Over the past few years, advances in training large language models (LLMs) have moved natural language processing (NLP) from a bleeding-edge technology that few companies could access, to a powerful component of many common applications. From chatbots to content moderation to categorization, a general rule for NLP is that the larger the model, the greater the accuracy it’s able to achieve in understanding and generating language.

But in the quest to create larger and more powerful language models, scale has become a major challenge. Once a model becomes too large to fit on a single device, it requires distributed training strategies, which in turn require extensive compute resources with vast memory capacity and fast interconnects. You also need specialized algorithms to optimize the hardware and time resources.

Cohere engineers are working on solutions to this scaling challenge that have already yielded results. Cohere provides developers a platform for working with powerful LLMs without the infrastructure or deep ML expertise that such projects typically require. In a new technical paper, Scalable Training of Language Models using JAX pjit and TPUv4, engineers at Cohere demonstrate how their new FAX framework deployed on Google Cloud’s recently announced Cloud TPU v4 Pods addresses the challenges of scaling LLMs to hundreds of billions of parameters. Specifically, the report reveals breakthroughs in training efficiency that Cohere was able to achieve through tensor and data parallelism.

This framework aims to accelerate the research, development, and production of large language models with two significant improvements: scalability and rapid prototyping. Cohere will be able to improve its models by training larger ones more quickly, delivering better models to its customers faster. The framework also supports rapid prototyping of models that address specific objectives — for example, creating a generative model that powers customer-service chatbot — by experimenting and testing new ideas. The ability to switch back and forth among model types and optimize for different objectives will ultimately allow Cohere to offer models optimized for particular use cases.

The FAX framework relies heavily on the partitioned just-in-time compilation (pjit) feature of JAX, which abstracts the relationship between device and workload. This allows Cohere engineers to optimize efficiency, and performance by aligning devices and processes in the ideal configuration for the task at hand. Pjit works by compiling an arbitrary function into a single program (an XLA computation), that runs on multiple devices — even those residing on different hosts.

Cohere’s new solution also takes advantage of Google Cloud’s new TPU v4 Pods to perform tensor parallelism. which is more efficient than the earlier pipeline parallelism implementation. As the name suggests, the pipeline parallel approach uses accelerators in a linear fashion to scale a workload, like a single long assembly line. Accelerators must process each micro-batch of data before passing it along to the next one, and then run the backward pass in reverse order.

Tensor parallelism eliminates the accelerator idle time of pipeline parallelism, also known as the pipeline bubble. Tensor parallelism involves partitioning large tensors (mathematical arrays that define the relationship among multiple objects such as the words in a paragraph) across accelerators to perform computations at the same time on multiple devices. If pipeline parallelism is an ever-lengthening assembly line, tensor parallelism is a series of parallel assembly lines — one making the engine, the other the body, etc. — that simultaneously come together to form a complete car in a fraction of the time.

These computations are then collated, a process made practical thanks to Google Cloud TPU v4 VMs, which more than double the computational power. The superior performance of v4 chips has enabled Cohere to iterate on ideas and validate them 1.7X faster in computation than before.

At Cohere, we build cutting-edge natural language processing (NLP) services, including APIs for language generation, classification, and search. These tools are built on top of a set of language models that Cohere trains from scratch on Cloud TPUs using JAX. We saw a 70% improvement in training time for our largest model when moving from Cloud TPU v3 Pods to Cloud TPU v4 Pods, allowing faster iterations for our researchers and higher quality results for our customers. The exceptionally low carbon footprint of Cloud TPU v4 Pods was another key factor for us.

Aidan Gomez
CEO and co-founder, Cohere

Why Google Cloud for LLM training?

As part of a multiyear technology partnership, Cohere leverages Google Cloud’s advanced AI and ML infrastructure to power its platform. Cohere develops and deploys its products on Cloud TPUs, Google Cloud’s custom-designed machine learning chips that are optimized for large-scale ML. Cohere’s recently announced their new model improvements and scalability by training an LLM using FAX on Google Cloud TPUs, and this model has demonstrated that transitioning from TPU v3 to TPU v4 has so far enabled them to achieve a total speedup of 1.7x. In addition to a significant performance boost, TPUs provide an excellent user experience with the new TPU VM architecture. Importantly, Google Cloud ensures that Cohere’s state-of-the-art ML training is achieved with the highest standards of sustainability, powered by 90% carbon-free energy in the world’s largest publicly available ML hub.

