# 213 Phrase CEO Georg Ell on the Arms Race in Language Technology

New Section

The conversation begins with the importance of multilingual generation and the evolving landscape in the language technology industry.

Importance of Multilingual Generation

• G, CEO of localization platform Phrase, discusses the significance of being in an industry experiencing rapid change and innovation.

• Broad themes include the adoption of generative technologies leading to second-order changes in business models and commercialization of software.

• Business leaders are increasingly focusing on language technology, raising questions about cost, return on investment, and business value.

Exploring Industry Changes

The discussion delves into how businesses perceive multilingual content generation as a solved problem and the challenges posed by generative and AI technologies.

Perception of Multilingual Content Generation

• Businesses are chasing increasing expectations as technology advancements make generative models appear magical.

• Innovation around generative and AI technologies is nonlinear and non-deterministic, posing challenges in scaling solutions for enterprises.

Market Framework Analysis

A market framework analysis is presented, outlining different layers from compute to service providers in the language technology ecosystem.

Market Framework Layers

• The "stack idea" framework includes layers such as compute (Nvidia, Google Cloud), foundation models (OpenAI), platforms (Phrase), use case technologies, and service providers.

Large Language Models and Application Layers

In this section, the speaker discusses the advancements in large language models and their integration into application layers for hyper-automation.

Advancements in Custom Models and Performance Scoring

• The company has released the ability to create custom models and quality performance scoring based on large language models.

• Various integrations have been implemented to build an Enterprise application layer for hyper-automation.

Next Generation Machine Translation (Mt) Development

• The speaker mentions the development of "Next Gen Mt" by combining GPT 40 and Next Mt into a powerful generative and machine translation fusion.

• This new approach enhances tag handling, additional analysis capabilities, and overall quality of translation solutions.

Concerns about Building In-House Models

• Expresses concerns about companies engaging in an arms race with well-funded entities when developing in-house models.

• Emphasizes the importance of focusing on the application layer rather than competing at the foundational model level due to potential industry shifts towards commoditization.

Cost Considerations and Model Integration

This part delves into cost dynamics related to large language models, emphasizing the value of building upon existing models rather than creating proprietary ones.

Value Proposition of Application Layer vs. Foundation Models

• Discusses how cost curves and capability progression favor building on top of existing models rather than investing heavily in foundational model development.

Cost Efficiency in Model Integration

• Challenges the notion that significant spending is required for Large Language Model (LLM)-based solutions, suggesting that innovative approaches can reduce costs compared to traditional neural machine translation solutions.

Market Demand for Multilingual Text Generation

• Highlights market demand as 40% of localization buyers prioritize multilingual text generation within language AI solutions.

Integration Strategies with GPT 40

This segment explores strategies involving GPT 40 integration within existing frameworks to enhance solution quality while managing costs effectively.

Strategic Integration Approach with GPT 40

• Describes a strategic model where GPT 40 serves as a core component within a flexible framework allowing for future integration with other large language models.

Quality Enhancement through Combined Assets

Discussion on Multilingual Generation at Source

In this segment, the speaker discusses the impact of multilingual generation at the source on solutions like theirs, emphasizing a shift towards hyper automation and hyperscale in content creation.

Implications of Multilingual Generation

• The traditional process involved human creation of source text followed by translation, but now machines generate content in 50 languages instantly, eliminating early quality scoring opportunities.

• Hyper automation and hyperscale are becoming crucial as generative AI and multilingual content at the source are increasingly used for real-time content generation by enterprises.

• Hyper automation leads to hyper personalization and hyperscale, creating a logical progression towards achieving value through automated processes.

Quality Scoring Challenges

• Generating content in multiple languages requires rapid quality scoring by machines to prevent errors from propagating quickly, altering the role of human reviewers in the process.

• The need for effective quality scoring tools like their selection engine becomes essential for ensuring high-quality multilingual content publication without extensive human intervention.

Evolution of Business Models with AI Acceleration

This part delves into how AI acceleration challenges traditional business models in service provision, emphasizing a shift towards paying for speed and quality rather than per word or minute.

Transformation in Business Models

• The industry is transitioning from per-word or minute payment models to valuing speed and quality due to advancements in technologies like theirs that enable efficient content generation and review processes.

• Accelerated AI technologies are reshaping business models within the industry, leading to a search for new revenue streams and changed business practices among service providers.

Enhancing Tag Handling with GPT-4

The discussion focuses on improving tag handling using GPT-4 technology, highlighting advancements made by their research team in embedding regex into prompts for better tag management.

