What are Main Use Cases of AI? |
| Published May 17 '22. Last edited Apr 22 '24 |
Uncertainty
Is AI a hype or is it real? Is it going to be more intelligent than human? Will our jobs be replaced by bots? If so, when? What skills do we need to focus on to better position ourselves in upcoming AI-age? I believe answers to all these questions have a lot to do with what are use cases of AI, how AI adds value to various industries of human socienty, which is an uncertainty that is being unfolded and I expect will unfold in the upcoming years. So I created the uncertainty of What are main use cases of AI on SaveNowClub, a platform to track uncertainties and their timelines by curating, taking notes, and summarizing information from the Internet as well as industry insiders, to track every key developments in timeline format under this important topic.
国内AI创新企业功夫源科技重磅推出「功夫量化」AI应用，这是针对金融行业的一次跨越性进击。
它能够在PB级金融数据海洋中，以秒级速度进行精准的信息搜寻，国内首款面向普通投资者的AI金融数据分析产品。
例如，用户可以轻松查询：
「今年3月每天收盘前30分钟逐笔成交数量最多的股票」
「中国石油和上汽集团在今年1月3号开盘后10分钟内每分钟逐笔成交数之差」
「2024年宁德时代三十分钟内涨幅超过1%的随后五分钟的涨跌幅度」
就在国内大模型竞争处于焦灼状态时，创业公司月之暗面却靠着kimi智能助手的“200万字长文本”率先出圈，甚至还催生了“Kimi概念股”。
Kimi的优势仅领先一个星期。随着阿里通义千问、百度文心一言等宣布免费开放200万—500万甚至1000万字的长文本能力后，kimi长文本推理能力瞬间被赶超，其“Kimi概念股”的热度也开始消散。
Zoom AI通过独创的「联邦AI」的技术路线，联合多个大模型，在特定任务上超越GPT-4，体现出了多个大模型互帮互助的强大能力，而且成本也能控制在GPT-4一半的水品。
去年底，从微软离职加入Zoom的华人AI大佬黄学东以Zoom CTO的身份发表了一篇技术博客，介绍了Zoom推出的联邦AI技术
——
差异化地利用不同成本的AI工具，让能力强成本高的AI完成难度大的任务；成本低能力弱的AI完成简单的任务，从而在完成质量和GPT-4几乎一致的前提下，将AI完成任务的推理成本降到了GPT-4的6%。
这一周，OpenAI视频AI工具Sora一出现，可谓是炸翻了天。
「饭碗保不住了」的恐惧，真实地击中了许多人。
不过，亚马逊的一位工程师Cameron Gould则认为，其实并不必对AI如此惧怕，它并不会导致我们失去工作。
有一些早期用户表示，最初对人工智能工具的兴奋很快就消失了。微软表示，除了支持AI转录会议内容功能的Teams外，一个月后大多数软件的 #Copilot 使用率下降了约20%。
种种因素使很多企业对全面应用AI仍处“观望”状态。
波士顿咨询公司的一项调查显示，尽管近90%的企业高管表示，#生成式人工智能 是他们公司今年的首要任务，但同时也有近三分之二的高管表示，这项技术至少需要两年时间才能超越当前炒作宣传的水平，其中约70%的人只关注小规模和小范围的AI应用试用。
Sam Altman: #ChatGPT is particularly useful for workers in these industries: coding, healthcare and education.
These are what I personally think are most thought-provoking takeaways regarding Large Language Models (LLM) from an one-hour YouTube video tutorial by Andrej Karpathy, Director of AI at Tesla. The tutorial used plain language to explain #LLM (Large Language Models) #UnderTheHood, how it works and how to build them. In addition, it also covers Andrej's vision of how LLM will evolve in next few years as well as a number of LLM's #cybersecurity vulnerabilities.
Regarding how LLM works, Andrej acknowledged that little is known in full detail despite the facts that we know how billions of parameters are dispersed through the neural network and we know how to iteratively adjust them to make the LLM better at prediction but we don't really know how the billions of parameters collaborate to do it. Therefore he recommended we think of LLMs as mostly inscrutable artifacts.
Regarding LLM's future, Andrej predicted an #LLM will evolve into an Operating System in a few years that
Google's #DeepMind Team releases #GraphCast weather forecast model that is more accurate than traditional weather forecast models.
由摩根大通（JP Morgan）人工智慧研究人员和大学学者所组成的团队，向ChatGPT和GPT-4这2个人工智慧大型语言模型提出了CFA风格的问题，以测试是否能够进行复杂的金融推理。
研究显示，ChatGPT也加入了“不合格”的行列。当被问及涉及大量选择题答案的一级模拟问题时，ChatGPT犯了很多“基于知识的错误”。GPT-4也犯了同样的知识性错误，但没ChatGPT那麽严重。
据报道，模型在CFA二级考试中的表现比一级差，主要是二级考试要求对股权投资和财报模型的推理分析。不过，研究人员表示，如果提示词足够好，GPT-4仍有能力通过CFA一、二级考试。
#LLM #UseCase: Medical Coding and Billing
中国湖北的快舟商业火箭公司自2016年开始向大数据平台输入发动机设计实验中的各种实验数据，比如发动机燃烧后内壁隔热层剩余厚度数据等等，经过不断的积累，可以利用该平台累计的数据对新设计的发动机做仿真试车，节省成本并加快进度。
Geoffrey Hinton demonstrated a use case of Generative AI using OpenAI GPT-4: solving a word problem (应用题) like:
The rooms in my house are painted blue or white or yellow. Yellow paint fades to white within a year. In two years time I want them all to be white. What shall I do and why?
