Audio Processing

AI safety beyond the frontier labs: uncensored local models

A recent Guardian report showed researchers asking frontier models to help plan violence and crime. Frontier labs (OpenAI, Google, Anthropic) responded that they’d already employed better safeguards since the research concluded. Ignoring the fact that you can still work your way around safeguards in many cases, they make no mention of local models.

I’ve experimented with highly capable local models - both from Google and from Alibaba - Gemini and Qwen - and found them to have similar safeguarding to hosted models. But then I checked the “Uncensored” versions of the same models, where researchers and enthusiasts take the base model and surgically isolate and remove the weights relating to guardrails and alignment. They report that the models answer 100% or 95+% of the unsafe questions asked of it, vs 1-5% before the further work.

Developer Agency in the Age of AI

As a bit of background, I have written about performance and security in posts that hit the front page of Hacker News, long before LLMs were in use. I’ve been setting up and securing Linux machines and servers for decades. I was writing in QBasic, a menu-driven “OS”, then simple games in C++ from around 12 (i.e. in the 90s). I am a big proponent of open source (FOSS)¹ and software anyone can run and work on, enabling me to have enterprise capability at no cost or low cost because I learned how to set it all up securely and to run at scale. I’ve brought this knowledge, of course, into the LanguageOps project. This is one area often missing from AI developed projects. People tend to underestimate the effort to operationalize a project.

AGI v1.0 Will Be Harness + Skills

Consider the trajectory of LLMs so far. Starting in Google’s translation research departments, following on from their groundbreaking neural network models, transformers were first and foremost a “language” model. They have grown to be able to convince people of their utility, much like the market stall owner would do pre-2000s, holding groups of up to 50 people in awe because of their fast talking, confidence and skill demoing whatever they were selling. The same works on shopping channels and more recently Youtube. Speak fast and confidently, cut all the dead air, and you got yourself the makings of a popular channel.

Council Translation: How Multi-Model Consensus Beats Single-Engine Output

Every machine translation engine has blind spots. DeepL struggles with certain idioms. Google Translate sometimes loses register. LLMs occasionally hallucinate. When your translation workflow relies on a single engine, you inherit all of its weaknesses.

The localization industry has accepted this as an unavoidable cost of automation. Post-editors spend hours fixing the same predictable errors, project after project. Quality teams develop mental checklists of “things to watch for” with each engine. It’s become so normalized that most platforms don’t even question it.

Cross-Lingual QA: Catching Errors Without Reading the Target Language

Translation quality assurance has a staffing problem. Finding reviewers who are native speakers of the target language, fluent in the source language, and expert in the subject matter is difficult. Finding them for 40 language pairs is nearly impossible.

Most organizations solve this by accepting lower review coverage. High-value languages get thorough QA. Lower-volume languages get spot checks or statistical sampling. The unstated assumption: some markets will receive lower-quality translations than others.

Video Localization Beyond Subtitles: The Dubbing Revolution

Subtitles are a compromise. They’re how the localization industry has handled video content for decades because real dubbing—with voice actors, recording studios, and audio engineering—costs too much for most content.

The numbers tell the story. Professional dubbing runs $75-150 per minute of finished video, assuming you already have a script. A 10-minute product demo costs $750-1,500 to dub into one language. Multiply by 10 languages and you’re looking at five figures for a single video asset.