# 221 Localization Is a Top ROI Use Case for GenAI with Lilt CEO Spence Green

AI Models and Human Verification in Translation

Introduction to the Podcast

• The podcast features Spence Green, CEO of Lil, discussing their vision for an end-to-end platform that integrates AI models with human verification.

• Spence shares his recent experiences at an AI conference in Las Vegas, highlighting the growing relevance of translation and localization in AI discussions.

Developments Since Last Podcast

• Spence reflects on the past three years since their last podcast, noting significant changes in both personal circumstances and industry dynamics.

• He emphasizes that Lil's business model remains focused on combining AI technology with human oversight to ensure quality standards for enterprise clients.

Shifts in Industry Approach

• There has been a notable shift from service-driven approaches to software-driven solutions in localization, influenced by advancements in AI.

• Companies are now more inclined to adopt comprehensive software solutions rather than relying on multiple services and vendors.

Adapting to New Technologies

• Spence discusses how Lil is adapting its offerings to include large language models (LLMs), which have become prominent since the rise of AI technologies.

• The platform allows users to choose from various models, including those from major providers like Google and Microsoft, as well as open-source options.

Continuous Training and Customization

• Lil's platform supports continuous training of approximately 177,000 models tailored for different content types and languages.

AI in Localization: Current Trends and Future Predictions

Overview of Current Models

• The speaker discusses the current series of models, which are single-digit billion parameter models, with plans for a new phase expected early next year focused on accuracy and performance cost.

• Clarification on model types: traditional neural machine translation (NMT) models can be multi-billion parameter encoder-decoder models, while GPT-style models are decoder-only frameworks ranging from billion to trillion parameters.

• For enterprise localization use cases like translation, quality control, rewriting, and summarization, smaller models may provide better performance cost without needing trillions of parameters.

Transitioning to New Technologies

• The transition from NMT to large language models (LLMs) is driven by client requirements rather than a singular shift; the platform remains agnostic to various offerings.

• Continuous training capabilities have been developed over ten years, allowing for efficient localization production. This includes solutions for clients operating in air-gapped environments.

Market Changes Post ChatGPT

• A significant change in market perception has occurred post ChatGPT; CEOs and CTOs are now actively exploring AI use cases within their organizations.

• There has been an increase in requests for proposals (RFPs), indicating a shift towards software-driven approaches in enterprise localization that had not previously been prioritized.

Translation Quality Improvements

• While state-of-the-art translation models have improved over recent years, there hasn't been a dramatic step function increase akin to advancements seen in conversational AI.

• The focus on AI for localization stems from a narrative shift where businesses recognize the potential benefits of adopting these technologies.

Predictions for 2024 and Beyond

• Six predictions were made regarding the future landscape of localization:

• Increased MTP drives rates leading linguists to leave the profession.

• Linguist shortages particularly affecting higher resource languages like those spoken in Nordic countries.

• Margin pressures resulting in mergers among top language service providers (LSP).

• Localization becoming one of the top ROI use cases amid generative AI implementation.

• MQM emerging as the default quality framework.

• CHD-style models surpassing traditional encoder-decoder styles concerning quality.

AI Transformation in Enterprises

Investment Trends and ROI in AI

• The discussion highlights significant investments in AI, particularly by companies like Nvidia, indicating a bullish market sentiment.

• A key question raised is the difference between current AI cycles and previous ones regarding actual ROI versus mere perception.

• The potential for positive ROI has existed for years; however, the current narrative emphasizes the vast software market's value, estimated at around $400 billion.

Leveraging AI in Service Markets

• There is an opportunity to apply AI models within large service markets that possess real differential pricing power, leading to substantial business growth.

• Industries such as localization, legal services, accounting, and customer support are identified as areas where AI can significantly enhance efficiency by automating labor-intensive tasks.

Change Management in Enterprises

• Resistance to change within enterprises often stems from perceived risks; thus, incentives must be established to encourage shifts in spending habits.

• The spectrum of AI adoption ranges from minimal efforts (like MTPE with traditional providers) to comprehensive transformations of entire workflows.

Executive Sponsorship and Career Incentives

• Successful change initiatives are more likely when there is executive sponsorship at the top level of an organization.

• A notable shift has occurred where executives now actively inquire about their company's AI strategies, reflecting a growing recognition of its importance.

Perception vs. Reality of AI Capabilities

• There exists a perceptual challenge among some stakeholders who believe that certain problems have already been solved by existing technologies.

Understanding the Shift from Consumer to Enterprise Use Cases

Quality Verification in Different Contexts

• The transition from consumer use cases to enterprise use cases introduces various challenges, particularly around quality verification, which varies based on business context.

• For some enterprises, quality may relate to regulatory compliance, especially in sectors like law enforcement where text must meet evidentiary standards.

Challenges of Scale and Integration

• Enterprises face complex systems infrastructure that requires integration with existing business workflows; creating thousands of web pages at scale is a common task.

• Cost efficiency is crucial for enterprises as they navigate these standard problems; AI and automation can assist but do not provide complete solutions.

Impact of MTPE on Linguist Employment

Market Dynamics and Pricing Adjustments

• The prediction that increased adoption of machine translation post-editing (MTPE) could lead linguists to leave the profession raises questions about market adjustments.

• A highly distributed freelance workforce lacks organization, making it difficult for individual freelancers to negotiate better rates against large buyers like language service providers (LSPs).

Creative Responses from Linguists

• As finding high-quality linguists becomes more challenging, many are leveraging tools to enhance their efficiency, including using AI-generated outputs.

• This shift has necessitated improved quality control mechanisms within organizations due to the influx of AI-generated content being submitted by linguists.

