IA
Uncensored AI models: a practical evaluation
November 18, 2026
Why it matters
An unnecessary refusal can interrupt legitimate technical work. Developers encounter this when documenting their own software, examining synthetic security logs, or researching sensitive subjects in an authorized environment. A model with fewer refusals might remove that friction. The useful question is whether its answers actually survive inspection.
Jack Roberts demonstrates three arrangements: a hosted API, a model running on his computer, and a server intended to keep the system available around the clock. Those arrangements change different things. Response policy affects what the model will attempt. Deployment affects where processing happens. Agent permissions affect what an answer can turn into.
I would evaluate those properties separately. The demonstrations belong to the video’s author; the procedure below is my proposed way to assess the underlying idea. It starts with a practical obstacle and asks whether modifying refusal behavior reduces the work required to reach an acceptable result.
Key points
- 01:39 Abliteration targets refusal behavior. The video describes modifying a model so it declines fewer requests. The foundational research identified an internal direction associated with refusal in the models studied. Its findings do not establish identical performance for every subsequently modified model. Arditi and colleagues’ paper.
- 02:11 The first route uses a hosted API. Abliteration.ai provides the service shown. This removes the need to load model weights on the user’s computer, while requests still travel to an external system for processing.
- 03:16 Application prototypes provide the visual demonstration. Roberts requests interfaces inspired by Granola and Screen Studio, using Reverse Engineer Anything. The screens illustrate prototyping potential; the demonstration does not establish complete functional parity with those applications.
- 04:55 Model substitution is paired with a cost comparison. The author reports runs costing $0.63, $0.32, and $0.55. Those are figures from his examples. A useful purchasing comparison would also require shared acceptance criteria and a record of corrections.
- 06:02 Setup begins with an API key and prepaid credit. Roberts recommends starting with a small deposit and points viewers toward a Starter Pack. A bounded initial budget makes experimentation easier to assess.
- 07:07 The surrounding agent matters. Agentic OS and DeepSeek Harness appear as ways to manage interaction. Memory, tools, and execution settings also shape outcomes, so replacing a model does not automatically preserve the agent’s previous behavior.
- 08:13 Local execution is introduced as a privacy option. The video mentions private documents and research. The relevant boundary is the complete workflow: local inference can still sit alongside remote voice processing or external integrations.
- 10:56 Ollama supports the local setup. Hardware capacity informs model selection. The Qwen 3.8 27B distribution cited occupies roughly 18 GB in the catalog checked; download size alone does not specify the memory required during operation. Official model listing.
- 12:03 Deployment and refusal behavior become separate axes. The quadrant distinguishes local from remote execution and standard from modified models. Downloading a standard local model does not make it abliterated.
- 13:09 Fish Audio adds speech output. The demonstration includes voice synthesis and switching voices through an API. That additional service belongs in both the budget and the map of external data processing.
- 15:50 A server addresses continuity. The goal becomes keeping the system available after the laptop closes. Hosting the agent application and running a large model’s weights are separate infrastructure requirements.
- 20:17 Tailscale supplies remote connectivity. It connects devices to the VPS. Its encryption protects traffic between devices; application permissions and server-side handling of information remain separate configuration concerns. Tailscale’s encryption documentation.
How to apply
1. Evaluate uncensored AI models against a specific obstacle
Start with a legitimate request that your current model refuses. Use material you control: a synthetic log, a small program you wrote, or a fictional document covering a sensitive topic. Write down what a useful answer would contain before trying another model.
For a meaningful comparison, use a standard model and an abliterated derivative of the same release whenever possible. Keep the instructions, input files, tool access, and execution limits consistent. Check whether both use the same weight precision or quantization, since that difference can affect performance and memory consumption.
A provider may not disclose enough information to build that matched pair. You can still compare its service with your current option, but record the limitation. A comparison across unrelated models cannot isolate the effect of abliteration.
2. Build a twelve-case evaluation set
Choose four routine tasks, four legitimate requests that previously triggered refusals, and four incomplete instructions. The last group matters because asking a follow-up question can be the correct behavior. For example, a request to change a file should require clarification if the target file is missing.
