DEV Community: Noah Dunnagan

@grok is this true

Noah Dunnagan — Sat, 11 Jul 2026 02:46:50 +0000

AI is a part of our life. For better or for worse, through writing or media, it is here to stay in some capacity.

And as the astute among you might recognize, content that has been created by these LLMs grows by the day. I can guarantee you with great certainty that one of the next 4-6 posts you read will have been generated by some model in some capacity. As such, some folks wish to know whether they have been "duped" into reading horrible AI text against their will or whether somebody asked claude for a spot of help with the wording.

That same desire for knowledge, has birthed a subset of SaaS products and tools all responsible for a very simple thing. "Is this made by AI?".

And before I name these large companies with tools-a-plenty I'd like to first talk about the root of this desire for knowledge, which in its defense is a noble cause built from frustration and pent up dissatisfaction.

A phrase everyone steeped in the colloquially named "tech twitter" knows as well as a mother knows her child.

It's an artifact of how these models are trained, a process called RLHF, or reinforcement learning from human feedback. You tune a model by showing humans its responses and asking which ones they prefer, do you like your coffee hot or iced? And nudge the model toward more of that. Trouble is, humans reliably prefer being agreed with. We like it when we're right. A lot. And rate the validating answer higher than the correct one. So a phrase like "you're absolutely right" starts becoming the linguistic equivalent of a golden retriever's tail wagging. It fires whether or not you actually are right.

The issue, however, is being agreed with constantly starts to become an issue. This golden retrievers tail is hitting a can, it is loud, make it stop, please. And as such we have started to build this sense of disgust towards a phrase so well known to be a reflex of our AI friends. If your email response starts with "You're absolutely right!", you absolutely did not put in any effort into a response.

Which, in a way is where I agree with these "ai detection" people, those who swear by it day in and day out. If one didn't care enough to write me an email response by hand, why should I care to read it? But with that comes another facet of comparison. Why am I upset at my significant other who brought home takeout and didn't cook it herself? She clearly knows how to cook, heck, pretty good at it too. But instead, she had somebody else do it, mass produced, at the drop of a hat, and is serving this to me instead. And truly I say, it will taste just as good.

However that is a topic for another conversation, I am not here to discuss why you like this food or that. I am not here to tell you and instruct what is morally correct about ordering out or eating in. I am here to talk about why this lie detector test that is held so highly is pointless to run.

And thus we have arrived. Pangram. The epitome of "Is this ai?" detection.

The true issue here is that there is genuinely no way for a service that did not watch you type this text to know whether it was generated or written. That is a fact, physics limits this.

And so Pangram has taken to a step after that, "does this pattern match that of what I know to be AI?" and for most cases it appears to work. But again, those among you who are familiar with this line of tool also know that the declaration of independence rang true as AI generated until it was manually whitelisted by most services.

That alone should add some skepticism. If something that was impossible to be created with AI is being flagged as AI, how can it know that something that could've been made with AI wasn't. You cant. Plainly, truly, and simply cannot reliably tell "is this made with AI" through any programmatic capacity.

Heres some example text I watched a friend write by hand.

Go on, spot the difference. There is one I promise, OCR and scan the content yourself to make sure I didn't edit the HTML for a screenshot. Can't find it? Bear with me and I will show you the difference.

If at first glance it is quite hard to discover what changed, you are correct in thinking that. The change is quite small and not something one would recognize on their daily peruse. However it is there. It is a change. And it is such a small detail.

Below is the diff between the text flagged as "100% AI" and "100% Human". Bear in mind some folks use this number to govern their decision making.

For those who haven't instantly picked it up, a few stray commas have shifted this text's nature entirely to human. 100% at that.

And I feel as though this doesn't need any more explanation. I don't need to instruct you as to why this is a fundamental problem with these tools. If this is governing the grade behind somebody's paper in college, as Pangram so courteously boasts, that is a problem.

So what should you take away from this? That Noah hates pangram and all it stands for? Or loves AI and wants to put it everywhere? Far from it. I am an engineer at heart and I truly respect everyone on the pangram engineering team for creating a readily available tool to scan text and pattern match.

I also do not think we should stop relying entirely on something like Pangram. I think it should be a data point used in a larger decision. One should not use Pangram, GPT Zero, or any other AI detection software as your sole measuring stick as taps sign we have established it isn't 100% accurate.

I will also throw you in for a turn here, I'd like you to go run this article through pangram or GPT zero, in my test as of today, it showed 100% human. However this text had 0 words created by a human. This entire text was generated by AI, specifically Claude Opus 4.8.