Advancements in Tag Handling

• Embedding regex into prompts along with additional techniques enhances GPT-4's ability to respect inline tags accurately during content generation, offering improved automation capabilities.

Discussion on AI Integration in Software Market

The discussion revolves around the integration of AI in the software market and its implications for Language Service Providers (LSPs) and tech platforms.

AI Impact on Software Market

• Large language models enable Tech Forward Thinking LSPs and tech platforms to offer comprehensive solutions by infusing AI throughout.

• Phrase focuses on strategic neutrality, collaborating with LSPs to create unique solutions combining services and software.

• Collaboration allows for tailored commercial or technological propositions, enhancing offerings to end customers.

Challenges and Opportunities with Large Language Models

Explores the challenges and opportunities presented by large language models in innovation journeys.

Large Language Models Insights

• Large language models are valuable but may not solve all problems as expected, leading to varied innovation outcomes.

• Innovation journeys are non-linear, emphasizing the importance of flexibility and adaptation in leveraging large language models effectively.

Evolution of RFP Trends in Language AI

Examines changes in Requests for Proposals (RFPs) related to language AI translation platforms over the past year.

RFP Evolution

• RFP trends show a shift towards specific requirements focusing on time, cost, quality improvements through large language models and AI.

• Emphasis on trust in technology output, data quality assurance, leveraging AI for quality assessment reflects evolving RFP criteria.

Customer Adoption Dynamics and Innovation

Discusses customer adoption dynamics regarding advanced technology solutions like large language models.

Customer Adoption Trends

• Enterprises exhibit varying levels of adoption maturity, with some pushing boundaries to drive innovation ahead of the curve.

Detailed Overview of Language Asset Management Solutions

In this section, the speaker discusses the development of a language asset management solution using AI technology to enhance efficiency and reduce costs in asset curation.

Leveraging Custom AI Training Model

• Automated asset curation offers up to 85% cost savings and 96% time savings.

• The solution focuses on cleansing, cleaning, and curating language assets efficiently.

• Clean language assets are essential for productivity and can be leveraged across departments.

Enhanced Portal Functionality with Single Sign-On

This part highlights the upgraded features of the portal, emphasizing single sign-on capability and customization based on cleaned language assets.

Customized Departmental Portals

• The portal now supports single sign-on for enterprise-wide access.

• It enables the creation of custom portals for different departments based on clean language assets.

• Custom models can be tailored for various functions like legal and marketing to streamline localization efforts.

Next Generation Machine Translation Framework

The discussion centers around next-gen machine translation (MT) by integrating advanced technologies like Phasex MT and OpenAI's GPT-4 to create a comprehensive framework.

Framework Components

• Next-gen MT fuses Phasex MT with GPT-4 to form a robust framework.

• Components include L&M product, foundational models like GPD 4, terminology handling, and tag management.

• Future iterations may incorporate additional foundational models and prompt building blocks for diverse language operations.

Automated Language Quality Assessment Solution

This segment introduces an automated Language Quality Assessment (LQA) tool designed to optimize quality assessment processes while reducing costs significantly.

Auto LQA Features

• Auto LQA aims to automate quality assessment processes traditionally done manually.

• It operates as a co-pilot alongside linguists rather than replacing them entirely.

• Initial results show substantial cost reductions up to 80%, with potential time savings ranging from 30% to 90%.

Balancing Human-Like Quality Assessment with Automation

Exploring the delicate balance between human-driven quality assessment and machine-driven automation in linguistic tasks such as translation evaluation.

Human vs. Machine Assessment

• Human quality assessment often lacks consensus (correlation factor: .48), leading to varied interpretations in translations.

• Machine-driven Auto LQA achieves a correlation factor of .46 compared to human assessments, indicating close alignment but facing trust challenges due to perceived low correlation.

Early Access Program and Platform Evolution

In this section, the speaker discusses the Early Access program's value and the company's transition towards a platform approach.

Early Access Program

• The Early Access program has been beneficial for companies, providing value to both Enterprises and LSPs.

• Benefits include cost reduction per word, faster processes, increased surface area for running LQA on content, and new features like analytics and automating updates.

Platform Evolution

• The company is moving towards being a platform company rather than selling individual products.

• Embracing hyper-automation requires composability, allowing customers to leverage various capabilities seamlessly across the platform.