By adopting Cloud TPUs, Cohere is making LLM training faster, more economical, and more agile. This helps them provide larger and more accurate LLMs to developers, and put NLP technology in the hands of developers and businesses of all sizes.

To learn more about these LLM training advances, you can read the full paper, Scalable Training of Language Models using JAX pjit and TPUv4. To learn more about Cohere’s best practices and AI principles, you can check this article co-authored with Open AI and AI 21 Labs.

Business leaders see much in Artificial Intelligence, including new ways to save money, serve customers, and figure out what to build. What’s often tougher is deciding how to engage. That’s understandable, since getting into AI often raises questions about costs, data integrity, project length, and similar issues of planning and execution.

For CIOs concerned with the deployment and use of IT at their companies, these are critical issues. Here is one way to think about this complexity, by breaking it down into three different areas; Early Automation, Learning and working, and System Views.

First, a bit of good news and myth dispelling for readers who think AI is some far off promise, still in the labs and not safe for work. The reality is that AI is very much here, and almost everyone in your workplace is using it every day. People work with AI when they use Google Search, Photos, the autocomplete feature in Docs, live translation in Pixel phones, and many other areas. It is in the products of many other companies too.

“IDC reports that the market for AI software, hardware and services is expected to break the $500 billion mark in 2023.”

Additionally, it’s increasingly clear that AI can be adapted as a virtuous process, not just a one-off purchase (though there is much to recommend that.) During a recent Alphabet earnings call, CEO Sundar Pichai noted that “investments in AI will be key” to its near-term strategy, with new techniques that make it faster and easier to train and build AI for a number of uses. Additionally, the company is offering AI-driven “insights, new tools, and automation” to its advertising clients. The striking thing in this was the way that developing AI in one area could lead to growth in many others.

So, how does an IT leader foster a growth process like this for their stakeholders? By leading people through the well-established stages of Awareness, Learning, and Extension. Here’s what I mean.

Awareness: Early Automation

Consumer-facing AI is particularly strong in communications functions like voice recognition, translation, and writing tips. It’s similar in business uses: One of the most effective early instances of AI in the workplace has been Contact Center AI (CCAI), which manages basic customer communications, automatically answering common questions and prioritizing calls that require human assistance. It is doing what automation has always done best, automating the rote stuff and leaving the higher-value imaginative activity to people. It has been used by governments, retailers, telecommunications companies, and others, in a wide variety of use cases.

These and similar language-centric products, like DocAI for extracting information from things like invoices, receipts, or AI that extracts information from business contracts, have a number of benefits. For one, the investment is relatively easy to control, unlike with a research project leading to a formal launch. The payoff is also clearer. In the case of contact centers, in particular, the automation relieves stress, wins loyalty and slows disaffection in a high-turnover area. In both cases, successful results build allies in the business, who can testify to the earlier benefits when it’s time to take on something more complex.

Perhaps best of all, creating interest in basic AI services for business, right now, means people become engaged in learning more, since they see the early benefits and wonder what else might be done.

Learning: The Human Factor

The Natural Language Processing (NLP) that goes into these ready made, “out of the box” AI products can, not surprisingly, be used on much more sophisticated levels. Twitter, for example, processes 400 billion different events in real time, and its staff queries this trove using advanced NLP, answering questions and improving customer experiences.

There is clearly an enormous gap between Call Center AI and processing Twitter’s 400 billion events per day, but it’s not noticed enough how quickly that gap is closing. Look at how many products, partners, and training resources have emerged in the past few years. The gap makes sense, insofar as both the means of AI (like large data sets, good algorithms, and sufficient computing) and the value of AI, are new.

Increasingly, as AI is incorporated into standard enterprise tools like spreadsheets and analytic tools, easier to use AI becomes a skill within reach for many (even as the advanced end becomes more complex, meaning this ease of use process will continue for some time.)

It’s so new that AI skills demand isn’t met by conventional education means, creating lots of good opportunities for both nonstandard skills training, and in-house learning in the workforce. Companies offering AI skills training could well gain a competitive advantage and retain staff better.

