12 Mar 2025 - tsp
Last update 12 Mar 2025
14 mins
When large language models (LLMs) first emerged, I was highly skeptical about their capabilities and questioned whether they could truly influence multiple domains beyond merely generating generic text or reproducing existing content. However, as I explored their practical applications and integrated them into my personal workflows, I realized their immense potential. Today, LLMs like ChatGPT and LLAMA have become indispensable in both my professional work as a scientist in quantum optics and electron microscopy and my personal projects. These AI tools represent one of the most transformative technological advancements since the rise of the internet, automated proof systems, and search engines, providing powerful assistance in structuring information, generating ideas, and refining communication.
Their ability to streamline tasks and foster creativity has significantly improved my efficiency. Whether summarizing research papers, refining email communication, or brainstorming project ideas, LLMs provide crucial support across various aspects of my work and hobbies. Below, I’ll explore some of the key ways I leverage these AI models in both professional and personal settings.
This is a brief overview of how I personally utilize LLMs in a professional context. If you intend to do the same, it is essential to verify with your employer regarding legal constraints and permissible information sharing. For instance, I am strictly prohibited from sharing any student-related data, their work, or confidential institutional information with an external LLM provider. In such cases, using locally evaluated LLMs, such as LLAMA, offers a viable alternative.
Reading through lengthy research papers can be time-consuming. ChatGPT not only summarizes key points but also understands the content of a paper in context. It utilizes a local configuration that aligns with my research interests, highlighting how papers relate to my field while still providing a summary even if they are not directly relevant. Ultimately, I still need to read the papers in detail myself, but this process significantly speeds up initial assessments.
I use vector embeddings to index arXiv articles automatically. ChatGPT not only rephrases key sections to match the style of my database embeddings but also helps in quickly filtering and discovering interesting articles. This allows me to locate relevant papers more efficiently and retrieve them later with greater ease, improving both search accuracy and accessibility.
I often struggle with phrasing emails in a socially appropriate manner. ChatGPT helps refine my wording, ensuring I convey professionalism and friendliness as I intend. However, it is crucial to ensure that no personal or sensitive information is inadvertently shared with third-party AI services.
When I receive complex or ambiguous texts, I sometimes verify my interpretation with ChatGPT to ensure I’m understanding them correctly. This is particularly useful when dealing with texts where different language styles, varying backgrounds, or complicated legal meanings could lead to misinterpretation.
During the writing process and before submitting a research paper, I use ChatGPT to check for logical inconsistencies, identifying weak arguments or missing links in reasoning.
While writing research papers, I use ChatGPT to rephrase paragraphs. Rather than accepting entire rewrites, I use AI-generated suggestions as inspiration to refine my own writing. However, it is essential to maintain one’s own writing style to preserve variety and originality. Relying too heavily on LLM-generated text may lead to homogenization, where both individuals and AI converge towards a single stylistic approach. To prevent this, I ensure that my final writing remains distinctly personal and diverse.
ChatGPT helps generate skeletal structures for applications, which I then refine and extend with specific functionality, ensuring full understanding of the generated code. We use this for Python, C, and JavaScript, as well as occasionally translating legacy MATLAB code into more practical languages like Python.
When adopting a new tool or API, ChatGPT provides a quick and structured introduction, helping me understand its functionality without extensive documentation searches. As frameworks grow increasingly complex, scientific work often requires quick hacks to solve specific problems rather than robust implementations. This approach differs from formal software development, where long-term maintainability is prioritized. If a temporary solution proves necessary for continued use, it must then be implemented properly. Many of these quick tests and experiments happen in Jupyter notebooks, where fast iteration is crucial. Coming from a software background, adapting to this style of rapid prototyping was initially challenging for me, and I found it problematic. However, I realized that in scientific computing, one-shot scripts that are only ever run once can be necessary. While many scientists struggle with writing structured code, this iterative ‘try and test’ approach is essential in exploratory research.
ChatGPT assists in brainstorming, structuring ideas into actionable projects, and performing feasibility analyses to determine the practicality of various ideas. This helps quickly discard totally unreasonable ideas, which is essential as a scientist is also an idea generator as part of their job.
