Even if Bing's AI is fundamentally GPT-4, it must got nerfed a lot. It's dummer than ChatGPT aldreay in many aspects, for example it doesn't understand "tanslate what I say later" and will directly translate the sentence. And in the recent days, it just stop working. It will search the web for what I want to translate, and freeze.
I suspect it has something to do with the RLHF part, or did they made the inference time per token shorter?
即使必应的AI基本上是GPT-4，它也被弱化了很多的版本。在许多方面它比ChatGPT还要笨太多，例如它不理解“稍后翻译我说的话”，而会直接翻译这句话。而且在最近几天，它更智障了。它会搜索我想要翻译的内容，然后就卡住了。
我怀疑这与RLHF部分有关，或者他们缩短了每个词的推理时间？
https://open.spotify.com/episode/3cemC6cxTLqgFtuVdRHYgf?si=PBU7jdmCR9i_AIVR49icrw&nd=1
A very good intro podcast to zk by the developer of starknet.
ZK, zero knowledge proof is a very also tool for many applications. The most brought up topic about zk is zk rollup on ETH. Fundamentally, blockchain is really not good at any general storage or computation, that's why programs on blockchain are called smart contract. In the long term, I still don't see the system built on blockchain could or should be used to handle too much general computation. What zk rollup brings to the system is only cheap computation with minimum accessible storage (so don't expect any server software running on it).
Zk is basically proof that you know something without actually revealing the information. The basic zk validity proof roll up process to L1 is: 1. the sequencer(s) aggregates all transactions into a block 2. the prover computes the block with a special paradigm and generate the results and the validity proof of the computation. 3. the prover submits the state difference/result and the proof to L1, there will be a verifier on L1 to verify those results.(verify takes very little computational resources)
This process separates the computation from the blockchain. It's crucial since, computation will be run on all nodes on L1, not scalable at all. Now you can only verify the results, you can also bundle them to further decrease the computational requirements. As comparison to optimistic rollup, which assumes all transactions processed are valid until the fault proof process, zk roll up doesn't have to wait for the days of roll up confirmation to L1.
Zk proof is the proof of the execution of transactions, so that is blockchain agnostic, meaning it could verify another chain from one chain. One major application is BTC <-> ETH atomic swap, without zk, you'll need some oracle (program outside the chain) to know whether the BTC address you want to swap actually did follow the swap process. Now with zk proof on the BTC chain's state change, you don't have to deal with all the problems come with the oracle. And other application is storage chain verification.
The contents in the podcast are very detailed. Some interesting topic includes why computation is cheap, but storage is not; how they went from ASIC to cpu;how the states are stored; the domain specific language to program zk.
#zk #zeroknowledge #zkproof #ETH #Blockchain #scale #scaling #AtomicSwap #BTC #rollup
https://vitalik.ca/general/2022/08/04/zkevm.html
"The different types of ZK-EVMs"
In the current stage, trade off between compatiblity and performance must be made when implementing zk-EVM. But different types of EVMs could co-exist to make the blockchain more scalable in different aspects.
Type 2's future market dominant makes most sense to me, it basically preserves the EVM environment while improves a lot on the prover time. Only the stuff from outside, like block structure and state trees are not compatible. Polygon and scroll are building towards this kind of EVMs.(Their product is about type 3 in current state)
“不同类别的ZK-EVM”
在当前阶段，实现zk-EVM时必须在兼容性和性能之间做出权衡。但是不同类型的EVM可以共存，以使区块链在不同方面更具可扩展性。
对我来说，类型2的未来市场主导地位是可预见的，它基本上保留了EVM环境，同时大大改善了证明时间。只有来自外部的内容，例如块结构和状态树不兼容。Polygon和Scroll正在构建这种类型的EVM。（他们目前的产品属于类型3）
https://twitter.com/luozhuzhang/status/1581385011604750336?s=46&t=Z3hVuhwSUrHhg3xDzw5U5w
https://www.youtube.com/watch?v=outcGtbnMuQ&t=3276s&ab_channel=OpenAI
GPT-4 demo, the biggest improvement imo is:
1. support image input; this is the huge part; cause users can draw something, and then ask GPT-4 to generate code or something else based on that.