The Economics of Translation Services

Supply and Demand Imbalance

• There exists a supply-demand imbalance in the translation market; while there is an expectation for lower costs akin to consumer services, this does not account for necessary human involvement in verification processes.

• Businesses must evaluate whether they can rely solely on automated translations or if they require human oversight for accuracy and context.

Evolving Skill Requirements

• Human translators are now expected to possess higher qualifications than before due to advancements in AI models that produce more accurate outputs requiring nuanced understanding.

Public Sector Language Sensitivity and AI Integration

Overview of Public Sector Challenges

• The discussion begins with a focus on the public sector's price sensitivity and its drivers, particularly in defense and intelligence contexts.

Civilian Use Case: National Weather Service

• The speaker highlights a collaboration with the National Weather Service to address language barriers in weather alerts, emphasizing the need for multilingual communication.

• A significant incident in 2021 where non-English speakers failed to receive hurricane alerts led Congress to mandate alerts in multiple languages.

• By mid-2022, the Weather Service began distributing hurricane alerts in Spanish, expanding to six languages nationwide for better accessibility.

• The initiative utilized AI translation models fine-tuned by meteorologists, allowing them to focus on their primary duties rather than translation tasks.

Defense and Intelligence Language Gaps

• There is a noted deficiency in language capabilities within government sectors concerning strategic regions like Russia, China, and Iran.

• This gap presents an opportunity for AI integration to enhance diplomatic and intelligence operations while addressing security challenges.

Recent Features: AI Analytics for Localization

• Introduction of new analytics features that provide insights into the effectiveness and ROI of translation services offered by AI technologies.

End-to-End Observability

• Customers are promised complete visibility over production processes, ensuring transparency regarding data handling and model usage.

Measuring AI Accuracy

• A key feature allows users to track baseline model quality versus customized model performance through metrics such as verified word fractions generated by AI versus human corrections.

AI Integration and Content Creation in Enterprises

Overview of AI Model Accuracy and Enterprise Content Mapping

• The discussion highlights a tool that provides insights into AI model accuracy across different languages, helping identify areas where models may underperform.

• A visual representation (map) shows content flow within the enterprise, detailing integrations and the volume of content processed through various connections.

• The speaker emphasizes the importance of having an end-to-end view of AI interactions within programs for effective management.

Multilingual Content Generation: Current Trends and Future Prospects

• The conversation shifts to multilingual content creation, questioning whether it can scale effectively beyond localization and translation efforts.

• After gathering bilingual or monolingual data from localization, businesses can generate new content tailored to specific needs, such as blog posts in regional dialects.

• Two primary use cases for this technology are identified: regional marketers needing quick content generation and HR teams creating internal documentation.

Verification Steps in Content Creation

• There is a built-in verification step in the platform to ensure generated content meets quality standards before publication.

• Future capabilities may allow users to differentiate between existing content for translation versus new content generation based on context-specific data.

Enhancements in Translation Workflow Tools

• The discussion addresses common complaints regarding user interfaces (UI), particularly around post-editing tools not being interactive enough for linguists.

• An interactive text completion interface is highlighted as a feature that enhances productivity by allowing real-time editing during translation tasks.

Transitioning to Document-Level Interfaces

• The original version of Lilt focused on segment-level translations but is evolving towards document-level interfaces to improve workflow efficiency.

Reconstructing Document Structures and Workflow Orchestration

Challenges in Document Reconstruction

• The process of reconstructing published documents involves understanding their underlying structure to recreate formats like PDF, HTML, or DOCX.

• Language-specific nuances present challenges, as different languages have varying sentence structures and punctuation rules.

Workflow Orchestration in Localization

• The speaker discusses the rise of workflow orchestration in the localization industry over the past 12 to 18 months.

• Introduction of "Connect Builder," a no-code workflow builder that integrates AI models with human feedback for efficient project management.

• This system allows for visualizing workflows and aims to replace outdated methods like spreadsheets and emails.

Future Directions for Workflow Builders

• Current limitations exist in handling complex file formats during pre and post-processing stages within workflow builders.

• The goal is to achieve a "zero touch" strategy where projects flow end-to-end with minimal human intervention, focusing on exception management.

Fundraising Environment and Market Trends

Current State of Fundraising

• Discussion on the challenging fundraising environment for SaaS companies over the past 18 months due to rising interest rates affecting capital costs.

• Notable mention of an NYSE TV interview indicating potential IPO opportunities for language industry companies amidst market fluctuations.

Shifts in Investment Dynamics

• A shift towards strategic investments from major tech firms (e.g., Google, Microsoft), often involving credits tied to services rather than straightforward cash funding.

• Early-stage funding remains focused on narrative-driven pitches rather than financial performance, complicating capital acquisition for scaling businesses.

Outlook on Market Normalization

• Anticipation that current market conditions will normalize after a period of volatility, returning closer to historical averages.

• Comparison between current AI developments and past internet revolutions suggests significant future applications are still emerging.

Future Initiatives and Model Developments

Upcoming Features and Models

AI Development and Open Source Models

Recent Developments in AI Models

• The team began releasing new AI models in July, with on-prem customers receiving them by the end of the month. The cloud version is now available.

• Historically, strong open-source models were scarce; previous machine translation models had to be built from scratch over the last decade.

Advancements in Open Source Large Language Models

• Newer open-source large language models are proving to be highly capable, allowing for post-training enhancements and effective distillation into smaller models.

• This accessibility enables non-big tech companies to develop competitive AI systems without needing extensive resources like data centers or significant capital.

Specialization vs. Generalization in AI Technologies

• Competing directly with major tech firms like Google or Microsoft on architecture search is challenging due to their vast resources; however, focusing on specialized AI applications can yield impactful results.

• Companies can thrive by tailoring technologies for specific business use cases such as localization and content creation rather than building broad horizontal technologies.

Upcoming Events