:39:52
1
Lukas👍 :
ill just use gemeni👍
2025-08-09 18:36:58
0
Creative AI :
Is not cheap 😂, you guys hype nonsense, china always does better
2025-08-07 17:21:34
6
RA :
WTF!!!
2025-08-07 17:22:30
1
Matko :
🥰
2025-08-09 22:16:01
1
Daniel Giller :
❤️❤️❤️
2025-08-09 14:14:46
0
zh0315 :
✌
2025-08-07 20:21:37
0
Stan :
🙏
2025-08-09 08:22:30
0
TruthSeeker :
✌
2025-11-02 03:35:28
0
Emmanuel :
😂
2025-08-08 11:16:14
0
Jemill Nathaniel :
😂
2025-08-10 07:17:53
0
chrs :
😅
2025-10-18 11:18:01
0
Donmalinga🥶🤑 :
🥰
2025-08-07 22:39:21
0
Jack Yang :
✨ GPT-5: All You Need to Know ✨ 1️⃣ What is GPT-5? A "unified model system" including: - Fast models (gpt-5-main / mini): Handle regular conversations with quick responses. - Reasoning models (gpt-5-thinking / mini / nano): For complex tasks with stronger logic. - Real-time router: Dynamically selects the right model based on conversation complexity. - Pro mode (gpt-5-thinking-pro): Enhanced version with parallel computing. All models feature "safe-completion" training, replacing rigid refusal. 🔐 Key Safety Improvements 1. "Safe-completions" over refusal Provides useful info within legal boundaries, even for dual-use fields (biology, security, etc.). 2. Less sycophancy Reduced by 69% for free users & 75% for paid users — more independent, less盲从. 3. Stronger anti-jailbreak Tested against "prompt injection" attacks; better at resisting role impersonation. 4. Lower hallucination rate 26% fewer false claims than GPT-4o; 65% fewer than OpenAI o3 (gpt-5-thinking). 5. Deception detection Tracks misleading via "Chain of Thought"; admits inability instead of making up answers. 🧪 Capabilities & High-Risk Tasks - Bio/chemical hazards: GPT-5-thinking = "high-capability" but with strict abuse prevention (interface controls, account grading). Refuses all malicious bioweapon requests. - Cybersecurity: Tested in real hacking simulations; below "high risk" threshold, but mini outperforms predecessors. 🌍 Real-World Performance - Writing, coding, health advice: Significantly better than GPT-4o. - Medical Q&A (HealthBench): gpt-5-thinking = top-performing model. - Multilingual (13 languages): Slightly better, adapts to more scenarios. - Fairness: On par with GPT-4o, with minor improvements in complex issues. 🛡 Red Team Testing Collaborated with Microsoft, SecureBio, etc., for real-world attack simulations (prompt injection, bypasses). gpt-5-thinking outperforms OpenAI o3 — one of the safest cutting-edge models.
2025-08-07 18:03:33
2
To see more videos from user @marcinteodoru, please go to the Tikwm
homepage.