Flexible Pricing Model and Customer Flexibility

This section delves into the company's shift to a flexible pricing model and enhanced customer flexibility within their platform approach.

Flexible Pricing Model

• Customers can no longer purchase standalone products but must buy the entire platform with fixed capabilities at a set volume. Additional capabilities can be purchased based on volume needs.

Customer Flexibility

• Enterprise customers have added flexibility in adjusting their base license up or down based on their specific requirements, offering varying levels of engagement with the platform throughout the year.

Multimodal Technologies Integration

The discussion focuses on exploring multimodal technologies integration within the company's offerings.

Multimodal Technologies

Digital Transformation Strategies

In this section, the speaker discusses the evolution of digital web applications and the shift towards multimodal approaches in strategic thinking.

Digital Evolution and Multimodal Approaches

• The extension of digital web applications into more multimodal forms is a current strategic focus .

• Two years ago, extensive Venture Capital funding was required for text-to-speech or speech-to-text technologies, but now it's as simple as an API call to a hyperscaler .

• Emphasizes the importance of capturing prompts as canonical items and exploring whether storing and processing images, sounds, and videos is necessary or if calling an API suffices .

Machine Translation Cost Analysis

This part delves into the cost comparison between machine translation methods like NMT and LLM.

Cost Comparison of Machine Translation Methods

• Recent discussions suggest that NMT and LLM have reached parity in terms of cost, although some argue that certain methods remain expensive .

• Contrasting viewpoints on cost: Conant suggests parity while others claim significant cost differences; ongoing exploration in this area .

Leveraging Existing Models for Efficiency

The speaker emphasizes leveraging existing models rather than building new ones to capitalize on evolving cost curves.

Leveraging Existing Models

• Focus on leveraging existing models like OpenAI to avoid single vendor dependency and adapt quickly to emerging technologies .

• Importance of proving not just basic translation but integrating it effectively into workflows with quality scoring for comprehensive solutions .

Changing Role of Linguists in Localization Industry

Discusses the changing role of linguists due to increased machine-generated content.

Changing Role of Linguists

• Shift towards more machine quality assessment due to high volume generated by machines; necessitates improved accuracy surpassing human capabilities .

Understanding the Impact of Partnerships in Language Services Industry

In this section, the speaker discusses the significance of partnerships in the language services industry and how leveraging add-on features through APIs can enhance workflows.

Leveraging Add-On Features Through APIs

• Add-on features through APIs are crucial in streamlining processes without needing to develop technology in-house.

• Partnering with companies like CaptionHub showcases the integration of speech-to-text, text-to-speech, and translation capabilities into platforms.

• Emphasizes the importance of partnerships like Lionbridge and others to facilitate high-quality work at scale.

Evolving Role of Language Service Providers (LSPs)

This segment delves into the evolving role of LSPs and their collaboration with technology providers to meet changing industry demands.

Evolution of LSP Partnerships

• Collaboration with top Global LSPs and tech-forward partners like Argos Vistatech signifies a shift towards tech-enabled solutions.

• Transition from product-centric approach to positioning as a platform for building solutions reflects a strategic shift within the industry.

Embracing Ecosystem Approach for Solutions Development

The speaker highlights the importance of an ecosystem approach in developing comprehensive solutions within the language services sector.

Ecosystem-Centric Strategy

• Acknowledges being part of an ecosystem where multiple service providers contribute to holistic solutions.

• Envisions collaborative press releases involving customers, technology partners, service providers, and LSPs as a testament to effective ecosystem engagement.

Implications of Digital Transformation on Localization Industry

Discusses how digital transformation is reshaping localization practices and expanding opportunities beyond traditional boundaries.

Shifting Focus Towards Digital Transformation

• Emphasizes moving beyond traditional localization towards broader digital transformation initiatives.

Engines in AI Solutions

The speaker discusses the engines used in their phrase language AI solution, highlighting the flexibility to choose different engines based on performance for various content types and language pairs.

Engines Flexibility

• DPL is a popular engine selected based on providing optimal results. -

• Different engines may be chosen depending on what is best suited for specific content types and languages. -

Innovation Cadence and Early Access Programs

The conversation shifts towards discussing the regular innovation rhythm of every three months, upcoming announcements, and plans for Early Access programs to engage customers more actively.

Innovation Rhythm

• Regular innovation rhythm with updates every three months. -

• June marks the upcoming update, followed by September and December releases. -

Early Access Programs

• Emphasis on expanding Early Access programs for customer engagement beyond discovery to Hands-On phases. -