LLM
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.
Read More
The Post-MT Enhancement Pipeline Your Competitors Don't Have
Machine translation output in 2026 is good. It’s not good enough.
Raw MT from DeepL, Google, or any major engine produces text that’s usually comprehensible but rarely polished. The translations are technically correct but feel off: wrong register, inconsistent terminology, awkward phrasing that no native speaker would use.
Most organizations address this through human post-editing. Translators review MT output, fix problems, and produce final text. This works, but it’s essentially using humans to compensate for systematic MT limitations—the same limitations, segment after segment, project after project.
Read More
Context-Aware Translation: Why Surrounding Segments Matter
Translation tools divide content into segments—usually sentences. Each segment gets translated independently. Segment 47 is processed without awareness of segments 46 or 48.
This architecture is efficient. It parallelizes well. It maps cleanly to translation memory matching. It’s been the industry standard for decades.
It also produces translations that feel fragmented.
The coherence problem
Read a professionally translated document closely. Often you’ll notice:
Pronoun inconsistency. “It” refers to different things in adjacent sentences, but the translations don’t maintain the reference chain.
Read More
Translation Memory Meets AI: The Hybrid Approach
Translation memory is a solved problem. You translate a sentence once, store it, and reuse it when the same sentence appears again. 100% matches translate instantly, at zero cost, with guaranteed consistency.
The trouble is with everything that isn’t a 100% match.
The fuzzy match problem
Real content evolves. A sentence that was “Contact our support team” in version 1 becomes “Contact our customer support team” in version 2. The TM has a 90% match. What do you do with it?
Read More
Style Guide Automation: AP, Chicago, Duden Without the Manual Work
The AP Stylebook runs over 600 pages. The Chicago Manual of Style exceeds 1,100. Every translation into English should theoretically conform to one of these standards—but no translator holds hundreds of rules in active memory while working.
Style guide adherence in translation has traditionally been aspirational. Teams declare they follow AP or Chicago, reviewers catch obvious violations, and countless small deviations slip through because no one can enforce that much detail manually.
Read More
LQA Automation: The End of Manual Error Hunting
Quality assurance in translation has a fundamental problem: it requires humans to read everything.
Linguistic Quality Assessment (LQA) evaluates translation quality through systematic error detection and categorization. A reviewer reads each segment, identifies problems, classifies them by type and severity, and scores the overall quality. This produces valuable data about translation performance.
It’s also exhausting, time-consuming, and doesn’t scale.
The review fatigue problem
LQA reviewers face cognitive challenges that undermine quality:
Read More
Translation QA Metrics That Actually Matter
“The translation scored 92%.” What does that actually mean?
Quality metrics in translation suffer from a fundamental problem: they’re often abstract numbers disconnected from what matters. A project can score well on mechanical metrics while producing translations that don’t work for their intended purpose. Or score poorly on pedantic criteria while delivering effective communication.
Better metrics connect quality assessment to actual outcomes.
The problem with single-number scores
A single quality score collapses complex information into one figure:
Read More