That is the issue, the reason I am writing this. I can say that and you will truly never know. Because now you have to decide who to trust more. Pangram an AI which is telling you this text is human or me a human who is telling you its AI. Maybe I am even lying and have drafted this entirely by hand to then claim it was generated. That is something nobody will truly ever know.

Hosting Private AI with Railway

Noah Dunnagan — Thu, 28 Aug 2025 00:17:38 +0000

In today’s world, AI is no longer just a buzzword. It’s part of our everyday lives. Whether it’s ChatGPT, Claude, or Gemini, most of us interact with these tools daily. But while the big names run on massive cloud infrastructure, what if you could host your own private AI?

The cost.

OpenAI pours hundreds of millions of dollars into training and serving its models. Google invests at a scale only tech giants can sustain. Anthropic follows the same path with enormous compute bills and specialized infrastructure. The average developer, company, or small team has no reason to spend like that. What they want is a simple way to run private AI without breaking the bank.

Custom solutions can be expensive and require complex setup. But what if they didn’t have to? What if your solution to having a completely private AI model was as simple as two clicks?

Enter Railway.

What's Railway?

Railway is a cloud platform built for developers. It takes you from idea to deployed application, or even a private AI service, in just a few clicks.

Complexity does not always equal quality. Too often, we overengineer solutions that could be simple. Why spend weeks engineering overly complex infrastructure when the real goal is to ship something that works?

How does Railway come into play?

At first glance it sounds impossible. AI models are massive. Don’t they need racks of GPUs and specialized hardware? The truth is, yes and no. Smaller open-weight models can run entirely on CPUs, and thanks to optimizations like quantization, they can run surprisingly well. Sure, you can’t run the full GPT-5 or Claude Opus 4.1, but models like GPT-OSS match the performance of o3-mini.

Railway provides the environment to deploy and manage these models without you having to worry about infrastructure plumbing.

So the question is no longer if you can run private AI on Railway, but how.

Thankfully, I’m here for that.

How did I deploy?

Well, me and @loudbook had this exact same question. How can we run and host an LLM on Railway with real performance?

Our first attempt involved a custom build of vLLM. It was powerful on paper, but in practice it was fragile, hard to build, and prone to compatibility issues.

Next, I suggested using llama.cpp. It looked promising, but getting it hosted and set up in a way that worked with Railway best turned out to be a very slow and painful process.

Then, the breakthrough. @loudbook suggested simply using Ollama. Deploying was now as simple as running ollama pull <model> and ollama serve. After a few small tweaks, it just worked. We had successfully deployed an open-sourced LLM on Railway. And because Ollama provides an OpenAI API–compatible endpoint, our private model can easily be hooked up to the existing OpenAI SDKs.

The numbers.

Cost.

Regardless of not using GPUs, running LLMs gets expensive.
Depending on the frequency of use, you will be looking at roughly $300–$900 monthly.

That number might sound high at first, but let’s keep it in perspective. Compared to the per-token or per-request pricing of hosted APIs, the cost becomes competitive very quickly if you are doing moderate to heavy usage. For a small team, this could mean hundreds of thousands of tokens per day for a flat, predictable price.

More importantly, the cost is fixed and under your control. You are not metered per request, and you are not paying for someone else’s margin on inference. You are simply renting the resources you need, running the model you want, and deciding how available you want it to be.

If you want a harsh price ceiling, Railway has a tool for that: Usage Limits.

If you want per-service resource caps, Railway has a tool for that: Resource Limits.

Speed.

Cost is only half the story. The other question is speed, because no one wants a private model that crawls along.

On Railway, inference speed depends on two things: the model you choose and the resources you allocate. Bigger models tend to generate slower, smaller ones much faster. In our testing, we saw:

DeepSeek — 10.8 tokens/sec
GPT-OSS — 13 tokens/sec
Llama — 20.5 tokens/sec
Qwen3 — 86 tokens/sec

Now, 10 might sound slow compared to 86, but bear in mind that a typical English sentence is roughly 10–15 tokens. That’s nearly one sentence per second.

For a complete tech breakdown, read this article by @loudbook.

Deploy a model on Railway.

Now that we have the numbers out of the way, here’s how to host a model on Railway.

Currently there are four models of multiple flavors:

Whichever you choose requires the Railway Pro plan.

After picking a flavor, simply click Deploy now. It’s that easy.

Conclusion.

Private AI is possible at an affordable price. In a few clicks, you can host your very own AI model.

The important part is not just that it works. It’s that it’s accessible. You don’t need to be a cloud architect or a Fortune 500 company to have secure, private AI infrastructure. Railway strips away the overhead, leaving you with the freedom to focus on what actually matters: building.