David Caswell, founder of StoryFlow Ltd., an innovation consultancy focused on AI workflows in news production, talked about #LLM #UseCase of #NewsProduction.
Overview of Anything-to-anything AI models
These newsroom tasks can be automated
UCLA math professor tried #GithubCopilot to write technical document on steps of mathematical arguments and found it helpful.
#PromptEngineering in the field of technical document creation.
在最新的「马斯克传」中，摘录了马斯克和他的工程师之间的讨论。
「机器人的目标应该是在不充电的情况下运行16小时。」这相当于2个8小时轮班的人力劳动，而且完全不间断。
它极大地降低了劳动力成本，使产品和服务的预算可能只是现在的一小部分。而且它让企业没有理由在5年内以7倍的成本来雇用一个人来生产产品和服务，做同样的工作。
assessment of potential use case of ChatGPT-like solutions for the industry, for more details, please contact us.
生成式AI未来的形态和重点。
对于企业私有大模型的形态，SambaNova也有与众不同的观点。
他们认为最终企业内部不会运行一个GPT-4或谷歌Gemini那样的超大模型，而是根据不同数据子集创建150个独特的模型，聚合参数超过万亿。
相当于把GPT-4等大模型内部的Mixture of Experts（专家混合）架构扩展到整个系统，称为Composition of Experts（专家合成）。
在企业运转的每个节点运行一个完整且经过专门调整的基础模型，分别用法律语料库、制造语料库、风险管理语料库、财富管理语料库、客户销售语料库、客户支持语料库等等不同数据训练。
这些专家模型之间通过一种软件路由或负载平衡器联在一起，收到推理请求后决定具体向哪个模型推送提示词。
这一策略与GPT-4和谷歌Gemini等做法形成鲜明对比，巨头大多希望创建一个能泛化到数百万个任务的巨型模型。
分析师认为技术上可能谷歌的做法性能更强，但SambaNova的方法对企业来说更实用。
没有任何一个模型或人能完整访问企业的所有数据，限制每个部门能访问的专家模型，就能限制他们能访问的数据。
AlphaMissense出山第一步，就是对全部7100万种可能的错义突变进行了分类。
结果是，这个AI成功将这些变异中的89%，分类为“可能致病”和“可能良性”。相比之下，人类专家目前的成绩是0.1%。
This is my takeaway of a Youtube video tutorial of using ChatGPT to perform #DataAnalysis.
In order for ChatGPT to perform a specialist role in a certain area, as part of #PromptEngineering skillset, you need to learn how to customize #ChatGPT to assume your role and achieve your goal in your specialist field. That means before using ChatGPT, you need to provide answers to the following two questions in Custom instructions for ChatGPT, a feature launched recently in July 2023.
Question 1: What would you like ChatGPT to know about you to provide better responses?
Example Answer:
Question 2: How would you like ChatGPS to respond?
Example Answer:
Other than allowing you to upload data file, I observe the following data analysis capabilities from ChatGPT in the demo
As a data engineer for over 10 years, features of ChatGPT that I see useful at this point are:
At the same time, I observe the following limitations as of now:
At its core, LLM or generative AI is supervised learning being used to repeatedly predict next word given a set of words as predictor variables. However, at application layer, LLM / generative AI has huge potential because it enables prompt-based AI which revolutionizes AI application development process.
#PromptEngineering, high-quality prompts can be generated using a customized #LLM, this tutorial walks through how to create such customized #LLM.
Finetune Falcon 7b/40b instruct with your own data - The step by step guide about how to train falcon model for generating high quality midjourney prompt, from prep training dataset to comparing final results;
This talk by David Rosenberg, Head of ML Strategy, Office of the CTO, Bloomberg covers #BloombergGPT, an experimental project by Bloomberg to create a ChatGPT-like large-language-model (#LLM) that serves both general purpose as well as domain-specific purpose.
BloombergGPT is a 50-billion parameter LLM built using 570 billion tokens of language data, half of data are public, the other half are private.
Areas that BloombergGPT performed better than peers are:
Get me the last price and market cap for Apple, expect output of get(px_last, cur_mkt_cap) for (['AAPL US Equity']). Without previous knowledge of BQL, with few-shot learning of 3 example pairs of input and output, BloombergGPT can produce expected output correctly given input subsequently.market cap of AAPL vs. MSFTdiscusses methods using #OpenAIAPI and #LangChain to iteratively summarize a long text or PDF.
华尔街使用 #人工智能 和 #自然语言处理 技术做 #选股 研究的使用案例。
State Street used NLP technology help them develop an edge in identifying hidden gems in stock market during #StockPicking. To me, State Street's AI use case here is a combination of semantic-search version of Google search on a customized collection of documents plus automatic categorization of search results into pre-defined categories enabled by supervised learning.
AI failed to replace radiologist yet as predicted by AI expert Geoffrey Hinton because AI candidate failed Fellowship of the Royal College of Radiologists (FRCR) examination.
Used car dealer CarMax and rural supplies provider Farmlands uses #LLM through Azure OpenAI Service to #SummarizeText into short summaries so that they are easier-to-read and faster-to-read.
Terms of Use: You are in agreement with our Terms of Services and Privacy Policy. If you have any question or concern to any information published on SaveNowClub, please feel free to write to us at savenowclub@gmail.com