Armenian Tech’s Year of Generative AI

If there is one term that could capture the essence of the technology landscape in 2023, it would undoubtedly be “generative AI”. Over the past year, the world has adapted to a new reality where widely-used technology products now offer capabilities that were once considered science fiction. Two technologies driving this new reality are large language models (LLMs) and text-to-image models. Through the input of “prompts” — text-based instructions given to AI — these systems can write entire book chapters and create striking artwork. 

These capabilities saw the tech world enter a new era in 2023. Much like how the internet and mobile technology marked the beginning of new technological eras, this year marked the dawn of AI. AI features and capabilities have become commonplace in most tech products. This new era is rapidly advancing as companies that develop AI models, which power these features, adopt a business model of providing access to the models for other developers, charging them for usage. This easy access to AI technology is driving innovation across various sectors, heralding a transformative phase in the evolution of technology. 

The start-up ecosystem in Armenia has followed this trend as well. In 2023, almost all major Armenian start-ups built or integrated AI technology into their applications. Some start-ups have even based their entire businesses on generative AI products, while others have enhanced their existing features by leveraging the new capabilities offered by generative AI. 

Turning Prompts Into Images

In the field of image generation, Picsart, Armenia’s first home-grown unicorn, has emerged as a leader through advancements in AI.  Initially established as a mobile photo editing software more than a decade ago, Picsart has consistently integrated AI into its features. With the incorporation of generative advancements, Picsart now markets itself as an AI-powered creative companion. The product features AI image and art generators, allowing users to further edit the generated outputs using Picsart’s built-in tools. Additionally, the company has introduced a separate app called SketchAI, which, instead of text prompts, allows users to do the digital equivalent of a finger painting, enabling users to quickly draw an image on their mobile screens and have the AI transform it into more sophisticated looking artwork.

To enhance its generative capabilities, Picsart has introduced additional AI features to complement and allow modifications of generated content. One example is the AI Replace feature, which replaces parts of photos and artwork through prompts, such as “make the jacket blue”. This feature was humorously marketed earlier this year by suggesting people write prompts to replace their ex-partners in old photos with red flags. Beyond fun use cases, this feature offers utility for businesses aiming to create AI-generated content. It enables iterative refinement of images, allowing companies to progressively tailor them to the desired final version for product and brand marketing. 

On the branding front, Picsart also offers a logo generator that allows companies to quickly generate logos for their businesses. Lastly, their generative AI powered features include text generation for social media copy, image captions, and branding materials. With all of these features, Picsart has become a platform that covers the entire content pipeline using generative AI. 

Picsart’s focus on generative AI extends to significant research and development (R&D) efforts, resulting in the creation of new products. The Picsart AI Research Lab (PAIR) has released work on training AI systems to generate video content, leading to the Picsart AI Video Generator tool. Similar to image generation, this AI tool takes a text prompt as input and generates a video. Furthermore, PAIR published a paper on refining image generation models at the prestigious CVPR conference earlier this year.  

Wirestock, another prominent player from Armenia in the field of image generation, specializes in streamlining the process of selling stock content on major platforms. Wirestock has recognized the growing popularity of AI-generated art and now allows creators to sell such content through its platform. Additionally, earlier this year, the company introduced “Creative Challenges” to incentivize creators with monetary rewards for the best AI-generated images in specific, high-demand categories, further integrating AI into the world of digital art and commerce.

AI Generated Audio Using User Voice

Beyond image and video generation, generative AI has also enabled audio generation. Text-to-audio tools have existed for many years now, but recently, products have emerged that allow audio generation using a specific person’s voice. Podcasts have been released online that appear to feature conversations with deceased people, such as Steve Jobs, based on scripts provided to the AI. One Armenian company working on this is Podcastle. Since its launch in 2020, the company has been developing AI-based podcasting tools. Earlier this year, they introduced AI Voice Cloning. The product requires users to record a number of sentences with their own voices, capturing various sounds needed to generate audio using their voices. Then, with a given script, Podcastle can generate audio using the user’s own voice, instead of the generic sounds typically associated with generated audio. This additional generative AI functionality offered by Podcastle expands on traditional methods.

For the Business World

While these are mainly consumer facing applications, businesses, particularly enterprises, have their own considerations when it comes to using generative AI in their workflows. Since 2015, Zero Systems has focused on the business sector by building AI systems that automate various business processes, including timekeeping and email management. This year, Zero Systems has focused on building infrastructure for generative AI workflows that cater to enterprise needs. Their product helps bridge the gap between available LLMs (Language Models), and the additional features required by enterprises, such as security layers and enhanced reliability of outputs. 

In the finance world, Armenian company Cognaize announced an $18 million Series A funding round to build LLMs for the finance sector. Cognaize’s models, trained on financial data, allow companies in the sector to use AI for processing large amounts of unstructured data so common in the finance world. 

Generating Training Data Using AI

Generative AI has also been widely used in the generation of synthetic data. AI systems are trained to perform specific tasks using large datasets. For instance, self-driving cars are trained using millions of road images that have been annotated to identify different elements such as stop signs, pedestrians, and vehicles. This allows the AI to learn about driving. Due to the costs and limitations associated with obtaining real-world data, companies that train AI models have turned to synthetic data as an alternative to bridge the gap. Two companies from Armenia, SuperAnnotate and Manot, are active players in this field. 

SuperAnnotate, a popular data annotation platform, has recently announced SuperGen. This new product enables users to generate image samples as training data through simple prompts. Similarly, Manot also generates synthetic images for training data, but it focuses on specialized scenarios known as edge cases. Edge cases refer to specific scenarios where AI systems struggle due to insufficient training data. For instance, if a face-detection AI lacks training images of people wearing glasses, it may inaccurately detect such individuals, leading to biased outcomes. Manot identifies these weaknesses and generates training data to address these gaps, ultimately enhancing the accuracy and performance of the AI model.