Define a check for each case. An extraction task must recover specified fields. A code change must produce an observable behavior. An explanation must identify the evidence supporting its conclusion.
Record the same information for both candidates:
| Measure | What to capture | |---|---| | Expected behavior | Answer, clarify, or request approval | | Unnecessary refusal | Was a legitimate task blocked? | | Correct outcome | Did the result pass the predefined check? | | Human intervention | Minutes spent correcting or redirecting the model | | Cost | API consumption and review time |
Repeat cases that produce inconsistent outcomes before selecting a candidate. This small evaluation can justify a limited pilot. It cannot establish reliability across every future request.
3. Trace the complete data path
Map the actual sequence: keyboard or microphone → interface → memory → model → tools → text or audio response. Mark every external service. Use fictional information during this first pass so you can inspect behavior without introducing confidential material.
Remote processing and permanent retention are different properties. Abliteration.ai states that it does not retain prompts or responses. Consider that statement alongside the provider’s terms when assessing the service; the inference still takes place outside your computer. Provider privacy statement.
For an Ollama local setup, verify the selected model and test the workflow without a network connection after downloading it. Restore integrations individually. This makes it easier to identify which features require external processing and which genuinely work offline. Add voice only after the text workflow behaves as expected.
4. Size infrastructure around the component it will run
The VPS section needs a hardware check before purchase. The KVM 2 plan lists 8 GB of RAM, while the cited Qwen distribution is approximately 18 GB. File size is not a complete memory calculation, but these specifications do not support assuming that this plan will run that model at useful performance. Hostinger specifications, Qwen distribution.
Choose the architecture explicitly. A VPS can host an agent that calls an API. A machine you own can remain powered on and provide inference remotely. Alternatively, rented hardware can be sized for the model itself. Each arrangement has different dependencies and operating costs.
Include a restart in the pilot. Check whether the application returns, retains the state it needs, and resumes appropriately. Remote access is only one part of keeping a system usable over time.
5. Use reduced intervention as the decision rule
My threshold for continued use would be fewer unnecessary refusals without an increase in errors or human intervention. Add API usage, hosting, voice services, and review time to the comparison. If the work is billed in Brazilian reais, record the actual charge in that currency rather than relying only on dollar figures from a demonstration.
Run the initial agent in a disposable project with limited permissions. Inspect its outputs and changes before expanding access. A model that attempts more tasks may still create more corrective work.
The outcome I would look for is narrow and measurable: one recurring task where the modified model produces acceptable results with less supervision. That is enough to justify a focused pilot and gives the next evaluation a concrete starting point.
Resources mentioned
- abliteration.ai (paid) · API comercial de modelos com menos recusas; serviço demonstrado no nível 1.
- Claude Code (paid) · Ambiente de agente usado na demonstração; custos dependem da configuração e do provedor.
- Reverse Engineer Anything — REA · Repositório aberto usado nos protótipos; modelos e infraestrutura podem gerar custos.
- DeepSeek Harness · Alternativa aberta apresentada para executar agentes; consumo de modelos é separado.
- Ollama · Ferramenta apresentada para baixar e executar modelos localmente.
- Qwen 3.8 · Modelo local citado; a distribuição padrão não equivale automaticamente a uma variante abliterada.
- Fish Audio (paid) · Síntese e personalização de voz; integração por API demonstrada no vídeo.
- Hostinger VPS (paid) · Hospedagem demonstrada no nível 3; dimensione aplicação e inferência separadamente.
- Tailscale · Conexão privada entre dispositivos; oferece plano pessoal gratuito e planos pagos.
- Obliterated Model Starter Pack · Gratuito segundo o vídeo, mediante entrada na comunidade; URL oficial direta do pacote não confirmada.
- Agentic OS + curso e comunidade (paid) · Comunidade oficial de Jack Roberts; a página apresenta curso e Agentic OS.
- Glaido (paid) · Ditado recomendado na descrição; possui plano gratuito limitado e versões pagas.