Extension: Building System Views

When a new technology lands and gains in popularity, people seek to find new uses for it, or build connections among its different uses. Networked computing is one example, but think also of the way cars were soon followed by trucks and fire engines, or the way the data services on wireless phones soon morphed into the App Economy. If something is useful, people look for ways to grow it.

How will AI grow? My colleague Dominik Wee recently wrote about ways that AI will soon change supply chains, change product design, and improve sustainability. Most interestingly, he talked about how customers in manufacturing were realizing savings and gaining insights when once separate quality control data was combined with system wide views of the quality process.

There are several reasons to think AI will promote many such system views. For one thing, successful AI promotes the collection of data from more places, at greater frequency, since that leads to insight (and the cost of data collection is dropping.) Additionally, AI is good at spotting patterns and interactions that are not currently known. As well, AI is used in prediction and scenario planning, which leads to better understanding of how large-scale systems interact.

This comes at a time when we have more ways of seeing the world, from satellites, sensors, social media, and much more. We have more awareness of interactions, and a demand to understand them, in everything from the supply chain crisis, to human rights and sourcing regulations, or in the business realities of partnering, and serving customers in all sorts of ways, online and in the physical world.

Whether by coincidence or design, the Age of AI is also an age when organizations see themselves more accurately with a rich web of connections, with their choices and actions having more resonance than ever. That awareness is both a competitive tool, and a call to greater responsibility, potentially affording more customer loyalty and a more satisfied workforce for those who get it right.

That transformation won’t happen everywhere overnight, but it seems to be happening at all sorts of companies. And the trend for AI to assist, to be studied and grown, and to provide a richer understanding of the world, is happening every time someone touches this technology, at whatever level they need.

Conversational AI is opening up a new world of possibilities in areas like customer experience, user engagement, and access to content. 

In Cloud AI, we’ve taken Google’s groundbreaking machine learning models in speech and natural language processing and applied them to the contact center space, radically improving the customer experience while also driving down operational costs. It’s shifting the focus of contact centers from the backroom, out of sight (and mind) to the boardroom and the strategic heart of the business.

Our solution, called Contact Center AI (CCAI), is an accelerator of digital transformation as organizations all over the world figure out how to support their customers during these challenging times. One customer that’s redefined the possibilities of AI-powered conversation using CCAI, is Verizon.

Whether through voice calls or chat, Verizon customers will no longer need to go through menu prompts or option trees; they simply say or type their question, and CCAI’s natural-language recognition feature finds the best way to assist them. No stilted speech or robot-like commands.

“These customer service enhancements, powered by the Verizon collaboration with Google Cloud, offer a faster and more personalized digital experience for our customers while empowering our customer support agents to provide a higher level of service,” said Shankar Arumugavelu, SVP & CIO at Verizon. 

CCAI is also driving cost savings without cutting corners on customer service. In the past, to improve customer satisfaction (CSAT), you had to hire more agents, increasing operating costs. Conversely, if you reduced the number of agents, your CSAT scores went down. Contact Center AI (CCAI) lets you have both.  

CCAI does this by expediting customer requests using virtual agents, assisting live agents, and providing insights on customer interactions to drive improvements back into the system. The result is a better experience for your customers—they get answers faster—and you get lower operating costs. It allows human agents to focus on higher value tasks in their customer interactions.

New products and features add richer customer experience 

We are continuing to invest in CCAI and today we are excited to announce the following new capabilities:

Dialogflow CX is the latest version of Dialogflow, a development suite used by over a million developers for building natural, rich conversational experiences into mobile and web applications, smart devices, bots, interactive voice response systems, popular messaging platforms and more. This new version of Dialogflow is optimized for large contact centers that deal with complex (multi-turn) conversations and it is truly omnichannel – you build it once and deploy it everywhere – in your contact centers and digital channels. Dialogflow CX features a new visual builder to create, build and manage virtual agents. It’s available now, in beta.

Dialogflow CX brings conversation state management to a whole new level

Lukasz Rewerenda
Principal Solutions Architect, Randstad (Netherlands)

Agent Assist for Chat is a new module for Agent Assist that provides agents with continuous support over “chat” in addition to voice calls, by identifying intent and providing real-time, step-by-step assistance. Agent Assist enables agents to be more agile and efficient and spend more time on difficult conversations, giving both the customer and the agent a better experience. It transcribes calls in real time, identifies customer intent, provides real-time, step by step assistance (recommended articles, workflows, etc.), and automates call dispositions.

