π€ Combating LLM hallucinations
Dear curious minds,
I realized that the first issue of this newsletter was published more than one year ago on April 27, 2023. It was not always easy to write a new issue week after week, but so far, I did not miss one issue and I can tell you this newsletter is here to stay. If there are any suggestions or feedback, I would love to hear from you to make the second year of this newsletter even better than the first.
In this issue:
π‘ Shared Insight
Don't Be Fooled by AI: How to Spot and Avoid LLM Hallucinationsπ° AI Update
From Paid to Free: Open-Source Upscaler Takes on Magnific AI
Forget the New ChatGPT Memory: Custom Instructions are All You Needπ Media Recommendation
Unlocking Your Future AI Assistant: PKM Tools and Ed Nico's Newsletter
π‘ Shared Insight
Don't Be Fooled by AI: How to Spot and Avoid LLM Hallucinations
LLM hallucinations refer to a phenomenon in which the model generates text that appears to be accurate, confident, and coherent, but is actually factually incorrect. This can occur because LLMs are designed to generate plausible text based on patterns in their training data, rather than factual responses. And the problem is even amplified by the goal of LLMs to provide an answer which satisfies you, and something where you do not realize that it is made up will more likely do that than the model stating that it cannot answer your request.
LLM hallucinations can be dangerous because users may trust and act upon information that is incorrect or misleading. This could lead to serious consequences in domains like healthcare, finance, or decision-making processes.
To minimize LLM hallucinations, the following strategies can be applied:
Verify information: Do not blindly trust the output of an LLM. Cross-check facts with reliable sources, especially when dealing with critical or sensitive information.
Provide clear and detailed prompts: Crafting clear, well-defined prompts can help guide the LLM to produce more accurate results.
Be aware of the LLM's limitations: Familiarize yourself with the known limitations of the models you are using. Examples are the knowledge cutoff or language limitations. This will help you spot potential hallucinations and understand when to be more skeptical of the output.
Use multiple prompts: Rephrase the same query in different ways and cross-check the outputs for consistency. Inconsistent responses may indicate hallucinations.
Ask multiple models: Send the prompt to multiple models and compare the outputs. This can be done super convenient with the browser plugin ChatHub, which was covered in an earlier issue.
Retrieval Augmented Generation (RAG): Using a set of documents or the results from a web-search as source to answer the request might help to answer the request factual. It is especially useful if the sources are cited in the answer of the LLM.
Ask for evidence or sources: When the LLM makes a claim, especially on an unfamiliar topic, ask it to provide evidence or cite sources. This works especially for LLMs with web access, like Googleβs Gemini or Microsoftβs Copilot.
Use LLMs designed for your specific use case: Some LLMs may be better suited for specific tasks or domains. For example, there are models built for medical questions, like Med-Gemini. Choosing an LLM designed for your particular use case can reduce the likelihood of hallucinations.
By following these strategies, you can minimize the risk of encountering LLM hallucinations and better utilize LLMs for reliable information and decision-making.
π° AI Update
From Paid to Free: Open-Source Upscaler Takes on Magnific AI
In November 2023 a tool named Magnific showed how AI can generate outstanding high-resolution images from one input images. The biggest downside was and still is the high price, as the monthly subscription starts at $39 per month and covers between 100 and 200 upscales.
After months of reverse engineering Magnific AI's upscaler, which utilizes MultiDiffusion, ControlNet tiles, and details LoRAs, the developer philz1337x has created an open-sourced alternative of the technology.
The free upscaler, can be used for free on Replicate or locally in Automatic1111 as described in this post on π.
![](https://substackcdn.com/image/fetch/w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F82ba744a-38d5-495b-8fdd-1a59fcd8f678_1963x1958.png)
There also exists a low-resolution version which enables you to go from 64x64 pixels to 37 Megapixels in two steps. Available via API on Replicate and code on GitHub as stated in this π post.
![](https://substackcdn.com/image/fetch/w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fee03baeb-cb85-471c-aa10-89c581a986a9_1690x1690.png)
Besides the open-source release, there also exists a paid version at ClarityAI.cc which brings a simplified user interface, image history and extra features.
My take: I was stunned by the results of Magnific, but could not justify paying such a high monthly subscription fee. Super cool to finally have an open alternative which I can try and explore for free.
Forget the New ChatGPT Memory: Custom Instructions are All You Need
The memory feature of ChatGPT which was announced by OpenAI in February is now rolled out to all ChatGPT Plus users as stated in a π post.
This feature can extract and store user-specific information across different conversations. This means that users can enjoy a more personalized experience when interacting with the LLM.
There are limitations on who and where the memory feature can be used. It is not available in Europe and Korea, not for Team and Enterprise users and also cannot be used in custom GPTs.
The memory feature can be enabled or disabled. The stored data can be modified.
Furthermore, you can now choose in the model selection dropdown menu to create a temporary chat which is not stored or used for training. Also, the memory will not be used or created. However, custom instructions are followed. Before, this privacy-aware option could only be set globally in the settings.
If you would like to know more about the memory feature from ChatGPT, you should take a look at the official FAQ.
My take: I was able to test the memory feature with my account registered to a German phone number via a VPN connection set to America. The feature extracts information which it remembers from far too many chats and is annoying after a while. My advice is to turn it off by default and maintain it manually or activate before you start a chat which you want to remember parts of. But be honest, custom instruction can cover this too and are the easier to control memory feature.
π Media Recommendation
Unlocking Your Future AI Assistant: PKM Tools and Ed Nico's Newsletter
In the digital age, where vast amounts of data are generated daily, having an efficient system to manage and utilize this information can significantly enhance our productivity and decision-making capabilities. And as our digital footprints grow, the notes we take today may very well fuel our future AI assistants, making thoughtful PKM practices increasingly crucial.
If this is a new topic for you, and you want to identify the right tool to get started with, you should take a look at the blog article The 4 Notetaking Styles: How to Choose a Digital Notes App as Your Second Brain from Tiago Forte.
To stay up to date, I highly recommend the new PKM Weekly newsletter by Ed Nico. It covers updates and news about tools like Logseq (my personal choice), Tana, Anytype (exploring it as my recipe book) and others. Ed maintained before two newsletters with exclusive content about Logseq and Tana. I love the new combined approach, as I get the detailed news about my tool of choice and can read the headlines about major changes and news for other PKM tools.
Disclaimer: This newsletter is written with the aid of AI. I use AI as an assistant to generate and optimize the text. However, the amount of AI used varies depending on the topic and the content. I always curate and edit the text myself to ensure quality and accuracy. The opinions and views expressed in this newsletter are my own and do not necessarily reflect those of the sources or the AI models.