MachineTranslation.com

Users recently asked something that stuck. "Is there a way to use MachineTranslation.com results directly in Claude or ChatGPT?" The question itself reveals something worth paying attention to. People aren't asking because they want another translation tool. They're asking because they already use Claude or ChatGPT for translation, and they've started to feel the gap. A single model translates confidently whether it's right or wrong. There's no internal check. A shifted register, a flattened idiom, a legal term that doesn't carry the same weight in the target language, it all comes out sounding fine. You'd only know if you speak the language yourself. That gap is what the question was really about. Not features. Verification. Shashank Jain, Technical Lead at Tomedes (translation company that developed MachineTranslation.com), wrote up how the team approached it. Including the technical reasons why running 22 models in parallel doesn't add latency, and what consensus actually means as a quality signal. → https://lnkd.in/gEns_M_z

Wordle proves you know words. LinguaBoard proves you know the world. 🎮 How many languages do you actually know? Try https://lnkd.in/gpakkeQ8 now.

"Save" is one word in English. In Spanish, it becomes five completely different words, depending on what you actually mean. We tested it across 6 contexts on MachineTranslation.com. Here's what happened: 💾 "Please save the file" → guardar (to store) 🆘 "The lifeguard saved the child" → salvó (to rescue) 💰 "I need to save money" → ahorrar (to save financially) 💺 "Can you save me a seat?" → guardar (same word, different meaning) 💬 "Save your breath" → ahórrate el aliento or no gastes saliva depending on which AI you ask 👑 "God save the King" → salve (to bless) The most revealing result was the idiom. ChatGPT translated "save your breath" literally (guarda tu aliento) which misses the meaning entirely. Mistral and Claude used the correct Spanish equivalent idiom, no gastes saliva. SMART surfaced the consensus. This is what contextual AI translation actually means. Not translating the word. Translating what you mean. Try it yourself → machinetranslation.com

We ran two hand-coded studies this quarter looking at what free and paying users actually translate, not just how much. The gap was sharper than we expected. Three findings: 01: The 54× gap Median free user: 12 words per session. Median paying user: 541 words per session. That's a 54× difference. And it holds across the full population, not a small sample. 02: Documents vs. text pastes 60.5% of paying users have uploaded a full document at least once. Roughly 1 in 10 free users have done the same. Paying users aren't just translating more. They're bringing real files. 03: The 150× job A median paying document runs 5,795 words. A median translation across all paying activity runs 38 words. That contract due Thursday isn't a bigger version of a typical translation, it's a different kind of job entirely. 150× longer. This is what drove us to rebuild our pricing around the job, not the subscription. Full breakdown, including what changed for free users → https://lnkd.in/ginPdff5

new LinguaBoard update 🧩 while you play, the sidebar now shows a language quote + a daily "did you know" fact today's fact: Esperanto has over 2 million speakers worldwide puzzle #59 is live → https://lnkd.in/gpakkeQ8

You spent time getting that translation right. The phrasing. The tone. The register. You made it work. Then you accidentally closed the tab. And it was gone. No history. No record. Nothing to go back to. It's No Buy July, so this isn't a sales pitch. Sign up for free on MachineTranslation.com and every translation you run is saved automatically. Your history. Your work. Always there when you need it. No credit card. No trial. No catch.

Boston, 1839. A newspaper editor wanted to write "all correct." He wrote "oll korrect" instead (on purpose, as a joke). Then shortened it to "o.k." 187 years later it's the most recognized word on Earth.

When six AI models translate the same untranslatable word and produce six different outputs, the problem is not the models. The problem is trusting any single one of them. "Kummerspeck" (the German compound for emotional eating weight) produced "grief bacon," "trouble fat," and "stress weight" in a single test. SMART surfaces the translation most models agree on. Not because consensus guarantees correctness. But because in the absence of a perfect answer, agreement is the most defensible standard available. Link: https://lnkd.in/gE7A2VSu

ASMR: things that should be obvious but aren't

"Translation is the circulatory system of the world's literatures." - Susan Sontag LinguaBoard #50 is live. 50 daily puzzles in. Play today → https://lnkd.in/gpakkeQ8