最近在学习本地运转人工智能模型，硬件还凑合（RTX４０９０），但觉得稍大一些的模型运转起来有些慢，所以决定测试一下小模型。也不干什么重要的工作，就是想问一些生活中的小事情。 今天测试的是Mistral，Chat GPT强烈推荐（我估计Chat GPT已经跟人类学坏了，它是故意的）。我先问了一下Mistral会不会说普通话，他说可以，但词汇量有限。我想，这不是什么大问题呀。于是问了下面一个问题： 馄饨馅儿怎么做才能鲜嫩滑爽？ 接下来请欣赏它的回答： 馄饨馅儿可以用豌豆、蛋、肉丝等材料制作，以下是一种经典的馄饨馅儿制作方法： 1. 将豌豆和水在汤锅里烧开，烧出皮肤。 2. 切成细丝，去掉豌豆皮。 3. 擦干豌豆丝并加入肉丝、蛋白、咸菜、辣椒粉、馄饨馅儿可以用豌豆、蛋、肉丝等材料制作，以下是一种经典的馄饨馅儿制作方法： 1. 将豌豆和水在汤锅里烧开，烧出皮肤。 2. 切成细丝，去掉豌豆皮。 3. 擦干豌豆丝并加入肉丝、蛋白、咸菜、辣椒粉、碧硝酸锠等调味料。 4. 将材料混合均匀，注意不要过混。 5. 用手或者吹管喷液做成小球形状的馅儿，用蛋黄烧焙后缀。 6. 将馅儿放入馄饨包装好，再使用馄饨皮包围它们。 7. 用油炸馄饨，直到外面黄金色并呈现气泡状。 8. 在蒸箱中蒸馄饨约10分钟，使其鲜嫩滑爽等调味料。 别的我就不多做评价了，你说这碧硝酸锠是什么鬼？ 我把Mistral的回答告诉了Chat GPT，这是Chat GPT的评论： ＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊ 哈哈，这回答也太离谱了，简直是“馄饨AI事故现场”！ 咱们逐条看看这“经典”做法有多出格： “将豌豆和水在汤锅里烧开，烧出皮肤。” → 啥叫“烧出皮肤”？豌豆都成洗澡小豆豆了... “切成细丝，去掉豌豆皮。” → 豌豆切丝？它是青椒还是竹笋啊，豌豆真冤！ “加入碧硝酸锠等调味料” → 化学老师快来！这是要让馄饨飘进化学实验室吗？“碧硝酸锠”听起来比味精厉害多了，吃完能升级成超人... “用手或者吹管喷液做成小球形状的馅儿” → 这是在包馄饨还是做3D打印？吹管喷液是什么厨艺流派！ “蛋黄烧焙后缀” → “后缀”？是代码写多了？你是把馄饨当 HTML 标签写了吗？ “用油炸馄饨，直到外面黄金色并呈现气泡状” → 馄饨油炸？？虽然有炸馄饨这种吃法，但和“鲜嫩滑爽”可八竿子打不着... “蒸馄饨约10分钟，使其鲜嫩滑爽” → 前面刚炸完，现在又蒸……这馄饨过得比我们一天还复杂。 总结一下：这回答估计是“人工智能外包给人工智障”写的！ Mistral 可能是没吃过馄饨，硬编了一套“科学料理”，结果馄饨馅儿变成了豌豆切丝 + 化学试剂 + 吹管喷液的奇葩产物。 要不要我给你写个正常点儿的“鲜嫩滑爽馄饨馅儿”做法？😄 ＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊ Chat GPT还挺幽默！😄 又问了一下Claude的看法，可能因为我使用的是英文叙述，Claude的回答比较严肃正板： ＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊ Oh wow! I can see why that surprised you. The Mistral model's Chinese response about making wonton filling is highly problematic and actually dangerous. Let me break down what's concerning: Some of the steps described are incorrect for making wonton filling More alarming is the inclusion of "碧硝酸锠" (which appears to be a made-up chemical compound that sounds like "potassium nitrate/saltpeter" mixed with other terms) The cooking instructions involve deep frying and then steaming, which is not how wontons are traditionally prepared This is a perfect example of the hallucination problem we discussed. Not only is this information incorrect, but following some of these instructions could potentially be dangerous if someone were to actually try adding chemical compounds to food. A proper wonton filling typically involves ground pork, shrimp, green onions, ginger, soy sauce, sesame oil, and other traditional seasonings - definitely no chemical compounds! This certainly validates your decision to uninstall Mistral, especially for a smaller model with fewer parameters. The hallucination risk is simply too high, particularly when providing instructions in other languages where the model might have even less training data to draw from. ＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊ 结论：慢就慢点吧，我还是会接着用DeepSeek R1、Gemma３、QwQ的 30B 左右的 Q4 版本，不然说不定将来某天会牺牲在这些小模型手里。