I use LLMs to analyze newspaper articles by manually or automatically inputting links. The AI provides a summary while also assessing the reliability, framing, and potential bias of the content. This helps determine whether an article is manipulative, biased, or potentially spreading misinformation. LLMs are highly effective in detecting linguistic framing and selection bias, making it easier to confirm any suspicions of an article’s intent. Additionally, they condense articles by removing unnecessary boilerplate text, leaving only the core content for efficient reading.
I use ChatGPT as a conversational partner to explore new project ideas, whether it’s building a machine, writing software, or developing a new workflow. It often helps to receive feedback, identify logical flaws, or organize a flow of unstructured thoughts systematically. Additionally, the ability of modern LLMs to remember and track the bigger picture provides valuable context and continuity in brainstorming sessions.
I automate the extraction of structured data from supermarket recipe PDFs using ChatGPT and LLAMA. The structured output is then inserted into SQL databases to track purchased items and analyze spending patterns for greater efficiency in both cost and waste reduction. Additionally, LLMs can generate cooking suggestions by creating new recipes based on available ingredients.
I use LLAMA combined with BERT encodings and PostgreSQL to create a private document retrieval system. Unlike traditional search engines that rely on keyword matching, document retrieval systems using vector databases work by embedding text into high-dimensional vector spaces. These embeddings allow for semantic similarity searches, meaning that queries return the most conceptually relevant documents rather than just those containing specific words. This makes it easier to locate relevant information, even if different terminology is used, and helps in efficiently organizing and retrieving stored knowledge.
Together with my partner, I use DALL-E and ChatGPT to generate playful illustrations of our shared fantasies. We both have vivid imaginations but lack the skills to create computer graphics or draw beyond a childlike level. This technology allows us to bring our ideas to life, and by utilizing the variations generated by LLMs and DALL-E, we can also explore new creative directions.
For my blog, I use DALL-E to generate images, Canvas mode for drafting, and ChatGPT to check for inconsistencies and logical fallacies. While I am unsure if AI-generated images are always the best choice, I use them in articles where I lack data or photographs to avoid empty space. Additionally, Canvas mode allows for collaborative editing with an LLM, enabling me to refine specific paragraphs, add relevant information, and discuss content dynamically, making it a valuable tool for structuring my writing.
For quick learning and idea refinement, I ask ChatGPT questions and engage in discussions about various topics. I don’t primarily use ChatGPT to query factual information but rather to discuss logical ideas and constructions. This approach enables highly accurate logical reasoning. While retrieving facts is possible using LLMs with browsing capabilities, their main strength lies in facilitating logical discussions and collaboratively constructing projects and topics.
I use a combination of LLAMA and a custom Python script to analyze and categorize emails automatically. The script utilizes IMAP to read the content without ever deleting emails. It reacts to certain inputs, such as appointment requests, by notifying me via XMPP and informing me about other scheduled events in my calendar. It also alerts me about urgent messages via XMPP and provides incremental daily and weekly summaries of my communication. This setup functions like a helpful assistant, akin to a secretary handling emails. While it does not replace reading my own mail, it can be effectively used as a spam filter. However, I am not a proponent of automatically deleting emails due to the risk of false positives.
I have integrated LLAMA with function-calling capabilities into my home automation system, allowing it to schedule tasks, browse the web, and access local files. This enables natural language interaction with the home automation system, even for flows that were not explicitly predefined, as long as the necessary capabilities exist. By leveraging a vector store and database, it also allows for creating and querying profiles using natural language. However, not all local files are accessible for privacy reasons. Additionally, scheduling tasks means that the LLM can invoke queries on itself to perform actions at a given moment. It can also respond to messages via XMPP, and I am currently developing a voice assistant using OpenAI’s transcription capabilities, though this project is still in its early stages.
I’ve utilized LLAMA here instead of ChatGPT, even though ChatGPT offers more advanced and higher-quality responses, because I want a long-term stable system that does not depend on cloud services or commercial providers.