2. larger memory, from 4k tokens to 32k tokens; it's crucial for analyzing long tax documents and codes. Seems like image input depends on that too.
3. A bit smarter than ChatGPT? Hard to say how much without direct comparison.
GPT-4演示，我认为最大的改进是：
1. 支持图像输入；这是很大的部分；因为用户可以画一些东西，然后让GPT-4根据那个生成代码或其他东西。
2. 更大的内存/记忆，从4k tokens增加到32k tokens。对模型分析长篇的 税务条文，代码等很重要，似乎图像输入也是依靠这个
3. 比ChatGPT稍微聪明一点？没有直接比较很难说有多少。
#AI #openAI #ChatGPT #gpt4 #future
Prometheus == GPT-4, the hype is gone 😅
https://www.youtube.com/watch?v=outcGtbnMuQ&t=3276s&ab_channel=OpenAI
GPT-4 demo, the biggest improvement imo is:
1. support image input; this is the huge part; cause users can draw something, and then ask GPT-4 to generate code or something else based on that.
2. larger memory, from 4k tokens to 32k tokens; it's crucial for analyzing long tax documents and codes. Seems like image input depends on that too.
3. A bit smarter than ChatGPT? Hard to say how much without direct comparison.
GPT-4演示，我认为最大的改进是：
1. 支持图像输入；这是很大的部分；因为用户可以画一些东西，然后让GPT-4根据那个生成代码或其他东西。
2. 更大的内存/记忆，从4k tokens增加到32k tokens。对模型分析长篇的 税务条文，代码等很重要，似乎图像输入也是依靠这个
3. 比ChatGPT稍微聪明一点？没有直接比较很难说有多少。
#AI #openAI #ChatGPT #gpt4 #future
https://openai.com/research/gpt-4
The section “Internal factual eval by category” is about how accurate the generated contents are. Improved from ChatGPT's 40-60% to 60-80%. In order to become a usable tool, this accuracy still needs to go much higher, since human reviews are super expensive.
"Internal factual eval by category" 那个部分是我最关心的地方，就是AI编故事的程度。我使用chatgpt和bing ai的经验都让我觉得他们还远远不能当工具，每当你指正它的时候，大概率它就开始编故事了。这个标准从之前的40-60%准确率到现在60-80%。还是远远不够，不然还是要耗费大量人工审查的精力。
https://www.youtube.com/watch?v=outcGtbnMuQ&t=3276s&ab_channel=OpenAI
GPT-4 demo, the biggest improvement imo is:
1. support image input; this is the huge part; cause users can draw something, and then ask GPT-4 to generate code or something else based on that.
2. larger memory, from 4k tokens to 32k tokens; it's crucial for analyzing long tax documents and codes. Seems like image input depends on that too.
3. A bit smarter than ChatGPT? Hard to say how much without direct comparison.
GPT-4演示，我认为最大的改进是：
1. 支持图像输入；这是很大的部分；因为用户可以画一些东西，然后让GPT-4根据那个生成代码或其他东西。
2. 更大的内存/记忆，从4k tokens增加到32k tokens。对模型分析长篇的 税务条文，代码等很重要，似乎图像输入也是依靠这个
3. 比ChatGPT稍微聪明一点？没有直接比较很难说有多少。
#AI #openAI #ChatGPT #gpt4 #future
https://www.youtube.com/watch?v=outcGtbnMuQ&t=3276s&ab_channel=OpenAI
GPT-4 demo, the biggest improvement imo is:
1. support image input; this is the huge part; cause users can draw something, and then ask GPT-4 to generate code or something else based on that.
2. larger memory, from 4k tokens to 32k tokens; it's crucial for analyzing long tax documents and codes. Seems like image input depends on that too.
