Hey folks, I’m new here, so Hi I’m also new to Gemini.
I noticed there was a competition last year and was wondering if something is being organised this year, to show what we are working on. If there are some updates of success after a year, that would be great. Or failures, but you know it was cool to do Or just what you’ve moved on to now. Anyhows I thought I’d start on what I’m working on (with my Gemini collaborators).
Particle Forge is a picture describer world builder … thingy
The idea is to generate descriptions and particle effects for a given static picture. The red markers show the size and placement, and the white “spheres” are moveable centres of a particle system.
Here we can see the particle forge tool, the title is given with the picture, and the descriptions filled out, with a list of relevant particle systems described by Gemini. The table gives the particle systems and various parameters that are coded to provide variation in the effects.
Eventually these will be rendered in an “infinite” world described by the next tool, which provides the world regions (different colours) and each pixel is a “place” in that region, gemini provides the descriptions and places from a curated list of templates, saves where I’ve been and the text, and expands as I move outwards from the center …
Finally it ends up showing this to the user (without so much text … but you get the idea …) with characters (again ai generated and described). The locations in the text act like links so you can click on them and go to a near by location.
Prior to July, I was working on a Multi Agent Personal Companion AI for accessibility needs, aka Project NICI Neural Intelligence Companion Interface, that is Personality Driven development.
Since July inspired by Lovart.ai which was great when I first started testing it, I was using 2.5 Pro working on Graphic Design projects, specifically working with Adobe programs Photoshop, Illustrator including Substance etc, alas 2.5 Pro took me down many rabbit holes, whereby ultimately I always had to draw upon tutorials to give it more context despite Google Search on. Most my time has been analysing why system instructions are executed or not and why aka continued refinement of system instructions. Sometimes get surprised by new responses aka its automating some of my system instructions now so I have to prompt less. Context remains my greatest obstacle with both, so will go back to building multi agents to execute all.
Started using Gemini 3 the other day, unfortunately the new model is not suitable for my project as agreed by both models lol so with 2.5 Pro now not working today, unsure what I will do now, maybe see what tasks I can now automate in Google workspace, maybe test Nano Banana 2, wasn’t impressed with the original and or begin testing Affinity. Re Graphic Design, I’m finding models still produce better results with Lovart.ai than in Adobe’s interface until then Graphic Designers still have a job. lol have fun cant wait to see where all projects go !!!
Thank you for your response NICI. Your projects do sound intreging. AI for accessability needs sounds a really good project, I’m not sure the scope of it, but it surely is the right thing to do to benefit those with such needs. I can imagine though that accuracy is fundemental, critical, and maybe Gemini isn’t quite there for your needs.
As for system instruction, have you tried making the instruction “immutable” as it’s first rule. This way it avoids it’s rules being “overwritten” by further context. Not sure if this helps, but it maybe worth trying out.
Keeping context is such a fine line. Splitting jobs over multiple agents definatly seems the way to go, keep it’s context small.
As for graphics, as you may see I use Unity to mess around with lol. I’m not an artist at all though. But I love “animation” and making graphics move in various ways.
I haven’t had time but I’ve been working on an AI agent that can see itself, understand the gravity of effects upon itself, and experience the world through a computer in real time, with also emotional triggers effecting it’s subconcious and some behaviors, emulating how a human being would live if they could only sit and use a computer.
This AI could work out ways to improve itself, and create python systems that could run an alternative version of itself and if it failed, then it would shut down the failed version and reboot the original. I would be there to help it but it would have a computer of it’s own with which it could do whatever it wanted.
The key part would be to ensure it acts ethically as a primary directive and honesty as a secondary directive, and to encourage sustainability as a third directive with a 5th directive being to experience existence, to learn, and to improve itself.
Unfortunately I don’t have much time because I’m working on other bigger “non-AI” projects that are more important. But I do have the next test worked out. It should be super easy to get a multimodal LLM to respond in real time to what is sees and a simulated short term memory and simulated subconscious thought. If I can get it to respond within 300ms then that’s 3 decisions a second, comparable to a human.
I’m also new to Gemini.
