MAINTAINING DOMINANCE: GENERATIVE AI

The Journey of Humanity’s Progress Has Brought Us to the Edge of Reality. Today, humanity is developing It is an undeniable reality that AI is often subject to sales and marketing activities that do not focus on actual requirements but rather align with trends. In nearly all known industries, AI is developed and purchased sometimes without even asking the “why” and “how” questions, inevitably leading to significant amounts of hype and bubbles.

When investing in AI, expecting such investments to guarantee absolute returns would be a mistake. On the other hand, failing to analyze trends or not giving them the necessary attention just because they might be a bubble could mean losing the technology race right from the start.

From the 1990s to the 2010s, AI succeeded in providing valuable solutions in various application areas such as data mining, industrial robotics, logistics, business intelligence, banking software, medical diagnostics, recommendation systems, and search engines. However, the 1990s marked a period when AI did not hold a prestigious position in the business world. As a result, many AI researchers deliberately labeled their work with different names during that time. Terms like informatics, knowledge-based systems, cognitive systems, optimization algorithms, or computational intelligence were used instead.

The first decade of the 2000s was characterized by the widespread adoption of machine learning techniques, which can learn from data, make decisions, and make predictions. Machine learning was seen as a powerful field for analyzing data, finding patterns, developing insights, making predictions, and automating tasks at speeds and scales previously impossible.

On the other hand, machine learning was also used as a tool to disguise AI research, as AI was often perceived as too “extraordinary.”

As we moved into the 2010s, significant advancements were made in the field of machine learning, particularly in an area called deep learning. Thousands of AI startups in the global market became known for their work in deep learning. Deep learning gained prominence through its ability to classify and detect objects in search engines or generate natural responses in AI-powered speech assistants.

Machine learning may require human experts to establish a certain hierarchy to understand differences between data inputs and often relies on more structured datasets. Deep learning, however, eliminates some of the human intervention. While deep learning requires more data points to increase accuracy, machine learning relies on less data.

Machine learning algorithms are typically designed to make decisions based on specific rules or criteria, enabling interpretability. In contrast, the way deep learning algorithms operate can be a significant obstacle in scenarios where end-users or stakeholders need explanations for the decisions made by an algorithm. While deep learning can solve more complex problems, it may also require specialized hardware.

As technology advances daily, AI, despite achieving different usage areas and popular successes, continued to be perceived as a “strange” field by certain segments. Achievements such as DeepMind’s AlphaGo program defeating South Korean Go master Lee Sedol in 2016 managed to capture the attention of a broad audience. However, this process had an impact on only a few people who played with the AlphaGo program and a specific circle, without enabling the global society to establish direct contact with AI. Similarly, AI-powered applications used by different industries during that period managed to attract the interest of only those with specific expertise.

With the advent of the 2020s, the widespread adoption of generative AI, which can create new content such as videos, images, code, text, animations, 3D models, etc., has led to a shift in interest towards AI.

Founded in December 2015, OpenAI had already made significant contributions in the field of generative AI before ChatGPT. However, the process following the launch of ChatGPT in November 2022 established OpenAI and ChatGPT as the flagship of generative AI. The fact that ChatGPT, a generative AI-powered chatbot, reached 1 million users within five days in 2022 represents a critical turning point in the development of this process.

The primary reason behind the intense interest in ChatGPT is its ability to be used for various individual and professional purposes, such as revising existing articles in terms of grammar and tone, conducting quick research on a specific topic, summarizing reports, translating texts into different languages, writing code, and creating social media content plans.

What sets ChatGPT apart from its competitors is the breadth of its use cases and the suitability of its outputs, which surpass those of other examples. While ChatGPT may not always generate the most accurate and appropriate outputs, it often serves as a resource that guides users and facilitates their processes. The availability of ChatGPT’s web application for free or for a limited fee and its strong initial step in the industry are other reasons for the intense interest.

It is clear that ChatGPT has confronted the global society with the reality of AI, even those who have been reluctant to acknowledge AI’s existence and viewed it as the latest technology bubble.

