Summary:
Josh Gould, Group CEO of TheBigWord – one of the world's largest language services providers, operating in over 80 countries with a workforce of 15,000 – explains why today's LLMs look more human than 2013's systems without actually being more accurate, how his company's orchestration layer routes each job to the model that is best at that language pair and domain, and why human linguists have shifted from translating to tuning, reviewing and editing while translation demand doubles every year.
Joshua (Josh) Gould has spent his career at the intersection of language technology and human expertise – including TheBigWord's multi-year collaboration with Google on human feedback for neural machine translation. TheBigWord provides translation and interpreting across borders, hospitals, courts, and healthcare systems, giving people language access when they need it most. Beyond language technology, Josh is a co-founder, investor, and board member at a healthy frozen plant-based food company, and a frequent podcast guest on AI, entrepreneurship, and the future of work.
Key Takeaways:
Models are old; scale is new. The architectures behind today's AI existed on paper for decades – chips, electricity, and internet-scale data made them real. The baseline of every model is still human.
LLMs didn't beat 2013. Current models don't score meaningfully better than the best neural Machine Translation of a decade ago – they look more human, while filling gaps with estimations that can make them slightly less accurate.
Orchestration is the moat. No single model is best at everything; TheBigWord's layer routes French–Dutch orthopedic content to one engine and English–Russian general content to another, tuned continuously by client and linguist feedback – pushing accuracy from ~94% toward 98–99%.
The linguists didn't disappear – their jobs changed. Translated content demand doubles every year; unit rates fell, volumes exploded, and translators became editors, reviewers and tuners. The predicted extinction never happened.
Tokens are the hidden bill. Interpreting (voice-to-voice) consumes six to ten times the tokens of written translation, and today's AI prices are loss-making subsidies: expect roughly a doubling once providers must be profitable – machine interpreting at 75% of human cost today could be 150% tomorrow.
Regulation helps the big and hurts the small. Compliance infrastructure is a sales advantage for a firm with 15,000 people and a burden that can put a bedroom startup in legal jeopardy; meanwhile European governments quietly race to adopt AI and will likely be forced to soften their own rules.
The football-team principle. If you're allowed eleven humans on the pitch and unlimited robots, you field eleven humans plus robots. AI supersizes people rather than replacing them – and at TheBigWord, that's a stated commitment.
Chapters:
00:04 The Evolution of Machine Translation
02:58 Moses, Hybrid Models and the Road to Neural MT
07:46 How Google Changed the Accuracy Game
13:17 The Role of Human Linguists in AI
17:30 Building the Orchestration Layer
20:21 The Real Cost of AI Tokens and Interpreting
22:36 The Economic Value of Immigration and Language Access
26:43 Cultural Nuances in Translation
31:05 Navigating Global AI Regulation
39:42 Regulatory Challenges in AI and Translation
Hyperlinks:
Anastassia Lauterbach - LinkedIn
First Public Reading, Romy, Roby and the Secrets of Sleep (1/3)
First Public Reading, Romy, Roby and the Secrets of Sleep (2/3)
First Public Reading, Romy, Roby and the Secrets of Sleep (3/3)
