真心講, 大家覺得deepseek 點

用嚟寫matlab code好好用

DeepSeek我淨係知道佢成日都話伺服器繁忙 . 同埋唔知係咪免費仔同收費仔分別啦 , Grok同Copilot 都已經可以做到圖片分析 , DS幾時先可以做到

They obviously changed from using ChatGPT synthetic data to Gemini synthetic data.
The reason is simple — o3 is too expensive and Gemini 2.5 is more reasonable.
But this over reliance in synthetic data for pre training
could be the major cause of its dramatically high hallucination rates (over 10%, still)

白痴仔
gemma 3n同llama畀你無視咗？

google同fb對上一次幾時放sota open source model？

gemini賺到笑，gemma放黎放去最大得27b，用下腦都知點解
facebook llama4 dead on arrival

連講都唔想

問非所答，問多2句就變左白痴

呢兩個都唔算？
gemma 3n唔單只細到電話都run到，而且仲要可以了解圖像，仲要實際做到嘅嘢多
利申我電腦就係用ollama run gemma嚟做coding agent

反而其實deepshit佢放嘅係v3，而唔係具推理能力嘅r1
而v3根本就係廢嘅，跑分勁嘅係r1。
雖然r1其實都係廢嘅
啲「思考過程」十足一個無溫書嘅學生試圖大包圍攞分咁

你係咪用左ollama上面既qwen3 8b distill/finetune當左685b果個

理解圖像只係multimodal，唔代表sota喎。兩年前有個llava一早玩過

that said 普遍大家都識right tool for the right job
人地model open source/weight放出黎，你又唔會少忽肉，話人地個野垃圾唔代表其他人都覺得垃圾

點解

曾經用追一排.. 開頭覺得個Reasoning 幾勁
不過server 成日busy.. 答答下又唔答得..
後面轉返用o3 舒服返晒

gemma點撚樣qwen
你係咪傻撚咗

我講左685b你都Sense唔到我係講緊deepseek r1 0528 證明同你討論唔會有得著

班支那閪係鐘意扮香港人認親認戚

r1邊有open weight過
而且deepshit多幻覺都唔係得我講

https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/tree/main

??????

佢後尾先open weight撚知咩
一開始公開俾人用嗰時都無open weight
但都無改佢幻覺嚴重嘅事實
gen出嚟啲嘢前言唔搭後語

Condom

多撚餘 叫佢翻譯野 翻翻下仲自己歸納 同 比意見
十份野8份都BUSY

Reddit鬼佬唔計啦
都係CC台手腳最識嘢

問下佢64，佢會hang機