DeepSeek v4 Flash to be tested on a prompt ChatGPT refused for moral reasons
yacineMTB · x · 2026-08-04
The author says they’re going to test DeepSeek v4 Flash on a task that ChatGPT refused on moral grounds.
It’s a small but concrete signal about model behavior differences, especially around safety refusals and what kinds of prompts each model will answer.
More from Models
- Kimi K3 turns 6,522 lines of canvas code into a jellyfish-like demo — MaziyarPanahi · 2026-08-04
- A joke sycophancy test exposes how easily LLMs praise almost any framing — charliermarsh · 2026-08-04
- DeepSeek opens V4 Flash beta after post-training boosts agent benchmark results — VraserX · 2026-08-04
- DeepSeek handled the prompt well enough to make one user consider $500K in GPUs — yacineMTB · 2026-08-04
- Gemma 4 26B runs locally on a 13-year-old Xeon CPU with zero GPUs — davidad · 2026-08-04
- Hugging Face weight dump shows a math SFT with AIME 2025 gains at 16k context — kastnerkyle · 2026-08-04