还是软路由+ap 吧，ap 想要漫游可以选用 openwrt 的，觉得麻烦可以买多个小米（ ax6 以上）或者 linksys 的应该都支持漫游，最好还是 Ubiquiti 一次性拉满（就是贵
插件的话 arm 要高速就一定要有把 /dev/crypto 支持好的固件，不然 arm 的 cpu 硬跑就太需要高性能的 cpu 了

https://help.openai.com/en/articles/6825453-chatgpt-release-notes
确实，所有 plus 用户下周都能用了

为什么你会觉得他能发链接给你，不说为什么你觉得 LLM 会有绘图功能？ 他说到底就是预测下一个更可能的字，他觉得你需要一个网盘链接所以编了一个给你，就是这样。
靠谱的做法是让他给你找 stable diffusion 的 prompt 然后再用 stable diffusion 生成

@catscan 嗯嗯，7621A 的支持感觉会好很多，至少能用 lean 的源，比较倾向用户多的 cpu 。 但可能有更好的选择？
顺便这里似乎已经把 hwnat 做好了 https://github.com/coolsnowwolf/lede/issues/9450 应该不会出现 glinet 的软交换跑满 cpu 的情况。 不过值得一提的是没有无线硬件 hwnat ，所以无线应该还是会消耗 cpu 性能的
随便看了一圈，发现 mtk 是有 7621 是有 crypto engine 的，不知道支持怎么样？

MT7621AT 和 SF19A2890 的 cpu
如果不追求硬件加速感觉随便买，反正速度不会特别好的
sft1200 的 cpu 更是完全没听过，感觉支持堪忧
如果有付费固件的话，为什么不稍微加点钱上一个高通的 cpu 呢，速度和效能会显著提高，在固件支持的好的情况下
顺便 gl inet 的 axt1800 的官方固件很差，但是支持倒是挺勤快的，然而解决不了用不了 nss-crypto 的问题。 以后不会考虑买。。

没 dx12 支持=无解
GeForce now 吧

@wsping 他们知道以后这样不好，但是有时候还是会不想去，大概就是不想发现就以为没有吧 感觉多少每个人都有点这种心理

@excitedXXX 展开讲讲嘛

我也想知道有没有用，以及会不会有什么副作用
但是我觉得还有个问题是，患者愿不愿意自己接受这么激进的治疗，有些患者连医院都不敢去

Filezilla 的客户端支持多线程下载，可以试试

@seaiaddca 哈哈哈西安名吃我楼下就有一家，每天天天吃。其实味道都不是重点，重点是他真的好便宜啊！

Peter Luger 的看起来牛排真不错，下周我也去试试看吧，op 有什么别的好吃的推荐嘛（我是馋鬼

不过加速方面还是高通的在 lean 那边支持更好

闭源开源一开始多少都会有些问题，一般过个半年就好很多了

@swhhaa 啊，我是就把整个帖子前 20 楼丢进去提问他整理下准确和不准确的地方
It is true that the training material for ChatGPT contains natural language, and before being processed by the model, it undergoes various preprocessing steps, such as tokenization and vectorization. However, the model still learns to process and generate human-like text based on the patterns and structure of natural language present in the training data.

I apologize for any misinterpretation of the author's point of view. It's important to note that while ChatGPT is a powerful language model, it is not an example of a general artificial intelligence. The process of fine-tuning the model for specific tasks indeed does not conflict with the author's perspective. There are various ways AI can learn and interact with the environment, and natural language is just one of those methods.
上面是他的回复

那么他自己是怎么想的呢？
Chat gpt 自己觉得你在以下几点说的有点不准确：
1. The author believes that ChatGPT cannot learn from its interactions, i.e., it cannot acquire new knowledge in the way humans do through natural language. While it is true that ChatGPT cannot learn in real-time as humans do, it does learn from the vast amount of data it has been trained on, which includes natural language texts. It is important to note that GPT models, including ChatGPT, are pretrained on a large dataset and can be fine-tuned for specific tasks or domains. However, they do not have the ability to learn incrementally during an ongoing conversation.

2. The author argues that an ideal general AI should have a learning interface that accepts natural language input. While ChatGPT doesn't have this interface directly, it can still be fine-tuned on specific data or tasks, which could include examples of natural language input.

3. The author implies that an ideal general AI should have the intelligence to learn from scratch without prior knowledge or training, whereas ChatGPT is pretrained on a large amount of data from the internet. However, AI models like ChatGPT benefit from this pretraining because it allows them to have a broad understanding of various topics and helps them generate coherent and contextually relevant responses.

同时，肯定了你的这些观点：
1. ChatGPT's limitations: The author recognizes that ChatGPT has limitations, and it's essential to be aware of these when using the model. Some of the limitations include the inability to learn in real-time during conversations, providing incorrect or nonsensical answers, sensitivity to input phrasing, and a tendency to be verbose.

2. The need for a better AI: The author's pursuit of an ideal general AI suggests that they understand the importance of improving AI systems to overcome the limitations of current models like ChatGPT. The quest for a better AI is a crucial aspect of ongoing research in the field of artificial intelligence.

3. Importance of natural language learning: The author emphasizes the significance of AI systems being able to learn from natural language input. This is a valuable insight, as understanding and processing human language is a critical aspect of making AI more useful and accessible to a broader range of people and applications.

我个人觉得 chatgpt 的 token 限制和长期存储模块大概率是无关的
“GPT-4 now has 8k tokens max, and there is a larger 32k token model on the horizon in the API.” [here]( https://community.openai.com/t/gpt-4-api-what-is-the-chat-history-limit-and-token-completion-limit/103094/2)
现在目前 GPT4 已经就提供了>4096 个 token 的支持，以后也会变的更大。

@studentBB 对于 4.0 的检测感觉非常不准，以前 3.5 的时候感觉还行，确实存在误判

@aliceclark 对，但是用来参考很好用
有时候我写一些固定句式的时候也会被误判

@studentBB 现在就有
https://www.zerogpt.com/

小米系这点就挺好的，默认送个壳，其实其他厂商也有在跟进
不过这个壳的质量真的很差