For customers in regulated industries, Agent Assist can remove the risk of agents providing inaccurate information (which can happen due to high agent turnover and limited training). Agent Assist can also surface the latest discount information, deals and special offers, which can be hard for agents to keep track of as this information changes frequently. 

Custom Voice, available in beta, is a new capability for CCAI and our Text-to-Speech API that lets you create a unique voice to represent your brand across all your customer touchpoints, instead of using a common voice shared with other organizations. By taking advantage of the custom Text-to-Speech model created with Custom Voice, you can define and choose the voice profile that suits your business and adjust to changes without scheduling studio time with voice actors to record new phrases.

At Google Cloud, we are committed to ensuring that our products and features are in alignment with our AI Principles. If you are interested in Custom Voice, there is a review process to ensure that your use case is aligned with our AI principles. Learn more about Custom Voice and how to sign up here. 

Bring Google’s groundbreaking speech models on-prem

For many customers, an all-in approach to public cloud is not an option, which is why we’re extending our AI capabilities to run on-prem. Last week we announced Speech-to-Text On-Prem, the first of our hybrid AI offerings, now generally available. Speech-to-Text On-Prem gives you full control over speech data and since it runs in your own data center, it’s easy to comply with your data residency and compliance requirements. At the same time, Speech-to-Text On-Prem uses state-of-the-art speech models from Google researchers that are more accurate, smaller, and require less computing resources to run than existing solutions.

Take a look at how businesses saved costs and improved customer experience with CCAI in the new commissioned study conducted by Forrester Consulting: “New Technology: The Projected Total Economic Impact™ Of Google Cloud Contact Center AI.” To learn more about CCAI and conversational AI, see how our customers are using the solution and check out these sessions at Next OnAir:

• CCAI State of the Union (keynote)
• Measuring and Improving Speech-to-Text Accuracy (breakout)
• Transform customer service with CCAI (demo)
• See the next generation of 3-1-1 in action (demo)

We meet daily with contact center owners and customer experience (CX) execs across all industries, geographies, and business sizes. Looking back at these conversations, it’s crystal clear that 2022 was a high-stakes year for call centers, with three primary challenges trending across all customers and continuing in 2023:

• Many organizations feel pressure to rapidly scale up their call center operations in response to macroeconomic changes. Uncertain conditions are forcing Contact Centers to be ever more cost-effective, and to find ways to generate revenue for the business.
• End users are increasingly demanding and less forgiving when it comes to CX. Users have a choice, and they expect brands to meet them where they are with superior experiences. Connecting with customers where they engage is one of the key components of superior CX—customers should not have to go through elaborate processes or unhelpful phone trees to get help but should rather have service available quickly and easily in their preferred channels. Consumers demand more intimate ways of connecting with brands and Conversational AI can create that critical interaction medium. 
• Organizations understand that AI can help address these challenges. However, many business leaders remain unsure how to successfully make the journey. A growing number of offerings are on the market, but many don’t deliver on their promise, with long and expensive integration requirements and unpredictable and underwhelming outcomes.

Helping our customers successfully address these challenges and opportunities was one of our top priorities last year and will continue to be a significant focus in coming months. In this blog post, we’ll review our Contact Center AI (CCAI) news from last year, as a primer for 2023. 

Looking back: Why 2022 was a big year for Contact Center AI

In 2022, we increased our strategic investment in CCAI, including expanding it to include a comprehensive, end-to-end contact center solution suite that is user-first, AI-first, and cloud-first. We launched Contact Center AI Platform, our Contact Center as a Service (CCaaS) offering, as part of the CCAI product suite that offers a modern, turnkey solution, designed with user-first, AI-first, and cloud-first design. During Google Cloud Next ‘22, we shared lots of great content on how organizations can use CCAI to improve customer experiences, including these breakout sessions:

• Delight customers in every interaction with Contact Center AI
• Power new voice enabled interfaces with applications with Google Cloud’s speech solutions

We also got a chance to hear how customers are using CCAI to better reach their own customers, including Wells Fargo and TIAA. We partnered with CDW to discuss Providing Better Customer Experiences and with Quantiphi in a webinar called “Elevating the Banking Experience with CCAI Platform.” Just recently, our customer Segra shared their success story.

