Back to all discussions

Manager   •   about 1 month ago

Gemini 3 Hackathon: Update on the Eligibility Review

Over the last week a number of you raised concerns about a few of our winners, and we want to address them in one place rather than in scattered threads. The bulk of the questions were about first place (Globot), with a couple of other threads we'll cover briefly first.

On judging:
Humans were involved at every stage of judging, including the finals.

On another winning team also placing in a separate hackathon:
Our rules require the project to be created during the Gemini 3 contest window. We confirmed that team's project was built inside our window. Submitting the same work to multiple concurrent hackathons isn't prohibited by our rules.

On Globot:
Most of the concern came down to two specifics: another project name (RoSP) appearing in the repo, and the public commit history starting with one ~12,000-line commit. Both are fair questions.

We sent the team many questions covering timeline, development process, AI tooling, and authorship, and asked them to answer with evidence. They responded in detail and have since made their past repository public and shared with us detailed accounts of their work. This included the predecessor RoSP repo.

The RoSP reference: RoSP was the team's earlier project name. Both the RoSP repo (first commit December 25th) and Globot were created inside the Gemini 3 contest window, which opened December 17th. The Globot repo is a snapshot of the RoSP codebase at the moment the team pivoted on December 29th.

The 12,000-line initial commit: A significant amount of those lines are lockfiles (package-lock.json alone runs several thousand), seed data, README/setup docs, and standard create-vite (React) and FastAPI scaffolds. The hand-authored application code is a much smaller fraction. The history "starts big" because they pushed the consolidated RoSP snapshot when they pivoted, then committed normally from there. The pre-pivot incremental history is now visible in the RoSP repo.

Based on the evidence we reviewed, Globot is eligible and will retain the Grand Prize

A real congratulations to the Globot team. They engaged with the review thoughtfully and quickly and made their work transparent. They built something the judges thought was excellent, and they earned the win.

Thank you to everyone who raised these concerns, and to everyone who participated. Integrity isn’t a nice-to-have, it’s a must-have.

The Devpost Team

  • 17 comments

  •   •   about 1 month ago

    Nice deflection. Most of us in the community can agree that the Grand Prize winner, while controversial for its commit history, is technically defensible as a top-tier project.

    However, focusing exclusively on Globot ignores the primary concerns being raised about the rest of the leaderboard.

    The 2nd, 3rd, and honorable mention selections are frankly baffling. It’s hard to trust a "human-involved" judging process when a 3rd-place winner is blatant promt injection.

    ​By only validating the Grand Prize, you are failing to address why several other winners are a joke compared to the caliber of submissions they beat. Integrity isn't just about following the rules it’s about maintaining a standard of quality that wasn't met here.

  •   •   about 1 month ago

    Community Response to the Eligibility Review — Unanswered Concerns

    [1/3]
    Devpost has posted their official response and cleared Globot to retain the Grand Prize. I want to respond carefully and factually.
    First, I appreciate that Devpost took the time to investigate and respond. But the response only addressed two of the many concerns raised: the RoSP project name and the 12,000-line first commit. I'd like to address both what was said and what was not.
    On what was addressed — logical concerns:
    Devpost states the RoSP repo's first commit is December 25th, and uses this to confirm the project was "created inside the contest window." But a commit date is not a creation date. The codebase contains azure_llm.py with api_version="2024-02-15-preview" — an Azure API version from February 2024. Code referencing a 2024 API doesn't become "newly created" by being pushed to a new repo in December 2025.
    The rules state: "New Projects Only: Projects must be newly created by the entrant during the Contest Period. The Project must be your original creation not a modification or extension of Your or anyone else's existing work." A codebase containing components dating back to early 2024 is, by definition, a modification of existing work.

    The response describes the 12,000-line commit as mostly "lockfiles, seed data, README/setup docs, and standard scaffolds." Standard scaffolds don't contain DJI drone sales prompts in Chinese. Standard scaffolds don't reference Imagine Cup 2026. Standard scaffolds don't come with a TELEGRAM_SETUP.md for integrating a sales bot.
    Devpost says "the hand-authored application code is a much smaller fraction." The concern was never about line count — it was about origin. The rules say: "Your Submission must: (a) be your (or your Team, or Organization's) original work product." A smaller fraction of code that comes from a pre-existing product is still not original work product.
    The response says the team "made their work transparent." They made the RoSP repo public only after being contacted by Devpost during the investigation. That is compliance under pressure, not transparency.