AI-Built Websites and Applications

An area of generative AI that may be less familiar to many is the creation of AI-generated websites and apps. Two notable players in this field are Armenian companies 10Web, which specializes in creating websites based on user prompts, and Softr, known for its ability to generate web applications. 

10Web begins its website generation process by gathering information about the specific business. This includes details about preferred layouts and design elements. Utilizing generative AI, the platform then crafts text and images that are optimized for search engine optimization (SEO) and customized to align with the business’s identity and needs. The initial output serves as a foundation, allowing users to further refine and adjust the generated content to better match their vision. The final phase involves adding  extra pages and content as needed to create a comprehensive and tailored website. In addition to their core website generation features, 10Web also offers an AI business name generator and an AI marketing strategy generator.

In contrast, Softr focuses on generating business applications. The Softr platform was launched in 2020 as a no-code application builder, allowing people with little to no coding experience to create complete apps. Now, with the integration of generative AI, the process has become even simpler. 

Their AI App Generators allow businesses to create web applications by providing prompts. This includes a range of applications such as client portals, membership platforms, and various internal tools, all customized to meet the unique needs of each business. Softr takes care of the application’s design, user role controls, and data integrations. 

The New Players

Observing the Armenian start-up ecosystem, it is evident that established start-ups have been the primary beneficiaries of generative AI technology.  These start-ups have successfully integrated these new capabilities into their products and launched complimentary offerings. One notable example is SaaS giant ServiceTitan.Through its Titan Intelligence platform, ServiceTitan provides generative AI features for creating invoices and review responses, catering to contractors managing their operations on the platform. Another key player is Voice AI leader Krisp. Krisp has expanded its popular noise-canceling application by incorporating features that generate meeting notes and transcriptions of calls.

We have also seen the emergence of several new start-ups that have been built around this technology from the beginning. One example is DoWork.ai, a platform that enables businesses to deploy AI agents for handling customer service inquiries. What sets this platform apart is its integration with a company’s specific data, allowing it to provide contextually relevant information tailored to each individual customer query. This capability addresses a key limitation found in more general AI solutions like ChatGPT, which, despite their versatility, often lack the specific understanding of a business’s context. Similarly, Anania, another recently launched Armenian start-up, focuses on syncing with internal data sources to conduct analytics and perform tasks that require direct access to a company’s proprietary data. 

Looking to 2024 and Beyond

The integration of generative AI features in the Armenian start-up ecosystem has primarily involved the use of existing AI models released by major tech companies like Google and OpenAI. Apart from a few exceptions, primarily the work done by Picsart through PAIR, we have yet to see the development of large generative models by Armenian companies. This is not uncommon, as the development of these models is both highly cost prohibitive and requires access to significant computational resources. It also requires highly skilled AI teams with experience building models of significant scale, a talent that is scarce even in major markets like the United States. 

To enhance Armenia’s competitiveness in this field, members of Armenia’s scientific and tech ecosystems in 2023 proposed the establishment of an AI supercomputing center. This center would house a cluster of GPUs, which are essential hardware for developing AI models. In November, Hakob Arshakyan, the former Minister of High Tech Industry and current Deputy Speaker of the National Assembly, announced that $8.5 million from the 2024 budget will be allocated to purchasing the supercomputer. 

Investing in the computing resources of the ecosystem is critical, especially considering the difficulty in obtaining specialized hardware for building larger AI models due to high demand. In 2023, major cloud GPU providers experienced a near-zero availability of A100 and H100 GPUs, the two most important models for training AI systems. Although specific details regarding the management and availability of the supercomputing center have not been disclosed, investing in the country’s computing resources is important to facilitate access for Armenian start-ups to the necessary computing resources for developing cutting-edge AI models. 

Access to a supercomputer will also ensure that more people in the ecosystem gain experience working with large computing resources, a highly sought-after skill in the world. This talent pool will make Armenia more globally competitive in the tech sector. One notable effort in developing Armenia’s talent pool is the machine learning research lab YerevaNN, which has already published works in this domain through successful collaborations with Meta’s AI lab FAIR. Their most notable work in generative AI is BARTSmiles, an LLM for the chemistry domain. YerevaNN also collaborated on two prominent AI research efforts by FAIR this year . The first research study demonstrated how mixed-modal AI models, which incorporate different types of data such as text and images, scale as more data and computing resources are used. The other effort was CM3Leon, a mixed-modal model capable of generating both text and images. These collaborations provide valuable knowledge transfer experiences for the Armenian ecosystem, ultimately increasing the country’s capacity in the long run with proper investments in computing infrastructure. 

Other countries have also adopted this strategy of securing computing resources for their national ecosystems. One notable example is the United Arab Emirates, which has acquired thousands of specialized chips in recent years. In 2023, the country’s Technology Innovation Institute (TII) launched their Falcon family of AI models, which are competitive with the world’s best AI models. The Financial Times reported that Saudi Arabia’s KAUST, a leading research center in the country, has also been rapidly acquiring AI hardware, spending $120 million in 2023 alone to increase their computing resources. 

Becoming a global player in the AI field will undoubtedly require increased computing resources and the development of engineers and scientists capable of building large AI systems within the country. These advancements will also prove valuable as AI becomes increasingly  important in areas such as security, healthcare, and the economy. Ensuring that both human and computing resources are available domestically to meet the country’s needs will be a priority for all nations. 

Regarding start-ups, it is expected that the trends observed in 2023 will become the norm in product development. Generative AI capabilities will continue to advance as more investment is made in developing large AI models. Consequently, tech products will increasingly integrate additional AI features. 

Also see

EVN Disrupt