diff --git a/index.Rmd b/index.Rmd index dd0c8d2..c9fb4dc 100644 --- a/index.Rmd +++ b/index.Rmd @@ -313,18 +313,22 @@ There are many places where help can be found on the internet. So in case the do Another type of online resource that has recently been gaining in popularity is generative AI, such as ChatGPT. Generative AI models can be interacted with by asking it questions, including questions about programming. The AI responds by providing examples of code, explanations about what the code does, and how to run it. Of course, most AI solutions are not limited to code and will also answer questions on history, biology, quantum mechanics, and will even play Dungeons and Dragons with you, including throwing dice. -Generative AI models can be a great tool to enhance learning, as they can quickly answer specific questions and give coding suggestions. However, many of the limitations of web search apply to generative AI models as well (in fact, most of these models are something of a smart web search engine, as they are trained on a lot of text found on the internet). Therefore, you need to be very critical of AI-generated answers. The code that the AI generates may seem like it would solve your problem, but it may also do something incorrectly, such as calling functions that are no longer available, or even making them up altogether. Many generative AI solutions, including ChatGPT, are unable to provide references for their statements, and when asked, will make up a list of references and links that do not exist in reality. They may also answer questions completely wrong, but the explanation that they provide usually sounds quite convincing, therefore it may mislead you or make you second-guess yourself. When generative AI models are confronted about a wrong answer, they often insist that it is correct, and the longer you talk with a generative AI, the more it will get facts mixed up with its own previous answers, as it remembers and learns from its own output. +Generative AI models can be a great tool to enhance learning, as they can quickly answer specific questions and give coding suggestions. However, many of the limitations of web search apply to generative AI models as well (in fact, most of these models are something of a smart web search engine, as they are trained on a lot of text found on the internet). Therefore, you need to be very critical of AI-generated answers. The code that the AI generates may seem like it would solve your problem, but it may also do something incorrectly, such as calling functions that are no longer available, or even making them up altogether. Previously, many generative AI solutions were unable to provide references for their statements, and when asked provided a list of references and links that did not exist in reality, though this has improved in the past months. They may also answer questions completely wrong, but the explanation that they provide usually sounds quite convincing, therefore it may mislead you or make you second-guess yourself. When generative AI models are confronted about a wrong answer, they often insist that it is correct, and the longer you talk with a generative AI, the more it will get facts mixed up with its own previous answers, as it remembers and learns from its own output. Generative AI tools can be chatbots, like ChatGPT, but they can also be tools that suggest code snippets as you write code, such as GitHub Copilot. The AI code suggestions are based on the same models and have the same pitfalls. But in addition, they may suggest code that was taken from software whose license is incompatible with the license of your own code, which could cause copyright issues. Some of the newer code suggestion models are able to provide references to where the code is sourced from, and the license it is under. Some of the currently active generative AI tools are: -* [ChatGPT](https://chat.openai.com/) - the original chatbot that started the generative AI trend. Made by a team of top AI researchers that formed into a company. It is unable to provide real references of its statements, and is very often extremely overloaded due to its popularity. -* [Perplexity](https://perplexity.ai/) - an alternative chatbot that is able to provide references for its statements (and you can even pick which ones it uses to give you answers). However, it still gives biased output and may get confused with its own answers. -* [Bing AI](https://www.bing.com/?/ai) - Microsoft's version of ChatGPT. It can also provide references, but they are more phrased as Bing search terms. -* [Google Bard](https://bard.google.com/) - Google's version of ChatGPT. Does not provide references, but is less overloaded compared to ChatGPT. +* [ChatGPT](https://chat.openai.com/) - the original chatbot that started the generative AI trend. Made by a team of top AI researchers that formed into a company. The newest version, ChatGPT-4 is able to provide sources and references if you specify this in the prompt. It also allows the usage of custom GPTs, better suited to specific tasks, like the Python and R Wizard GPTs. +* [Perplexity](https://perplexity.ai/) - an alternative chatbot built using ChatGPT-4o and [Anthropic's Claude 3.5](https://www.anthropic.com/claude) models that is able to provide references for its statements (and you can even pick which ones it uses to give you answers). However, it still gives biased output and may get confused with its own answers. +* [Microsoft Copilot](https://www.bing.com/?/ai) - Microsoft's version of ChatGPT, also built on the ChatGPT-4 framework. It can also provide references for you if you ask. It generally allows you to access the newest version of ChatGPT for free, instead of having to pay for GPT Plus on the ChatGPT website. +* [Google Gemini](https://bard.google.com/) - Google's version of ChatGPT. It has recently improved a lot compared to Google's older models. Will also provide sources or references, if you ask for it. Can integrate with your Google Workspace (Gmail, Drive, YouTube), for example, it's a great way to summarize a YouTube tutorial into bulletpoints or clear steps. * [Amazon Codewhisperer](https://aws.amazon.com/codewhisperer/) - code suggestion AI, free to use, but works only with some code editors. +Note that all of these generative AI tools are built on proprietary models, but there are open source alternatives such as [Meta's Llama 3.1](https://llama.meta.com/) that you can use for your own applications. + +Every day, more generative AI tools become available, increasingly embedded in the tools we use, like Google Gemini replacing 'Ok Google' on your phone and some web-browsers shipping with built-in models, like Brave browser shipping with Leo AI based on [MistralAI's Mixtral model](https://mistral.ai/technology/#models). Despite the fact that generative AI is increasingly embedded, it remains important to check whether the results it provides are factual, accurate and if the output is compatible with the license of your own code. + ### Question and answer forums However, it may also happen that you discover a *bug* or something that you would qualify as abnormal behavior, or that you really have a question that no one has ever asked (corollary: has never been answered). In that case, you may submit a question to an appropriate Stack Exchange (e.g. [Unix & Linux for Bash questions](https://unix.stackexchange.com/), or contact the author of the package you are using (often by filing an issue on the package's GitHub page). diff --git a/index.html b/index.html index 5d78016..68e9f80 100644 --- a/index.html +++ b/index.html @@ -7,7 +7,7 @@ - + Linux terminal & Bash @@ -24,8 +24,8 @@ } }); -