Through these customer interactions, three key priorities have surfaced as we look forward to 2023: Elevate the customer experience, bring new forms of AI to drive new automation and accelerate time to value.

Looking forward: Elevate CX, integrate new forms of AI, accelerate time to value 

1. User-first: Meet them where they are with elevated Customer Experience.

As we have learned, users expect that brands meet them where they are and on their own terms and expectations. To do that, brands must integrate with and adopt the latest user-centric technologies and product best practices from consumer mobile and web apps. Enterprise B2C can’t exist anymore in a parallel world of different and often inferior user experience. Google has over 20 years of experience in building such consumer experiences, with multiple products successfully serving billions of users. Bringing these capabilities and experiences from our consumer products and research teams to our cloud offerings was a key component for our product offerings in 2022 and is a big part of our key investments in 2023. Moreover, a vast majority of CX user journeys start with a query on Google Search or YouTube. Connecting with the users at that point, even before they reach out directly to the contact center is a win-win, saving money for the brand and delivering immediate value to the user. By focusing on the user we created a superior  integrated omnichannel experience.

2. AI-first and cloud-first: Quality contact center growth depends on transforming to modern, Cloud, AI solutions.

For contact centers to evolve, they need to transform from cost centers to revenue generators. That requires modern Cloud and AI solutions. Conversational data spans across all parts of the contact center, opening new ways to generate value. Cloud capabilities of privacy, security and scale can enable personalized CX across channels, enabling key omnichannel experiences. From a study by McKinsey: “Cross-channel integration and migration issues continue to hamper progress. For example, 77 percent of survey respondents report that their organizations have built digital platforms, but only 10 percent report that those platforms are fully scaled and adopted by customers. Only 12 percent of digital platforms are highly integrated, and, for most organizations, only 20 percent of digital contacts are unassisted.” Traditional telephony technologies are becoming commoditized and struggle to keep up with ever more complex rule based systems. Leaders in applicative AI and Cloud technology are stepping up as the new partners for brands who understand they need to take the leap to the next generation CX solutions. .  

3. Accelerating time to value while future proofing investments with predictable and measurable value

Reducing upfront implementation investment and accelerating time to value can be a challenge for contact center solutions. Scaling Cloud and AI can provide a faster path advanced conversational AI, can help address these challenges. Let’s look at three examples:

• Out of the Box(OOTB) integrated transcription, chat and voice summarization, and topic modeling — This saves customers money by reducing agent handling time for every chat and call, as well as providing valuable insights that can be used for quality management, contact center optimization and automation, agent and user churn prediction, business insights, and revenue opportunities.
• AI based chat and voice calls steering  paired with info-seeking virtual agents — Together these deliver higher Customer Satisfaction at scale while reducing cost –  by significantly reducing waiting queues and being routed to the wrong agent, as well as automating away total handling time.
• Reduced time to full automation — Reduce the complexity of conversation modeling, prebuilt components and APIs for shorter time to value and more predictable outcomes, and metrics driven ML-Dev & QA tools and playbooks.

With these new capabilities, our customers can now see results as soon as they implement CCAI. We’re excited to get our customers to where they want to be faster!

And there you have it: a quick overview of CCAI and its progress in 2022 and what’s coming in 2023. For more details, check out the documentation or our CCAI solutions page.

As AI-powered tools become increasingly prevalent, prompt engineering is becoming a skill that developers need to master. Large language models (LLMs) and other generative foundation models require contextual, specific, and tailored natural language instructions to generate the desired output. This means that developers need to write prompts that are clear, concise, and informative. 

In this blog, we will explore six best practices that will make you a more efficient prompt engineer. By following our advice, you can begin creating more personalized, accurate, and contextually aware applications. So let’s get started!

Tip #1: Know the model’s strengths and weaknesses

As AI models evolve and become more complex, it is essential for developers to comprehend their capabilities and limitations. Understanding these strengths and weaknesses can help you, as a developer, avoid making mistakes and create safer, more reliable applications.

For example, an AI model that is trained to recognize images of blueberries may not be able to recognize images of strawberries. Why? Because the model was only trained on a dataset of blueberry images. If a developer uses this model to build an application that is supposed to recognize both blueberries and strawberries, the application would likely make mistakes, leading to an ineffective outcome, and poor user experience.