Creating fantasy worlds and storybook-style graphics was my primary motivation for using Stable Diffusion. However, I have also explored generating fake photographs out of curiosity and found it to be quite straightforward. In addition, I leverage it to dynamically generate images for applications I develop.
I use LLMs to quickly draft code frameworks and functions as well as quickly lay out system architectures. However, I have to ensure that I fully understand and extend the generated code. In my private projects, stability and maintainability are my highest priorities since I write a large amount of code and cannot afford to constantly refactor old implementations. Therefore, robustness is key to ensuring long-term functionality and efficiency. While LLMs provide an incredible speed boost, I follow the same principle as in my professional work: quick hacks can be useful for rapid experimentation, but any long-term solution must be implemented properly. The usage of ChatGPT has significantly accelerated development for my private projects while allowing me to maintain structured, reliable code.
For complex problems, I rely on ChatGPT, particularly the 4-o3-mini model, which has outstanding reasoning capabilities. As a physicist, I frequently contemplate theories about how the universe operates, and ChatGPT’s logical analysis helps refine these ideas. Additionally, I have numerous engineering ideas, and ChatGPT proves invaluable in assessing their feasibility, identifying potential flaws, and structuring them into coherent plans. Its capacity for logical reasoning not only aids in technical problem-solving but also enhances creative daydreaming by expanding and refining concepts in a structured way.
While I am not an artist, I have access to a variety of machines that allow me to create intricate decorations. I often have clear ideas of how I want designs to look but lack the artistic skills to bring them to life manually. By leveraging tools like Stable Diffusion and DALL-E, I can generate visually appealing yet manufacturable designs, such as gothic engravings, whimsical fantasy illustrations cut from wood, and intricate milled or engraved patterns. These tools enable me to transform conceptual ideas into tangible pieces that go beyond purely functional machine parts and equipment.
Although my creations may lack the refined touch of a professional artist, they are more than adequate for my personal projects and creative space. This approach allows me to experiment freely, refine my designs, and ultimately produce unique, customized artwork with the help of AI-driven inspiration.
This is an example not from my personal usage but from my social circle.
Two of my friends use ChatGPT as a supplementary tool alongside their psychotherapy. One of them, who has borderline personality disorder (BPD), turns to ChatGPT during downward spirals caused by BPD splitting. By describing their situation, they receive a structured logical analysis of typical human behavior, which helps them break the negative thought cycle and regain a sense of stability. The other friend, who has Asperger’s syndrome, relies on ChatGPT to interpret emotional content in messages and interactions that might otherwise be unclear. They also use it to identify hidden meanings and potential social cues in advance, making it easier to navigate conversations and structure their responses. Additionally, ChatGPT assists them in managing aspects of daily life that can become chaotic due to ADHD.
While AI should never replace professional mental health care, both of them continue working with their therapists. However, ChatGPT has proven to be a highly effective supplementary tool for emotional processing, reasoning, and improving social interactions.
LLMs like ChatGPT and LLAMA have significantly transformed how I approach both professional and personal projects. Whether in scientific research or creative endeavors, AI has become an indispensable tool, enhancing efficiency, creativity, and problem-solving. While it is essential to critically assess AI-generated content and verify its accuracy, the advantages of integrating AI into daily workflows are undeniable.
Running large-scale LLMs on personal hardware remains challenging due to significant resource demands. While LLAMA 2.x performs well up to the 7b models, running a 70b model or larger requires substantial hardware investments. As a result, highly capable models currently necessitate reliance on commercial services, which, despite ongoing concerns, often prove more cost-effective than purchasing dedicated hardware. Ideally, I advocate for self-hosting systems whenever feasible—running LLAMA on a high-end GPU is my preferred approach. However, for large-scale AI models, utilizing commercial services like OpenAI’s ChatGPT remains the most practical option.
If you’ve found innovative ways to use AI in your work or hobbies, share your experiences to contribute to the ever-expanding possibilities AI offers!
This article is tagged:
Dipl.-Ing. Thomas Spielauer, Wien (webcomplains389t48957@tspi.at)
This webpage is also available via TOR at http://rh6v563nt2dnxd5h2vhhqkudmyvjaevgiv77c62xflas52d5omtkxuid.onion/