3. A bit smarter than ChatGPT? Hard to say how much without direct comparison.
GPT-4演示，我认为最大的改进是：
1. 支持图像输入；这是很大的部分；因为用户可以画一些东西，然后让GPT-4根据那个生成代码或其他东西。
2. 更大的内存/记忆，从4k tokens增加到32k tokens。对模型分析长篇的 税务条文，代码等很重要，似乎图像输入也是依靠这个
3. 比ChatGPT稍微聪明一点？没有直接比较很难说有多少。
#AI #openAI #ChatGPT #gpt4 #future
https://alpaca-ai-custom3.ngrok.io/
我想翻译点东西，长一点点的都慢的要死，不知道是不是跑在核显上😅
luv letter做的拼多多利用androind漏洞严重侵犯用户隐私的详细分析视频，知乎网页没了，算补档吧
现代社会都是政府有不低干涉程度的社会主义国家（包括朝鲜），不存在纯资本主义的国家（可能从来就没这东西）。不过社会主义国家之间的差别也是可以根据个人的观感天差地别的。
"麦卡锡主义、日本共产党" 都是牵涉人数超级少，而且也没什么特别大规模的压迫的事件（不是说这就好），用来和墙比，是太看不起墙吧。
我个人觉得这20年，甚至这10年内的社会文化，人口素质都已经和以前差别太多了，不能用几十年前的内容来参考。特别是种族的问题上，发达国家的种族歧视 和 地域歧视比起来，哪个更严重也很有争议性
因为“政客们这样吵架，演给你” 已经是非常接近最优解的政治了。想百姓过好日子又不付出？历史已经无数次证明是做梦，只是可能经济高速发展，人均生活条件都有提升，掩盖了很多问题。
“假设完全没墙，是有被外来思想入侵，甚至颜色革命渗透的风险的。让有自己分辨能力的可以翻墙，去自由交流。” 颜色革命也有有那个土壤才革命得动，堵不如疏，丑化民主自由普世价值这些概念就是这样的后果。国内从人民到政府都是那个样，怎么疏；不就是逼有能力有想法的人润吗
确定不是在反串吗？ 很多观点比较墙内平行世界
我觉得他们做得很合理，高通胀只能这样压制，不然社会更混乱。而且他们的态度上也一直是“通胀不降，我们就不转向”；现在的问题就是，消费服务就业（除了少部分金融互联网）市场还是相当火热
不太喜欢在nostr上发图片，因为图片不在relay上存。另外nostr的relay保存和转发文字已经够吃力了。
https://twitter.com/pCtOtNJDtrc03al/status/1632463262749949952/photo/1
基本如上，另外这个漏洞在android 13才修复，但android 13之前升级上去的系统不行，已经被感染了。建议还是用ios更安全
是的，真的删了。药丸，还是ios安全很多，android的漏洞这么多年都修不完
用FDIC自己的保险基金去补差额。另外他们准备搞一个Program，让以后出现相同流动性问题的用其他资产抵押借钱，现在已经有250亿。
"Depositors will have access to all of their money starting Monday, March 13," said a joint statement from the Treasury, Fed, and FDIC. "No losses associated with the resolution of Silicon Valley Bank will be borne by the taxpayer," saying the bank-funded FDIC insurance fund will absorb any costs.”
https://www.axios.com/2023/03/12/silicon-valley-bank-regulators-backstop
他们宣传是这样，之后不够会不会纳税人出钱填/补充他们那个保险基金，还有另外那个program，就不清楚了。
题外话，我觉得SVB资产配置已经很安全了，现在似乎也没明确说要破产，市场以后还有信心的话，更透明保险运作也不是不行。我感觉监管机构到现在的做法都很好。 一大堆比它泡沫大得多的金融 房地产业，要直接刺破泡沫暴死的话，社会肯定大乱； 即使是Nostr的BTC Maxi，估计也没几个人能承受
https://www.zhihu.com/question/587624599/answer/2928285639
尼玛，比我想象中流氓软件还流氓