Open WebUI is a web app for chatting with LLMs, like chatgpt.com or claude.ai. I’ve run it for the last 6 months, and I’m done with it. I’m building my own replacement.

Why I chose it

I wanted an open-source and self-hostable UI, so I don’t support the big AI companies. I chose Open WebUI because it seemed like the most feature-rich and least resource-hungry web UI for LLMs. The alternatives didn’t fit. LibreChat requires a stack of 6 containers. I wanted a single container. Most other alternatives are native desktop apps. I wanted something that syncs between my phone and my laptop.

It’s a resource hog

Open WebUI takes nearly 1GB of RAM on my 2GB VPS. Its backend is Python, and its frontend is a single-page app, and both are heavy.

It also bundles Whisper (speech-to-text) and embedding models by default, which makes even the slim container image nearly 1.5GB.

The bugs are awful

Over the last few months, I’ve hit some insane bugs.

• Model selector clears itself when you create a new chat
• Exponential growth of data
• The UI hard-refreshes on it’s own, despite being a SPA app
• Reasoning blocks aren’t rendered

The defaults and UX are awful

Without the bugs, I still hate the defaults. The configuration is unmanageable. Each model has to be configured independently.

When you start a chat, it tries to connect to every MCP server. Why? Some are slow, and I have a lot of them.

You can only have one Web Search provider at a time. I bounce between Kagi and Exa because they behave differently, so I added them as MCPs instead, which lets me use both.

I have to configure each model I use or switch to. Every time I try a new model, I configure it to have web search, each of my MCPs, Code Execution, and Native Function Calling. These are all disabled by default, and can only be enabled on one model at a time. It’s such an unsustainable way to manage your configuration!

The buttons don’t make sense.

CMD + R means “refresh the page”, right? Wrong! It resends your last message!

Chat search does not work

I’ll often search the sidebar chat for something and can’t find the one I need.

What am I going to do about it?

I won’t pay a big AI company for this — they already have plenty of money, and I don’t like the politics of any of them.

I tried looking at the open-source or smaller alternatives, and I’m dissatisfied.

You have the desktop apps like Jan or AnythingLLM, which aren’t versatile and are unavailable on my phone.

Web-hosted apps like Open WebUI or LibreChat are bloated.

I want something self-hostable, on the “cloud”, and lightweight. I’ll use an LLM provider in a data center, because it’s faster and more energy efficient.

So going to build my own: I’ll only use cheap open-weight models, and vibe-code my ideal LLM harness.

I was trying to set up a new retool instance on docker when I encountered a 401 Unauthorized error:

$ docker compose up -d
[+] Running 3/3
 ✘ retooldb-postgres Error failed to resolve reference "docker.io/library/postgres:15...                          0.8s 
 ✘ https-portal Error      context canceled                                                                       0.8s 
 ✘ postgres Error          context canceled                                                                       0.8s 
Error response from daemon: failed to resolve reference "docker.io/library/postgres:15.12": failed to authorize: failed to fetch oauth token: unexpected status from GET request to https://auth.docker.io/token?scope=repository%3Alibrary%2Fpostgres%3Apull&service=registry.docker.io: 401 Unauthorized

I had just signed up for a new docker account and had not verified my email yet. Docker will return 401 Unauthorized if your email is not verified.

I’ve hit this issue a few times before, always after starting a new job. Quite frustrating!

I found a poster with a math puzzle that points to a reward website. I couldn’t figure it out, so maybe the prize is still available!

Here it is all typed out:

prefix(
    [
        3!
        + argmax(
            softmax(
                [log(2), pi + e]
            )
        )
        + log*(16)
    ] ^ [
        BusyBeaver(4)
        + ⌈H100 flops / A100 flops⌉
        + 1
    ]
)

.ai/givemeprize

Stop here if you want to work on the puzzle yourself!

My Breakdown

Below is everything I figured out before I got stuck.

• argmax refers to numpy.argmax returns the index of the maximum in an array
• softmax is more interesting. Wikipedia’s definition is complicated, but pytorch explains it concisely around normalizing probabilities to sum to 1.
• log is maybe log₂ because of the computer context, maybe log₁₀, but probably log_e because numpy and pytorch both assume log_e.
• log* is confusing. I’ve never seen that syntax before. Is that natural log, aka log_e? I’m guessing it’s log₂ because the argument is 16, which makes a nice and round answer of 4, like the rest of the problem.
• BusyBeaver(4) is either the maximum shift 107 (likely) or the maximum score 13 (original)???
• H100 flops and A100 flops are the flops on the H100/A100 Nvidia chip. There are so many variations. But “H100 flops/A100 flops” is surrounded by upper brackets, which is ceil. Thus, I only need a best guess. Running through a few numbers, I get around 3.04 to 3.5, so ceil is 4.
• prefix is weird, but I’m guessing it’s the integer portion of the number, not the decimal portion.

Here’s my reduced form:

prefix(
    [
        6
        + argmax(
            softmax(
                [0.693147181 , 5.85987448]
            )
        )
        + 4
    ] ^ [
        107
        + 4
        + 1
    ]
)

.ai/givemeprize

I don’t need to run softmax, because I know the second number will be bigger and argmax will return its index. Argmax is zero-based, so that reduces to one.

prefix(
    [6 + 1 + 4] ^ [107 + 4 + 1]
)

.ai/givemeprize

Again, reducing by simple math:

prefix(
    [11] ^ [112]
)

.ai/givemeprize

Maybe the answer was 11.ai/givemeprize, but now it 404’s. They’re not based in SF, so I doubt it.

Open questions:

• What is a prefix function that takes only one number?
• Should BusyBeaver(4) be 107 or 13? 107 is very large for an exponent, but it’s technically possible. 11 ^ 112 is 112 digits long, and the maximum domain name length is 253 characters long.
• What is log*? I could try log₂ or log₁₀. log_e is also possible, but I’ve only seen it written as ln.

That leaves these options:

I considered looking at recently registered domain names. I could filter by domains only containing numbers and under the *.ai TLD. I could hit them all at /givemeprize and see who responds at all and without 404. Unfortunately, that information is behind paywalls.

I love puzzles, but I’m done with this one. I don’t think I get it. Maybe I missed something.

The Last Bits

Someone left the correct answer in the comments, but didn’t show any of the work, which is frustrating cause I want to understand!

However, folks at lobste.rs guided very kindly and I figured out my mistakes!

• log* is iterated log. I could have learned this if I used ChatGPT or WolframAlpha instead of Google.
• BusyBeaver(4) is 13, my less likely answer
• this makes the exponential 10^18
• prefix in the context of 10^18 is the SI prefix, like EXAbyte or TERAflops. So the answer here is exa.

I somehow got a lot more traction on HackerNews way after I posted.