Hey all, we have a quick update for everyone who participated in the Hacktoberfest Open-Source AI Challenge: Ghost in the Machine.
We were blown away by the creativity of the submissions for our final Hacktoberfest challenge. From haunted terminals to heartfelt projects honoring loved ones, you all brought the spirit of the season to life with open-source AI. With 21 winners to select across our overall prize and 20 partner categories, we need a bit more time to carefully review everything and make our final decisions.
We'll be back with the winner announcement by the end of next week. Thank you so much for your patience, and for putting so much effort into your builds. 👻🌼
In the meantime, check out the Presage SmartSpectra SDK Challenge running now!
Happy coding!
Top comments (19)
You sure there isn't a Ghost in your development team that is delaying this?
Jokes aside, couldn't wait to see the Winners :D
Nah, they have a sloth (Sloan). They need a ghost or something like a Genie 🧞♂️ 🤣
Joke apart, the number of submissions is indeed overwhelming.
The delay is totally understandable.
Hats off to the DEV Team 🥰
It's because they realized I made a new open web protocol (an SDK) to allow a centralized site to act as a dummy cross-origin permission manager for browser-storage-partition to store and share open-weights models allowing airgapped edge compute via web browser with webGPU ans WASM-fallback, and it never touches my (or any) server except to trigger the local browser native request for shared strorage permission... request one time per site (e.g., authorize), and did so accidentally, and their minds blew up, as did mine 🤯.
Looking forward to the announcement!
Pls. take your time, we have your back 🥰
A good judgement is often far better than a quick one ✊
Good luck 😃
Thanks! I'll need that luck to win this one 🥰
There are like millions of submissions, LOL 🤣
But to be honest, this is the first DEV Challenge projects I've built that I'm using now every single day, and improving it as I speak. So it's already a win for me.
Once it's feature rich enough, and stable enough, I'm planning to release it.
So this is one of those projects I won't lose either way!
I'll make a demo video soon and post here on DEV.
You'll see. It's really useful 🥰
Your submission is one in a million! There may be millions of submissions, but yours is truly special. Never forget that! ❤️
I'm really looking forward to this video. I'm always excited to watch your videos, haha! Please make sure to mention me when you post it so I don't miss it.
Wishing you lots of good luck! Whatever happens, I'm sure everything will work out for the best. 😊❤️
Thanks! That means a lot 🥰
building ossuary for dead domain is the real challenge for me, making your ai refuse to lie.
You can't make it "not lie" because AI is a calculator, really. You can, however, give it preset expectation and do 'verdicial enforcement'.
For example: PathMap Alchemy uses small models to build a PubMed dataset. Then AI is told to find a literature based discovery and extract 10 supporting verbatim quotes from the dataset.
At this point, if AI "lies" and gives 6 verbatim quotes and 4 paraphrased or misstated it is a total failure. Suppose AI misstated a dosage on a drug or a formula... could kill someone if trusted.
So, use JS/php/python or whatever to do a btye-perfect check for the 10 verbatim quotes. On fail, retry with feedback loop with limit, and keep the dynamic content (feedback) at end of prompt and resending dataset exactly as before for RAG optimization.
It does work. I can make 4,000,000+ plant-derived molecular pathway discoveries using Gemma 4 26B. If I had a mac studio M2, I would anti-patent the future of medicine in a month.
Use programmatic validation, never trust AI outright, nor any human for that matter... we are just biological organic-intelligence calculators... ugh...
The 'biological organic-intelligence calculators' line is hilariously accurate. 😂
But I think you misunderstood how OSSUARY actually works under the hood. I don't just 'preset expectations' for the AI— I completely stripped the AI of its voting rights.
The AI is never allowed to determine if a website is dead, alive, or pivoted. If I let the AI decide, it would hallucinate. Instead, I built a strict programmatic Gate:
cannot_determineerror. I'd rather give the user nothing than a hallucination.Your PathMap Alchemy approach is brilliant for medical data where you need byte-perfect quotes for PubMed. For my project, the 'byte-perfect check' is DNS + Wayback CDX. The AI is only allowed to speak after the programmatic truth is verified.
I am 100% on the same page: programmatic validation > AI guesswork. Hope you get that Mac Studio M2 to scale PathMap!
Thanks for explaining! I see... so if AI only speaks after verification of truth you can guarantee a good response/dataset. Do you capture a bad dataset as well where AI was found to be inaccurate initially and add a fail code? You could probably soft-train a model for that. I can see several case uses where for a small fee you can get an honest and accurate validation of a domain or its public details or history/status.
This -->
🫣
Thanks for the update, and no rush! This challenge was a lot of fun. I built a tiny hand-drawn ghost that warns you before you run a scary PowerShell command 👻
Command Casperr
Always happy to hear feedback !
Anyone working on the Presage hackathon yet? I think I found a way to run it no egress (airgapped) on a tablet/handheld using a double-wrapper. I am gonna do a game platform for pediatric BPM/HRV metrics because it is gard to get them to sit still, airgappimg gives it a prayer for HIPPA compliance (though who knows).
The wrapper will be open-source and conceptually built to build a local runtime for a primary app with a sandboxed partner/helper runtime (could also add local AI airgapped this way like bringyourownbrain.org project).
If that gives you an idea for your project, I should have the double-wrapper code on Git this weekend. 👍
The obsession with wrapping remote APIs into multi-layered agent abstractions is creating massive hidden latency and failure cascades in production.
If you profile real-world workloads, bare-metal local inference (via direct C++ / llama.cpp or vLLM bindings) consistently beats fragile cloud wrappers on both predictability and cost. Tooling should eliminate abstraction penalties, not compound them.
21 categories and that many winners? You absolutely get a pass for the delay. The mix of "haunted terminals" and heartfelt projects sounds incredible—take all the time you need to judge them fairly. Can't wait for the announcement!
Take your time🙌