  •   •   about 1 month ago

    [2/3]
    On the RoSP repo itself:
    The team made the RoSP repo public as part of their response. I looked at it. The README's first line reads:
    Imagine Cup 2026 参赛作品 - "演示专用 (Mock-Ready)" 版本
    Translation: "Imagine Cup 2026 Competition Entry — Demo-Ready (Mock) Version"
    The repo contains:
    submitted_materials/ described as 最终 Imagine Cup 交付材料 (Final Imagine Cup Submission Materials)
    IC26 Official Rules and Regulations Dec 2025.pdf
    TELEGRAM_SETUP.md for integrating a Telegram sales bot
    telegram_bot_with_ollama.py confirming Ollama usage
    Azure OpenAI (GPT-4o), Azure AI Search, Azure Cognitive Services — consistent with IC26's requirement for at least two Microsoft AI services
    Project folder still named ai-sales-mvp/
    All documentation in Chinese
    This is not an "earlier project name." This is a complete submission to a separate competition with its own originality requirements. The same codebase was then translated to English and submitted to the Gemini 3 Hackathon.
    The rules state: "The Project must be your original creation not a modification or extension of Your or anyone else's existing work." A project submitted to IC26 and then resubmitted here is, by definition, an extension of existing work.

    What was not addressed at all:
    The rules state: "Entrants must develop a new application that uses the Gemini 3 API." The demo video and web app UI display "Powered by Gemini 2.0." PR #18, submitted on deadline day (Feb 9), is titled "Replaced Gemini 2.0 Flash with Gemini 3 Flash Preview." The project was running on Gemini 2.0 until the final day. Not mentioned.
    The tagline claims a 2M token context window. Google's official Vertex AI documentation states Gemini 3 Pro has a 1M token context window. This hackathon's own resources page also states 1M. No official Google documentation supports a 2M claim for any Gemini 3 model. Not mentioned.
    The rules state: "If a Project integrates any third-party SDK, APIs, data and/or any information belonging to a third party... you must indicate it in your submission description." The codebase contains .claude and .cursor directories and gpt-4o-mini as a fallback model — undisclosed tools including direct Google competitors. Not mentioned.
    The rules state: "A Team must consist of only Eligible Individuals, have all team members added as members of the Project on Devpost." Key contributors visible in the commit history are not listed on the Devpost team. Not mentioned.

  •   •   about 1 month ago

    [3/3]
    The community identified the third-place submission as containing prompt injection targeting evaluators. Not mentioned.
    The rules state judges "may or may not be listed individually on the Hackathon Website." They chose "may not" — no judges panel was ever disclosed. The response states "Humans were involved at every stage of judging, including the finals" but provides no names, no count, and no process details. This is an assertion without evidence.
    The grand prize includes "social promotion of the winning project." Zero social promotion has occurred. No blog post, no news release, no social media announcement. For comparison, the GKE Hackathon (4,773 participants) received a full Google Cloud Blog post. Not mentioned.

    The Winner Verification Requirement:
    The rules state: "THE AWARD OF A PRIZE TO A POTENTIAL WINNER IS SUBJECT TO VERIFICATION OF THE IDENTITY, QUALIFICATIONS AND ROLE OF THE POTENTIAL WINNER IN THE CREATION OF THE SUBMISSION." Was this verification performed? Did it cover the concerns listed above?
    I'm not asking for the result to be overturned. I'm asking why a response titled "Update on the Eligibility Review" doesn't address the majority of the eligibility concerns that were raised — concerns grounded in the hackathon's own rules.
    The community raised far more than two concerns. The community deserves more than two answers.
    RoSP repo: https://github.com/Vector897/RoSP_Hackthon2026

  •   •   about 1 month ago

    [Follow-up]

    Devpost has now given their answer. But this hackathon was not run by Devpost alone.
    The delay updates explicitly stated: "our judging panel includes members of the Google DeepMind team." The prize page promises "social promotion of the winning project." The hackathon carried the Google DeepMind name across 35,000+ participants.
    Omar Sanseviero, Developer Experience Lead at Google DeepMind, hosted the official build + Q&A session for participants: https://www.youtube.com/watch?v=JbwBuFYXl4E
    Google DeepMind has published zero blog posts, zero news releases, and zero social media announcements about the winners. For any other Google hackathon, this would be standard.
    The question is simple: Does Google DeepMind endorse Devpost's eligibility review?
    35,000 developers participated on the strength of the DeepMind name. They deserve a response from DeepMind — not just from the platform administrator.