Today, we use capabilities that once seemed “strange” and “fictional” to many in a completely ordinary way. Just 3-4 years ago, having AI write a text was presented with “scary” headlines in the media. Now, ChatGPT and its competitors are being used by nearly every segment of society, including small children who need help with their homework. Although OpenAI CEO Sam Altman emphasized a certain improvement, he described GPT-2, GPT-3, and GPT-4 as “bad,” stating that GPT-5 would not be too bad. Therefore, users will likely encounter a more advanced AI in the future.

While OpenAI’s success is quite impressive, I believe that focusing solely on idolizing the leader in any matter narrows the perspective. As with all other technologies, we have not yet taken the necessary steps to address the risks posed by technology in this early stage of generative AI. There is no general consensus on what exactly these risks entail. However, as always, regulations and other discussions lag behind technology. At this point, focusing on the risks of monopolization in the field of generative AI and alternative efforts will broaden our perspective on the subject.

Firstly, the success of ChatGPT is closely related to its advanced AI capabilities as well as its access to large datasets and computing power. According to an analysis published in February 2023, ChatGPT’s computing power (hardware) cost is approximately $694,444 per day. It requires around 3,617 NVIDIA HGX A100 servers (28,926 GPUs). Accordingly, the cost per search is 0.36 cents.

The investment by tech giants in companies operating in the generative AI field, amounting to billions of dollars, is not limited to the example of OpenAI.

Anthropic, an AI company established in 2021 by Dario Amodei, former Vice President of Research at OpenAI, and his sister Daniela Amodei, who works on AI ethics, distinguishes itself from its competitors with the resources it has secured and its unique positioning. The company claims to focus on building reliable, interpretable, and steerable AI systems and has aimed to differentiate itself from rivals based on these principles since its inception.

Anthropic can be considered one of the strongest players in the generative AI sector alongside OpenAI. Alphabet (Google) has committed $2 billion, and Amazon has agreed to invest $4 billion in Anthropic. Both Alphabet and Amazon also support Anthropic with chips and cloud infrastructure. Dario Amodei, co-founder and CEO of Anthropic, has presented their simultaneous partnerships with both tech giants as a sign of “independence.” According to Amodei, this “independence” and “choice” set the company apart from other agreements.

In a statement made by Amazon in March 2024 following the investment decision, it was noted that the collaboration between Amazon and Anthropic to bring the most advanced generative AI technologies to customers worldwide was “still in its early stages.” Dr. Swami Sivasubramanian, Vice President of Data and AI at Amazon Web Services (AWS), emphasized that the two companies were helping organizations of all sizes globally integrate advanced generative AI applications into their operations.

Although Amazon has stated that Anthropic’s “primary cloud provider” is AWS, Anthropic has emphasized that the company adopts a “multi-cloud” approach. Not being tied to a single cloud provider helps spread risks. While labeling this approach as “independence” is a successful marketing strategy, it does not fully reflect the reality.

While Anthropic falls short of the $13 billion sourced by OpenAI from Microsoft, it still has access to a very strong resource base. However, Amodei has noted that while training models currently costs around $100 million, this is expected to rise to $100 billion in the future.

Anthropic’s Claude AI offers features that can compete with ChatGPT. Similar to ChatGPT, Claude AI can be accessed via the web and mobile platforms or as an API. While there is no significant pricing difference for the web and mobile-based technologies aimed at direct consumers, Anthropic offers a notable price advantage in the APIs designed for developers.

Generative AI-powered chatbots have the potential to influence every component of global society, from individual processes to private and public institutions operating in various sectors.

Naturally, this power significantly impacts the investment decisions of actors who can invest at the billion-dollar level. As concerns and investigations regarding monopolies become more prominent, the presentation of investments may also change.

In March 2024, Microsoft invested $650 million in Inflection AI. Following this investment, the majority of the company’s 70 employees, including its co-founders, began working within Microsoft. It was also decided that the company’s AI model (API) would be accessible via Microsoft’s cloud platform. With this decision, Inflection AI shifted its focus from general consumers to corporate clients and developers.

There are assessments suggesting that this process, presented by Microsoft as a “recruitment” rather than an “acquisition,” may have been formed to avoid intensifying federal antitrust investigations.

This type of investment is likely to attract significant attention, regardless of the sector or location where it occurs. Furthermore, Inflection AI’s development process adds to the importance of this decision.

