Jack Roberts demonstrates three ways to work with less-refusal-prone AI in his tutorial: call a hosted model API, run a model on your own computer, or keep an agent available through a remote server. He connects those ideas to DeepSeek Harness, Ollama, Qwen, Reverse Engineer Anything and Tailscale. The most useful lesson is not that a model will "do anything." It is that where the model runs, where the agent runs, and what tools it can reach are three separate decisions.
Start with a hosted API if you need to compare capable models without buying hardware. Use Ollama locally when your computer has enough memory and your data should stay on that machine. Use a VPS plus Tailscale when you need an always-on agent you can reach remotely. A small CPU VPS can host the agent or call a model elsewhere; it is not automatically capable of running the large model shown in the video. None of these routes removes the need to review tool access, data handling and costs.
Watch Jack Roberts' Tutorial
The video promotes Hostinger, Fish Audio, a starter pack and a paid course/community. Discount codes, prices and model access can change; the setup guidance below links to the vendors' current documentation rather than treating the offers as technical evidence.
Source Note: Demo Versus Documentation
This article uses the supplied transcript and the video. On 8 November 2026, we checked the official Abliteration API docs and pricing, Ollama download and model page, Hostinger's KVM limits, Tailscale installation, and the named GitHub repos. We did not reproduce Jack's app builds, benchmark the models, audit the hosted provider's data practices or verify his quoted project costs. The linked short URLs are creator offers, not independent documentation.
Three Routes, Different Trade-offs
| Route | What runs where | Best first use | Watch for |
|---|---|---|---|
| Hosted API | Model inference on a provider's servers; your agent may run locally. | Quickly compare models on a small, non-sensitive task. | Prompts and files leave your machine; rates, provider policy and key security matter. |
| Local Ollama | Weights and inference on your computer. | Offline or lower-exposure experiments when hardware fits. | The model may fit on disk but not in usable RAM/VRAM; connected agent tools can still send data out. |
| VPS plus Tailscale | An always-on agent on a rented server; model inference can be local, on another machine or via API. | Reach an agent or home model from a phone without opening a public model port. | A VPS is not your computer and is not necessarily a GPU host. Review network permissions, recurring bills and backups. |
What an Abliterated Model Actually Changes
"Abliteration" describes a family of weight-editing techniques intended to reduce a model's tendency to refuse certain requests. That is narrower than removing all guardrails. Results depend on the base model, the edit and the task; a lower refusal rate does not prove equal quality or accuracy. The provider itself describes a hosted less-refusal-prone model alongside a policy layer. Jack's "no refusals" and cost comparisons are demonstrations and estimates for his selected projects, not a general promise.
His app recreations are also a reminder that capability and rights differ. Reverse Engineer Anything explicitly frames its use around lawful research, interoperability, education and software a user owns or may examine. Building a new tool inspired by a workflow is different from copying a competitor's code, branding, assets, protected data or paid access. Keep the project within your rights and document the boundaries before turning an agent loose.
Route 1: Try a Hosted API
Jack uses abliteration.ai as his hosted example and shows a model used through an agent harness. The service's docs describe OpenAI- and Anthropic-compatible API surfaces; its live pricing separates input, cached input and output tokens, with prepaid credit available. His $63/$32/$55 project figures should be read as his own run costs, not a controlled comparison with an identical Opus run. Check current pricing before funding an account.
- Create a provider account and inspect its model names, rate limits, privacy terms and billing controls.
- Start in the playground with a harmless prompt. Add a small prepaid balance only when you understand the charges; leave automatic top-up off for the first test.
- Create an API key and store it in the harness's secret store or an environment variable, never in a prompt, screenshot or repository.
- Connect the harness using the provider's documented endpoint. Give it a small local test directory and review its tool permissions before running a build.
- Compare a fixed task against your current model, recording output quality, corrections, time and actual charges.
DeepSeek Harness is an official open-source agent harness, but installing it is a separate choice from choosing a model provider. A hosted API still receives the task content you send it, regardless of which harness sits in front.
Route 2: Run a Model on Your Computer
Download Ollama for your OS, then use its quickstart to pick a local model. On a machine that can support it, the official Qwen3.8 page gives ollama run qwen3.8. Its 27B default package is about 18 GB on disk; runtime memory and context need additional headroom. Do not infer that an 18 GB file will work well with exactly 18 GB of RAM or VRAM. If it does not fit, select a genuinely smaller model from Ollama's library rather than forcing this one.
- Check free disk space, system memory and GPU/VRAM. Note whether you are using CPU only.
- Install Ollama from the official download page and open its app or terminal.
- Choose a model whose published file size and expected runtime memory fit your machine. Pull and run it; start with a short prompt and measure speed and memory use.
- For an agent, follow Ollama's coding-agent integration or its Anthropic-compatible API docs. Verify which model the agent actually calls.