  • Private user

    Private user   •   about 1 month ago

    At this point just let it go, they won't ever give any explanation(which tbh we ain't even entitled to) that is not AI generated and won't change their mind, just stop wasting time on hackathons and that's that.
    The results were whatever, they literally just got picked by feeding a handful of projects to an LLM and let it choose based on how good the prompts were at fooling it/convincing it they deserved to win(as it was very evident by the 3rd place winner) so next time either just don't join a hackathon or instead of working on the actual project focus on polishing the prompt so it targets whatever criteria they pretend to be evaluating(because I am sure not even the videos were watched)

  •   •   about 1 month ago

    What the person above wrote is 100% true, there were posts by people mentioning that YouTube analytics showed 0 views from any region other than theirs, also the fact that the second place has no backend code but oh well, this is apparently the standard Devpost wants their future hackathon to follow then so be it, tbh might just make an agent to spam @#$% prompt injection projects into every single devpost hackathon from now on... Oh and also professionally... Telling clients to stay the @#$% away from Google and their tech.

  •   •   about 1 month ago

    Since Devpost has confirmed the Grand Prize is well-earned and the team 'made their work transparent,' I'd like to ask a supportive follow-up question:

    When can the community expect the social promotion that is part of the prize package? A blog post from Google DeepMind, a news release, a social media announcement — as was done for the GKE Hackathon, which received a full Google Cloud Blog post with detailed winner profiles for 4,773 participants. Surely 35,000+ participants and $100,000 in prizes deserve at least the same.

    The winners deserve their moment. I'm sure they're looking forward to it.

  •   •   about 1 month ago

    >The RoSP reference: RoSP was the team's earlier project name. Both the RoSP repo (first commit December 25th) and Globot were created inside the Gemini 3 contest window, which opened December 17th. The Globot repo is a snapshot of the RoSP codebase at the moment the team pivoted on December 29th.

    This doesn't prove anything. If you know git well enough, you may know that commit history can be re-written for whatever dates developer wants whenever they want. So the RoSP repo being public doesn't mean anything at all since they can override commit history when you want them to make it public. What was the reason they hide previous repo in the first place if it was already in acceptable dates?
    Also it is obvious that prompt injection used for an another hackathon project. What is their defense or devpost team's explanation ?

  •   •   about 1 month ago

    That's exactly why commit dates alone should never have been used as the basis for an eligibility decision. But even setting the date question aside entirely — even if we accept December 25th at face value — the README still says "Imagine Cup 2026 参赛作品." The demo still says "Powered by Gemini 2.0." The tagline still claims 2M. The model was still swapped on deadline day.

    And Devpost described the 12,000-line first commit as "standard scaffolds." Standard scaffolds are generated by tools like create-vite and FastAPI in seconds. They don't come with DJI drone sales prompts in Chinese, a TELEGRAM_SETUP.md for a sales bot, or an Imagine Cup submission folder. If your scaffolding contains another company's product code and another competition's entry materials, it's not scaffolding. It's a pre-existing project.

    None of these require examining commit dates. They just require reading.

  • Private user

    Private user   •   about 1 month ago

    Don't worry guys, these hackathons aren't for genuine developers or real-world problem solvers like us. They’ve become a theater for whoever can write the most polished, scripted description to impress super dummy performers and organizers who clearly don't know code and have no grasp of real-world technical constraints. When winning is about gaming the system rather than technical integrity, it's clear who this platform is actually built for.

  •   •   about 1 month ago

    Shawni, thank you for the update and for engaging with the community on this. I want to acknowledge that managing a review of this scale — $150,000 in prizes, 35,000+ participants, and significant community scrutiny — is not easy, and I appreciate the effort.
    I'm writing this follow-up not to relitigate what's already been addressed, but to help close the remaining gaps in a way that's constructive for everyone. To make sure the community and I are not misunderstanding the review's conclusions, could you help me understand the following:
    The review confirmed that the RoSP reference and the 12,000-line commit were investigated. However, several other concerns raised by the community — including model version discrepancies in the demo video, the 2M context window claim, undisclosed third-party tools, and the third-place prompt injection issue — were not mentioned in the response. Were these reviewed and found to be non-issues? Or were they outside the scope of this particular review?
    For instance, the community has documented specific items such as DJI-related code in what was described as standard scaffolding, Imagine Cup 2026 submission materials in the RoSP repository, and a model swap PR submitted on deadline day. Were these specific items part of the review?
    If these points fall outside the scope of the eligibility review, a simple confirmation of that would also be sufficient — the community simply needs clarity on what was and wasn't examined.
    If possible, could you respond with numbered points addressing each of the unresolved concerns listed in my earlier comments in this thread? That would make it significantly easier for the community to track which issues have been resolved and which remain open. It would also help prevent further speculation, which I believe is in everyone's interest. It would also be helpful to share the same response in the #gemini-3-hackathon-general-chat Discord channel, as many participants have been following the discussion there as well.
    Thank you again for your time. I believe we all want the same thing — a process that participants can trust.
    P.S. I recognize the irony of asking for structured, specific responses in a thread that discusses prompt injection as one of the concerns. But I trust that a human-led review process can distinguish between a genuine request for transparency and a formatted prompt.