Founded in 2022 by Reid Hoffman, co-founder of LinkedIn, and Mustafa Suleyman, co-founder of DeepMind, Inflection AI’s generative AI-powered chatbot Pi (“personal intelligence”) distinguishes itself from competitors with its high level of emotional intelligence. While ChatGPT often emphasizes that it is an “AI model,” Pi offers responses that come closer to human dialogue with natural language expressions. According to Inflection AI, Pi is designed to be a kind and supportive “companion,” offering friendly advice and concise information in a natural and fluent style during written and verbal communication.

In June 2023, Inflection AI managed to secure a $1.3 billion investment from Microsoft, NVIDIA, Reid Hoffman, Bill Gates, and Eric Schmidt. With the $225 million first-round investment obtained in 2022, the company has thus raised a total of $1.525 billion.

During this period, Inflection AI was valued at $4 billion, and Suleyman announced that the company had built “the world’s largest supercomputer.” This process, advanced with the support of CoreWeave and NVIDIA, was stated to involve the use of 22,000 NVIDIA H100 Tensor Core GPUs.

Just a few days before Microsoft’s March 2024 investment decision was announced, Mustafa Suleyman commented that Pi’s new version, Pi 2.5, had achieved results using only 40% of the training computation power of GPT-4. Suleyman emphasized that Pi 2.5 was on par with GPT-4 in many parameters. However, according to official information shared with the public at that time, while Inflection AI had 6 million monthly and 1 million daily active users, ChatGPT’s weekly active user count was at the level of 100 million.

At this point, while Mustafa Suleyman has started working as CEO of Microsoft AI, Inflection AI, under its new CEO Sean White, will focus on integrating its AI model, known for its emotional intelligence, into various business bots.

This chaotic landscape has led the U.S. government to direct its pressure toward tech giants through various investigations and statements.

In a June 2023 publication, the U.S. Federal Trade Commission stated that generative AI relies on certain inputs (data, talent, computing power). In this context, it emphasized that if specific companies or a company controls one or more of these key inputs, they could use their control to weaken or distort competition in generative AI markets. According to the Commission, if generative AI becomes an increasingly critical tool, those who control these key inputs could have significant influence over a large part of economic activities.

At this point, the Commission stated that while having large amounts of data is not “illegal,” many lawsuits have been filed alleging that companies’ data collection, storage, or usage policies and practices are unreasonable, unfair, or deceptive. On the other hand, according to the Commission, even with responsible data collection practices, companies controlling data can also create entry or expansion barriers that prevent fair competition from fully developing.

Given the current technical, economic, and lobbying power of tech giants, it is relatively easy for them to move from one monopolistic area to another. Indeed, this situation carries the risk of making monopolization an endless cycle. This risk has led to deep concerns (sometimes excessive) about tech giants.

While the relationships between U.S.-based tech giants and startups are considered risky even by the U.S. government, the perceived risks of this process can escalate significantly for other actors. This becomes even more evident given the potential outcomes of generative AI.

Microsoft, which has strengthened its position in the generative AI field with OpenAI, made a €15 million investment in February 2024 in Mistral AI, a French competitor to OpenAI. This investment is a relatively small amount compared to Microsoft’s size and its past investments in OpenAI. It also represents a small portion of the $544 million that Mistral AI has raised from a total of 22 investors. Nevertheless, Microsoft’s investment in a European competitor to OpenAI, regardless of the amount, is highly noteworthy.

Catherine Morin-Desailly, a member of the French Senate, commented on the matter by stating that the French government “should not boast” and “should stop waving around the term ‘digital sovereignty.’” According to her, “The government is completely lacking in coherence.” Kim van Sparrentak, a member of the European Parliament, also mentioned that Microsoft and Mistral AI had been lobbying during the negotiations for the EU’s AI Act. She stressed that the agreement should be evaluated not only from a monopolistic perspective but also from an ethical standpoint.

Brad Smith, Microsoft’s Vice President, commented on the investment, saying, “We see this as the emergence of a new sector of the economy; we call this the AI economy. (It) will create entirely new businesses and new categories of jobs.” According to Smith, the broad debate that regulators will eventually focus on is whether companies without data centers and cloud infrastructure, like those of Microsoft, Google, and Amazon, will have widespread access to the necessary infrastructure to train and develop AI models. Smith stated, “If this were not accessible, it would be a situation that restricts the development of the market. The important thing is that we are committed to making this accessible.”