It’s important to note that AI models have the ability to be biased. This is due to AI models being trained on data that is collected from the real world, and so it can reflect the inequitable power dynamics inherently rooted in our social hierarchy. If the data that is used to train an AI model is biased, then the model will also be biased. This can lead to problems if the model is used to make decisions that affect people by reinforcing societal biases. Addressing these biases is important to ensure that data is fair, promoting equality, and ensuring the responsibility of AI technology. Prompt engineers should be aware of training limitations or biases so they can craft prompts more effectively and understand what kind of prompting is even possible for a given model.

Tip #2: Be as specific as possible

AI models have the ability to comprehend a variety of prompts. For instance Google’s PaLM 2 can understand natural language prompts, multilingual text, and even programming codes like Python and JavaScript. Although AI models can be very knowledgeable, they are still imperfect, and have the ability to misinterpret prompts that are not specific enough. In order for AI models to navigate ambiguity, it is important to tailor your prompts specifically to your desired outcome. 

Let’s say you would like your AI model to generate a recipe for 50 vegan blueberry muffins. If you prompt the model with “what is a recipe for blueberry muffins?”, the model does not know that you need to make 50 muffins. It is thus unlikely to list the larger volume of ingredients you’ll need or include tips to help you more efficiently bake such a large number of muffins. The model can only go off the context that is provided. A more effective prompt would be “I am hosting 50 guests. Generate a recipe for 50 blueberry muffins.” The model is more likely to generate a response that is relevant to your request and meets your specific requirements.

Tip #3: Utilize contextual prompts

Utilize contextual information in your prompts to help the model gain an in-depth understanding of your requests. Contextual prompts can include the specific task you want the model to perform, a replica of the output you’re looking for, or a persona to emulate, from a marketer or engineer to a high school teacher. Defining a tone and perspective for an AI model gives it a blueprint of the tone, style, and focused expertise you’re looking for to improve the quality, relevance, and effectiveness of your output. 

In the case of the blueberry muffins, it is important to prompt the model using the context of the situation. The model might need more context than generating a recipe for 50 people. If it needs to be aware that the recipe must be vegan friendly, you might prompt the model by asking it to answer by emulating a skilled vegan chef. 

By providing contextual prompts, you can help ensure that your AI interactions are as seamless and efficient as possible. The model will be able to more quickly understand your request and it will be able to generate more accurate and relevant responses.

Tip #4: Provide AI models with examples 

When creating prompts for AI models, it is helpful to provide examples. This is because prompts act as instructions for the model, and examples can help the model to understand what you are asking for. Providing a prompt with an example looks something like this: “here are several recipes I like – create a new recipe based on the ones I provided.” The model can now understand the your ability and needs in order to make this pastry,

Tip #5: Experiment with prompts and personas

The way you construct your prompt impacts the model’s output. By creatively exploring different requests, you will soon have an understanding of how the model weighs its answers, and what happens when you interfuse your domain knowledge, expertise, and lived experience with the power of a multi-billion parameter large language model. 

Try experimenting with different keywords, sentence structures, and prompt lengths to discover the perfect formula. Allow yourself to step into the shoes of various personas, from work personas such as “product engineer” or “customer service representatives,” to parental figures or celebrities such as your grandmother, a celebrity chef, and explore everything from cooking to coding!

By crafting unique, and innovative, requests replete with your expertise and experience, you can learn which prompts provide you with your ideal output. Further refining your prompts, known as ‘tuning,’ allows the model to have a greater understanding and framework for your next output.

Tip #6: Try chain-of-thought prompting

Chain of thought prompting is a technique for improving the reasoning capabilities of large language models (LLMs). It works by breaking down a complex problem into smaller steps, and then prompting the LLM to provide intermediate reasoning for each step. This helps the LLM to understand the problem more deeply, and to generate more accurate and informative answers. This will help you to understand the answer better and to make sure that the LLM is actually understanding the problem. 

Conclusion

Prompt engineering is a skill that all workers, across industries and organizations, will need as AI-powered tools are becoming more prevalent. Remember to incorporate these five essential tips the next time you communicate with an AI model, so you can generate the accurate outputs that you desire. AI will forever continue to develop, constantly refining itself as we use it, so I encourage you to remember that learning, for mind and machine, is a never ending journey. Happy Prompting!