Generative AI tools burst onto the scene in November 2022 with the release of GPT by OpenAI. The era of large language models had arrived! Overnight, the world was plunged into machine-run environments that could only previously be depicted in Hollywood. The excitement surrounding the potential benefits of generative AI (https://news.mit.edu/2023/explained-generative-ai-1109), from boosting worker productivity to advancing scientific research, is undeniable. While the rapid growth of this new technology has enabled the swift deployment of powerful models across various industries, the environmental consequences of this generative AI “gold rush” remain elusive, let alone manageable.
As a user of generative AI tools for work and scientific research, I have consistently believed that these tools serve as powerful aids to enhance my work. Whether it’s using generative AI agents to write a series of software functions with an intuitive front end, accelerating my productivity, or providing in-depth analysis of diverse topics from multiple data sources, the integration of these tools into my daily routine has enriched my work. A wise mentor once advised me during my scientific journey (a long time ago) that one doesn’t need an extraordinary intellect; all you need is the ability to find data and connect disparate sources effectively — just what LLM systems are generally designed to undertake!
It was inevitable that Astronomy and Astrophotography would adopt AI tools. I couldn’t have predicted the profound impact they’re having on the field. These aren’t professional astronomy tools, large-scale data processing engines, or result inference systems. Instead, they’ve gradually emerged and enriched the field’s norms. As reasoning engines have improved, their understanding of my hobby is astounding. Recently, I wanted to plan a fun imaging session and asked Google Gemini, “Hey Gemini, can you provide me with an imaging target list for my telescope tonight so I can try to image as many Messier objects as possible?” The response was quite remarkable, as detailed in Appendix 1.
However, such scheduling and data interpretation make sense in light of the development of other LLM systems. Essentially, the above search is a reinterpretation of different data sources, finding information, and organizing it—a modified “Google” search. Stallarmate recently introduced it Stella AI system built into the which allows the use of natural language to control and manage your connected imaging setup or observatory. The system will undertake a number of very advanced operation from setup and implementation of different sequences as well as having a knowledge of what you have done before it can also recommend how to collect more data to enable the optimal signal to noise and there for generate the best data. Being an avid user of Stellar mate this development excited me but the need for internet connectivity whilst remote is still problematic.
So currently I am able to utilize AI driven systems to define my optimal imaging routine for a night as well as sequence and generate the specific subframes and number of sub frames that I need to generate the optimal data. Data processing is not immune of the tools either. Many of the basic image processing workflow components, such as background gradient removal, sharpening, noise removal are now covered by excellent ML tools Such and Russell Croman tools (http://RC-Astro.com) or GraXpert (https://graxpert.com/). Although these tools are not utilizing reasoning / generative AI they utilize the ML capabilities to learn how data should look and what is the optimal method for the data processing. The RC-Astro tools have altered every users workflow with their ability to simplify what was some of the most complicated components of data processing namely the sharpening and appropriate noise filtering.
So where will this end? Very recently Pi2LLM appeared (https://github.com/scottstirling/pi2llm) this tool built by Scott Starlings likely to become as big as other tools from the descriptions the ability to “- Get recommendations on your next processing step. – Ask for a detailed description of your astronomical target, which LLM Assistant will generate based on the astrometric data. – Request a summary of the processing steps applied to a finished image. – Ask general questions about PixInsight processes in the context of your current image. – Customize the System Prompt as desired” is one step away from Nirvana a tool that not one suggest but also implements.
The ability of software to undertake all of the steps in a logical progression for then ext generation of imaging platforms. The explosion of small imaging platforms such as Dwarf Labs, Viaonis and ZWO SeeStar etc. have opened this hobby as well as night sky exploration to many more people. But as of today these systems do not include AI components I can only image what happens if they do. But this is a logical progression full scheduling, full image manipulation and processing all through AI. This could be transformative to the hobby. Imaging the scene of an open dark sky site and a army of SeeStars all thinking and communicating with each other….. But lets not pick on or bash the diminutive devices our behemoth dinosaurs already have the ability to fully adopt AI in their workflows it is potentially our mental inflexibility that limits our ability to grasp and embrace this technology or is it the goto – none-goto or 35mm Film vs CCD fights all over again?
I have been asked if such tools will stifle the innovation and ability to ultimately be a unique individual as all images will start to be processed in a robotic singular manner. But we should remember that these are tools and not the be all and end all of the processing task. I am sure that the APOD images will come from personally processed data and not the robotic systems for years (maybe months). I for will be embracing these tools to their fullest. If they simplify my hobby and make it easier for myself to generate images that I find appealing then ‘hold my beer’. The more these tools are used the more they will improve as they learn form the community they will surpass the normal users abilities in many cases and I for one will be helping them learn.
However we adopt these tools and how they fall into the workflows they are here to stay embrace them or not they won’t care but our hobby has changed drastically in the last 30 years, none goto, push augmented goto, full goto, plate solved augmented goto; film cameras, CCD and CMOS; Manual guiding, full autoguided system and finally the rise of the full automation with tools like NINA or Stellarmate. Each of these were disruptive as the last and AI/ML is just the next to come along.