今天测试的是Mistral，Chat GPT强烈推荐（我估计Chat GPT已经跟人类学坏了，它是故意的）。我先问了一下Mistral会不会说普通话，他说可以，但词汇量有限。我想，这不是什么大问题呀。于是问了下面一个问题：
馄饨馅儿怎么做才能鲜嫩滑爽？
接下来请欣赏它的回答：
馄饨馅儿可以用豌豆、蛋、肉丝等材料制作，以下是一种经典的馄饨馅儿制作方法： 1. 将豌豆和水在汤锅里烧开，烧出皮肤。 2. 切成细丝，去掉豌豆皮。 3. 擦干豌豆丝并加入肉丝、蛋白、咸菜、辣椒粉、馄饨馅儿可以用豌豆、蛋、肉丝等材料制作，以下是一种经典的馄饨馅儿制作方法： 1. 将豌豆和水在汤锅里烧开，烧出皮肤。 2. 切成细丝，去掉豌豆皮。 3. 擦干豌豆丝并加入肉丝、蛋白、咸菜、辣椒粉、碧硝酸锠等调味料。 4. 将材料混合均匀，注意不要过混。 5. 用手或者吹管喷液做成小球形状的馅儿，用蛋黄烧焙后缀。 6. 将馅儿放入馄饨包装好，再使用馄饨皮包围它们。 7. 用油炸馄饨，直到外面黄金色并呈现气泡状。 8. 在蒸箱中蒸馄饨约10分钟，使其鲜嫩滑爽等调味料。
别的我就不多做评价了，你说这碧硝酸锠是什么鬼？
我把Mistral的回答告诉了Chat GPT，这是Chat GPT的评论： ＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊ 哈哈，这回答也太离谱了，简直是“馄饨AI事故现场”！ 咱们逐条看看这“经典”做法有多出格： “将豌豆和水在汤锅里烧开，烧出皮肤。” → 啥叫“烧出皮肤”？豌豆都成洗澡小豆豆了... “切成细丝，去掉豌豆皮。” → 豌豆切丝？它是青椒还是竹笋啊，豌豆真冤！ “加入碧硝酸锠等调味料” → 化学老师快来！这是要让馄饨飘进化学实验室吗？“碧硝酸锠”听起来比味精厉害多了，吃完能升级成超人... “用手或者吹管喷液做成小球形状的馅儿” → 这是在包馄饨还是做3D打印？吹管喷液是什么厨艺流派！ “蛋黄烧焙后缀” → “后缀”？是代码写多了？你是把馄饨当 HTML 标签写了吗？ “用油炸馄饨，直到外面黄金色并呈现气泡状” → 馄饨油炸？？虽然有炸馄饨这种吃法，但和“鲜嫩滑爽”可八竿子打不着... “蒸馄饨约10分钟，使其鲜嫩滑爽” → 前面刚炸完，现在又蒸……这馄饨过得比我们一天还复杂。 总结一下：这回答估计是“人工智能外包给人工智障”写的！ Mistral 可能是没吃过馄饨，硬编了一套“科学料理”，结果馄饨馅儿变成了豌豆切丝 + 化学试剂 + 吹管喷液的奇葩产物。 要不要我给你写个正常点儿的“鲜嫩滑爽馄饨馅儿”做法？😄 ＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊
Chat GPT还挺幽默！😄
又问了一下Claude的看法，可能因为我使用的是英文叙述，Claude的回答比较严肃正板： ＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊ Oh wow! I can see why that surprised you. The Mistral model's Chinese response about making wonton filling is highly problematic and actually dangerous. Let me break down what's concerning: Some of the steps described are incorrect for making wonton filling More alarming is the inclusion of "碧硝酸锠" (which appears to be a made-up chemical compound that sounds like "potassium nitrate/saltpeter" mixed with other terms) The cooking instructions involve deep frying and then steaming, which is not how wontons are traditionally prepared
This is a perfect example of the hallucination problem we discussed. Not only is this information incorrect, but following some of these instructions could potentially be dangerous if someone were to actually try adding chemical compounds to food. A proper wonton filling typically involves ground pork, shrimp, green onions, ginger, soy sauce, sesame oil, and other traditional seasonings - definitely no chemical compounds!
This certainly validates your decision to uninstall Mistral, especially for a smaller model with fewer parameters. The hallucination risk is simply too high, particularly when providing instructions in other languages where the model might have even less training data to draw from. ＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊＊
结论：慢就慢点吧，我还是会接着用DeepSeek R1、Gemma３、QwQ的 30B 左右的 Q4 版本，不然说不定将来某天会牺牲在这些小模型手里。