  •   •   30 days ago

    UI and UX Design strategy was carrying a heavy weight in the result, can anyone help me understand how these winning projects are capable to meet this criteria?

  •   •   29 days ago

    its so easy to see its like 5000 project and they saying human no body in this time window can review a 5000 project its doesnt matter if its a valid window time if its created in that time window of the hackthon and then if its a real project or just do a webpage and then some lines and if ita a strong enaugh so the ppl who understand this have gotta know if u have a great idea and do a webpage and do a good video u will be fine and the other once who tryed todo a real programming day and night but have not created that video they gonna sing to u bella ciao why its unfair already from the beginning because they said we will not install or test ur project we will see the youtube video and that's it so why u till us to coded if its about the idea not the developing that idea into the real word that is painly and it hurts and u cant blame any one for this its a new expirence for every body but whate i really ask google specially for this hackthone to start a process to give many devlopers mutch more a price or like google ads reward or any thing like this cause we see in this hackthone many users with mutch more and better work then thos and they is unseen this will make it more fair and more respecting the champoins

  •   •   28 days ago

    A brief follow-up note on the "transparency" claim in the official response.

    The official update on April 30 stated: "the team... made their past repository public... This included the predecessor RoSP repo."
    On 4 May 2026 at 08:30 UTC, the RoSP_Hackthon2026 repository at https://github.com/Vector897/RoSP_Hackthon2026 returned HTTP 404 (archived: https://archive.ph/ODTOW). The repository remains inaccessible. The Vector897 user profile remains accessible, and the Globot repository remains accessible. Only the RoSP repository — the one specifically cited in the official response as transparency evidence — has been removed.

    For independent verification, both repositories were forked prior to removal: RoSP_Hackthon2026 fork at https://github.com/akagb666/RoSP_Hackthon2026 (archived: https://archive.ph/PixlZ), and Globot fork at https://github.com/akagb666/Globot-bkup-evi- (archived: https://archive.ph/ebVO0).

    A cross-reference of last-commit timestamps shows RoSP_Hackthon2026's last commit was January 12, while Globot's last commit was February 10 — one day after the Gemini 3 Hackathon submission deadline.

    The RoSP repository contains a README describing it as "Imagine Cup 2026 参赛作品" (Imagine Cup 2026 Competition Entry), a "submitted_materials/" folder described as 最终 Imagine Cup 交付材料 (Final Imagine Cup Submission Materials), the IC26 Official Rules and Regulations Dec 2025.pdf, and a project folder still named ai-sales-mvp/. Per the README and submitted_materials folder, this codebase was positioned as an Imagine Cup 2026 entry. Between January 12 and February 10, the codebase was developed into the Globot repository, which was then submitted to the Gemini 3 Hackathon.

    This sequence is relevant to the rule cited in my earlier comment: "Projects must be newly created by the entrant during the Contest Period... not a modification or extension of Your or anyone else's existing work."

  •   •   27 days ago

    Keep up the pressure!

    Whether or not they eventually decide to do a fair evaluation or not is not something that we can control - but calling out this @#$% when we all of us put in months of effort and waited a few more months for the result is absolutely a disrespect for every participant. I have been to several hackathons where the entire community praised and applauded the winners and never have I seen a case where everyone clearly knows that the entire judging process was a short cut.

    We should probably take this to LinkedIn and start demanding answers over the points raised. If they do have reasonable justification then its fair and sqaure but until then we should not let this rest. They are probably just waiting for us to give up and let this go.

    @Devpost - If you are arraging a hackthon of this size with proper rules - either commit to it, or keep a max count of projects or just stop hosting hackathons altogether.

  •   •   27 days ago

    And what about @DeepMind? Wasn't a 3rd-place prompt injection enough - is the entire result and review prompt-injected too?

    "Where are your 'humans'?"

Log in or sign up for Devpost to join the conversation.