15 subscribers
Mettez-vous hors ligne avec l'application Player FM !
Podcasts qui valent la peine d'être écoutés
SPONSORISÉ
#218 How Large Language Models Replace Neural Machine Translation with Unbabel’s João Graca
Manage episode 428499235 series 2975363
João Graça, Co-founder and CTO of language operations platform Unbabel, joins SlatorPod to talk about the present and future of large language models (LLMs) and their broad impact across all things translation and localization.
First, the CTO explains how Unbabel was founded to address language barriers for people using services like Airbnb, combining MT with human validation to improve translation quality.
João believes that LLMs are quickly replacing neural MT models as much more R&D is going into LLMs vs NMT. He highlights that LLMs can handle more complex tasks like automatic post-editing, source correction, and cultural adaptation, which were previously difficult to achieve with traditional models.
He also tells the backstory of the company's decision to develop TowerLLM. João shares how Unbabel's approach involves using open-source LLMs, fine-tuning them with multilingual data, and applying techniques like retrieval-augmented generation to improve translation quality in production settings.
Despite the advancements, João acknowledges that human intervention is still necessary for high-stakes translation tasks.
The podcast concludes with the hiring environment for AI talent and the future directions for LLM development, with João expressing optimism about the continued progress and potential of these models.
Chapitres
1. Intro (00:00:00)
2. Background and Motivation Behind Unbabel (00:00:34)
3. Research Contributions (00:04:13)
4. NLP and LLM Impact (00:07:10)
5. RAG Approach (00:09:12)
6. Adapting Production Processes (00:11:04)
7. Evaluating Model Usage (00:12:42)
8. Evolution from Neural MT to LLMs (00:13:56)
9. Comparing Price (00:15:44)
10. Why Unbabel Decided to Build TowerLLM (00:16:43)
11. TowerLLM Development Process (00:18:49)
12. Multilingual Model Performance (00:23:07)
13. Model Usage and Commercial Restrictions (00:25:25)
14. Quality Testing Process (00:26:24)
15. TowerLLM Challenges (00:29:20)
16. Future of Translation Technology (00:30:19)
17. Areas of Application for LLMs (00:32:09)
18. Understanding xTOWER (00:34:49)
19. AI Pipelines (00:37:23)
20. Language Coverage (00:38:59)
21. Hiring Environment (00:40:43)
22. Acceleration of LLMs and AI Progress (00:41:51)
237 episodes
Manage episode 428499235 series 2975363
João Graça, Co-founder and CTO of language operations platform Unbabel, joins SlatorPod to talk about the present and future of large language models (LLMs) and their broad impact across all things translation and localization.
First, the CTO explains how Unbabel was founded to address language barriers for people using services like Airbnb, combining MT with human validation to improve translation quality.
João believes that LLMs are quickly replacing neural MT models as much more R&D is going into LLMs vs NMT. He highlights that LLMs can handle more complex tasks like automatic post-editing, source correction, and cultural adaptation, which were previously difficult to achieve with traditional models.
He also tells the backstory of the company's decision to develop TowerLLM. João shares how Unbabel's approach involves using open-source LLMs, fine-tuning them with multilingual data, and applying techniques like retrieval-augmented generation to improve translation quality in production settings.
Despite the advancements, João acknowledges that human intervention is still necessary for high-stakes translation tasks.
The podcast concludes with the hiring environment for AI talent and the future directions for LLM development, with João expressing optimism about the continued progress and potential of these models.
Chapitres
1. Intro (00:00:00)
2. Background and Motivation Behind Unbabel (00:00:34)
3. Research Contributions (00:04:13)
4. NLP and LLM Impact (00:07:10)
5. RAG Approach (00:09:12)
6. Adapting Production Processes (00:11:04)
7. Evaluating Model Usage (00:12:42)
8. Evolution from Neural MT to LLMs (00:13:56)
9. Comparing Price (00:15:44)
10. Why Unbabel Decided to Build TowerLLM (00:16:43)
11. TowerLLM Development Process (00:18:49)
12. Multilingual Model Performance (00:23:07)
13. Model Usage and Commercial Restrictions (00:25:25)
14. Quality Testing Process (00:26:24)
15. TowerLLM Challenges (00:29:20)
16. Future of Translation Technology (00:30:19)
17. Areas of Application for LLMs (00:32:09)
18. Understanding xTOWER (00:34:49)
19. AI Pipelines (00:37:23)
20. Language Coverage (00:38:59)
21. Hiring Environment (00:40:43)
22. Acceleration of LLMs and AI Progress (00:41:51)
237 episodes
Tous les épisodes
×1 #237 DeepL Voice Launch and Interpreter Job Surprise 30:20
1 #236 The Year in Review and 2025 Predictions! 44:45
1 #235 The Interpreting SAFE AI Task Force with Katharine Allen and Dr. Bill Rivers 47:03
1 #234 RWS and BLEND Appoint New CEOs, Microsoft and Amazon Speech AI 28:14
1 #233 Charles Campbell on Juntos and the Future of Localization in Latin America 32:56
1 #232 Reddit Rises on AI Translation, TransPerfect Buys Studio 27:00
1 #231 AI Impact on Media Localization with ZOO Digital CEO Stuart Green 58:35
1 #230 Christian Elongue on Advancing the Language Industry in Africa 41:54
1 #229 Leading Localization from Asia with EC Innovations’ Sijie Wei 38:16
1 #228 OpenAI’s Advanced Voice Mode and Interpreting, AI Dubs for Instagram 29:26
1 #227 Leading a Boutique LSP with GORR Co-founders Gregor and Olivera Rosulnik 50:32
1 #226 MTPE Use, Pricing, AI Adoption, and Other Findings from the 2024 ALC Survey 29:36
1 #225 Smartcat's Series C, YouTube Dubs, Viva Translate Shuts Down 23:15
1 #224 On AI for Deep Content Understanding with Zetta Cloud Chief Strategist George Bara 43:18
1 #223 Evaluating Translations Objectively with ISO 5060 Lead Christopher Kurz 46:00
Bienvenue sur Lecteur FM!
Lecteur FM recherche sur Internet des podcasts de haute qualité que vous pourrez apprécier dès maintenant. C'est la meilleure application de podcast et fonctionne sur Android, iPhone et le Web. Inscrivez-vous pour synchroniser les abonnements sur tous les appareils.