WUR Geoscripting -WUR logo

+

Linux terminal & Bash

@@ -1469,16 +1469,16 @@

ChatGPT and generative AI

answers. The code that the AI generates may seem like it would solve your problem, but it may also do something incorrectly, such as calling functions that are no longer available, or even making them up -altogether. Many generative AI solutions, including ChatGPT, are unable -to provide references for their statements, and when asked, will make up -a list of references and links that do not exist in reality. They may -also answer questions completely wrong, but the explanation that they -provide usually sounds quite convincing, therefore it may mislead you or -make you second-guess yourself. When generative AI models are confronted -about a wrong answer, they often insist that it is correct, and the -longer you talk with a generative AI, the more it will get facts mixed -up with its own previous answers, as it remembers and learns from its -own output.

+altogether. Previously, many generative AI solutions were unable to +provide references for their statements, and when asked provided a list +of references and links that did not exist in reality, though this has +improved in the past months. They may also answer questions completely +wrong, but the explanation that they provide usually sounds quite +convincing, therefore it may mislead you or make you second-guess +yourself. When generative AI models are confronted about a wrong answer, +they often insist that it is correct, and the longer you talk with a +generative AI, the more it will get facts mixed up with its own previous +answers, as it remembers and learns from its own output.

Generative AI tools can be chatbots, like ChatGPT, but they can also be tools that suggest code snippets as you write code, such as GitHub Copilot. The AI code suggestions are based on the same models and have @@ -1491,23 +1491,41 @@

ChatGPT and generative AI

+

Note that all of these generative AI tools are built on proprietary +models, but there are open source alternatives such as Meta’s Llama 3.1 that you can use for +your own applications.

+

Every day, more generative AI tools become available, increasingly +embedded in the tools we use, like Google Gemini replacing ‘Ok Google’ +on your phone and some web-browsers shipping with built-in models, like +Brave browser shipping with Leo AI based on MistralAI’s Mixtral +model. Despite the fact that generative AI is increasingly embedded, +it remains important to check whether the results it provides are +factual, accurate and if the output is compatible with the license of +your own code.

Question and answer forums