- Test with synthetic files first, then check every connector, voice service and web tool. "Local model" does not mean the entire workflow is offline.
Ollama also offers cloud models; its docs distinguish those from local downloads. If you select a cloud model, prompts go to Ollama's servers. Jack's Qwen3.8 demonstration is a local-model example, not proof that all model options or optional voice features stay on-device.
Route 3: Keep an Agent Available Remotely
Jack's third level uses a Hostinger VPS and Tailscale so he can reach an agent when his laptop is closed. This can be useful, but the hardware claim needs a reality check. Hostinger lists KVM 2 at 2 vCPU, 8 GB RAM and 100 GB disk, with no GPU allocation in that plan table. Qwen3.8 27B is about 18 GB just to download. KVM 2 is therefore a reasonable candidate for a lightweight agent, remote interface or API client, not a sensible default for hosting that 27B model. It could call a hosted API or a separately provisioned home/GPU machine through a private connection.
- Choose a VPS only after deciding which component will run there. Compare the live plan's RAM, CPU, storage, region, renewal price and backup options.
- Provision an up-to-date Ubuntu image and set a unique admin credential. Keep SSH and system packages updated; use a separate, least-privileged account for the agent.
- Install Tailscale on the VPS and the device you will use to connect. Follow Tailscale's Linux instructions and authenticate both into your own tailnet.
- Confirm the VPS is reachable on its tailnet address. Keep any model API and agent web UI off the public internet unless you deliberately add authentication, TLS and access controls.
- Start the agent with a low-risk task, a spending limit and backups. Confirm what happens after a reboot and how you revoke access if a device is lost.
Tailscale protects the connection between authorized devices; it does not decide whether the agent may read a file, spend money or send a prompt to a third-party model. Those permissions belong to the agent and service configuration.
Tools by Role, Not by Hype
| Tool | Job | Practical caveat |
|---|---|---|
| Abliteration | Hosted less-refusal-prone model API. | Cloud inference: check data terms, provider policies and token bills. |
| DeepSeek Harness | Agent interface with tools and plugins. | Its model backend is a separate configuration; inspect plugins and approvals. |
| Agentic OS / Jarvis | The creator's voice-led agent interface and course/community. | A separate creator product, not required to run the models or Ollama. |
| Reverse Engineer Anything | Inspect authorized software behavior and binaries. | The repo does not grant rights to third-party code, assets or licenses. |
| Ollama + Qwen3.8 | Download and serve a model locally. | Select a size your RAM/VRAM can actually run; distinguish local from cloud tags. |
| Tailscale | Private-network access to another device or VPS. | Network access is not agent authorization or a data-retention policy. |
| Fish Audio | Cloud speech, transcription and voice generation. | Using its API sends audio/text to a service; clone only voices you have rights and consent to use. |
Copyable Setup Brief
Use this before pasting a starter pack into an agent. It asks for the real data path and hardware fit, then stops before purchases or credential changes.
Choose a route that fits
Compare hosted, local and remote setups on one approved task.
Sources and Creator Offers
| Role | Links | What to check |
|---|---|---|
| Video and creator material | Full tutorial · starter pack · course/community | The video demonstrates Jack's workflow. Short links and offers may change; review their destinations and terms. |
| Models and APIs | Abliteration docs · pricing · Qwen3.8 on Ollama | Current model IDs, compatibility, file size and charges. |
| Agent tools | DeepSeek Harness · REA repo | Install instructions, tool permissions, updates and lawful-use boundaries. |
| Remote hosting | KVM limits · Tailscale install · creator's Hostinger offer (code JACKROBERTS) | Actual hardware, eligibility and renewal price; the promotion is not a hardware specification or savings guarantee. |
| Voice and extras | Fish Audio docs · creator's Fish link · Glaido offer · free resources | Consent, audio data path, fees and offer conditions before connecting a real voice workflow. |
Video Chapters
The idea and the hosted route
| Time | Segment |
|---|---|
| 00:00 | Uncensored AI |
| 00:39 | Three levels and warnings |
| 01:52 | What is an abliterated model? |
| 02:11 | Level 1: hosted API |
| 03:00 | Linear recreation |
| 03:22 | Granola and Screen Studio demos |
| 04:52 | Changing the model in a harness |
| 05:11 | Cost comparison |
| 05:53 | abliteration.ai setup |
| 06:21 | Starter pack |
| 06:54 | Reverse Engineer Anything |
| 07:10 | Agentic OS and DeepSeek Harness |
Running on your own machine
| Time | Segment |
|---|---|
| 08:07 | Level 2: local |
| 09:19 | Local Jarvis demo |
| 10:08 | Comparing local model capability |
| 10:50 | Install Ollama |
| 11:17 | Choose a model |
| 12:10 | The local versus hosted matrix |
| 13:00 | Fish Audio voice |