June 25, 2024

C-3PO’s babies

Hi there, 

Generative AI is ushering in the next era of translation. 

Advanced language models can better grasp context and nuance than earlier approaches to machine translation, improving AI translation accuracy.

Big companies even think it's good enough to rely on now. 

For example, this OpenAI customer at a company with 50,000+ employees is “moving away from people-based translation to more LLM-based translation”:

The buyer also points to OpenAI’s growing share of the crowded translation market. 

Expect a shake out. 

A search on language translation and localization brings up 100+ VC-backed companies on the CB Insights platform, with the majority founded before 2020. 

Sam Altman has talked about “steamrolling” certain types of companies and this looks to be one area that GAI is fundamentally disrupting.

CB Insights data indicates some startups may be short on runway, with multiple players 2+ years out from their latest raise. For example:

• Smartling — $160M Series E in December 2021
• Lilt — $55M Series C in April 2022
• Lokalise — $50M Series B in December 2021 (with extension in 2022)

And some we observe are shrinking, per CB Insights headcount data. Below is the trend for Lokalise. 

Specialized startups vs. generic models

At the same time, specialized AI translation startups like DeepL and Unbabel are betting they can carve out a niche vs. foundation models from companies like OpenAI and Anthropic and competing big tech products. 

These specialized startups emphasize their focus on compliance, accuracy, purpose-built models, and workflow tools as differentiators.  

In May, Germany-based DeepL raised a $300M Series B round at a $2B valuation just over a year after hitting a $1B valuation. It reports over 100,000 business customers.

DeepL was using OpenAI for some capabilities while also developing its own LLM, per reports in 2023. Unbabel has also built dedicated translation models on top of Meta’s Llama.

This DeepL customer needed a “deeper integration” and higher-quality translation than offered by Google Translate and Microsoft Translator.

Looking ahead

One limitation of current AI translation is coverage of languages other than English, as this Llama customer highlights. 

Lower-resourced languages, or those with limited training data available, will continue to pose a challenge to developing comprehensive AI translation capabilities.

This might be one area where small language models may be quite useful as the CEO of Veon, Kaan Terzioglu, highlighted in the company’s March earnings call.

We dig into the potential for small language models in our recently published Enterprise AI Roadmap report.

I love you.

Anand

@asanwal 

P.S. We’re hiring for our research unit. Be at the front of what’s next in tech. Check out the roles here.

Get started with CB Insights

CB Insights' emerging technology insights platform provides all the

analysis and data from this newsletter. Our data is the easiest way to discover and respond to emerging tech. 

Was this email forwarded to you? Sign up here