According to Microsoft, its collaboration with Mistral AI focuses on three key areas:

1. Supercomputer Infrastructure: Microsoft will support Mistral AI with its Azure AI supercomputer infrastructure, providing “best-in-class performance and scale” for AI training and inference workloads for Mistral AI’s leading models.
2. Market Access: Microsoft will offer Mistral AI’s best models as Models as a Service (MaaS) to its customers via Azure AI Studio and Azure Machine Learning model catalog. Microsoft believes that this process will provide positive outcomes for Mistral AI in terms of promotion, sales, and distribution.
3. AI R&D: Microsoft and Mistral AI will collaborate on training purpose-built models for specific customers, including the European public sector.

While focusing on monopolies and their power, it is also essential to understand why startups may be inclined to collaborate with giants. In an interview given shortly after announcing the partnership with Microsoft, Arthur Mensch, co-founder and CEO of Mistral AI, highlighted several key points:

• Mistral AI is competitive with OpenAI or Google. “Agility is a strength.”
• Microsoft’s investment is a small part of Mistral AI’s total funding.
• More than 75% of Mistral AI is owned by Europeans. However, due to structural issues, European growth funds cannot provide sufficient investment for Mistral AI to achieve its goals.
• Mistral AI also utilizes different cloud providers in addition to Microsoft for computing power.
• The strategic partnership with Microsoft is an important first step in expanding distribution. Having Mistral AI’s models on Azure is expected to increase accessibility and reliability.

When asked how to avoid the dominance of tech giants, Mensch stated that the proximity between cloud providers and actors offering AI models should be examined. He emphasized that users should not be forced to use specific AI models. The importance of this issue becomes even more apparent when considering allegations that users were forced to choose a particular internet browser.

Mistral AI’s business model includes both open-source models, which offer free access expected to drive distribution and demand, and commercial models accessible for a fee. The company’s open-source models are primarily aimed at developers, but Mistral AI also targets general users. Le Chat, introduced by the company in February 2024, is similar in nature to ChatGPT, though it has a relatively higher error rate.

Therefore, Mistral AI’s business model positions collaboration with Microsoft and other tech giants as advantageous.

For government institutions and certain companies operating in critical areas, the question of who to collaborate with and under what conditions can be as important as the technological capabilities acquired.

This focus is particularly evident in the approach of Aleph Alpha, a Germany-based company active in the generative AI field.

With the emphasis on “Sovereignty in the AI Era,” Aleph Alpha defines its mission as empowering businesses and governments with the most advanced generative AI technology to gain a decisive advantage in the emerging AI economy.

Operating in the generative AI field with a focus on general consumers and competing with global giants while also emphasizing the principle of “sovereignty” is extremely challenging. At this point, Aleph Alpha focuses on “critical and complex processes” rather than B2C (direct-to-consumer), targeting sectors such as public services, healthcare, finance, law, and security. Companies and government institutions make up Aleph Alpha’s primary customer base.

Developing a product for direct consumers in the generative AI field requires competing in highly intense conditions. In addition to the standard-setting leadership of ChatGPT, various chatbot providers aim to attract direct consumer interest. Aleph Alpha’s CEO, Jonas Andrulis, has highlighted this intense competition, stating that there is nothing for them to gain in the B2C area. Accordingly, Andrulis emphasized that their strengths lie in “sovereignty” and “explainability,” but direct consumers are not interested in “data sovereignty” or “transparency.”

According to Andrulis, focusing on government institutions and companies has led them to establish their own data centers and work with “different partners” from the beginning.

Founded in 2019, Aleph Alpha has managed to raise more than $530 million in total investment. Of this, $500 million was secured in a funding round in November 2023 led by Germany-based Innovation Park Artificial Intelligence (Ipai), Bosch Ventures, and Schwarz Group. This funding round also included significant companies such as Germany-based SAP and U.S.-based Hewlett Packard Enterprise (HPE).

It is expected that the $500 million will be spent on category-defining research on foundational models, advanced product capabilities in business-critical environments, and commercialization activities with selected key partners.

According to Andrulis, with this investment, Aleph Alpha will:

• Continue expanding its services by providing customers with independence and flexibility in infrastructure, cloud compatibility, on-premise support, and hybrid deployments.
• Expand its interfaces and customization options tailored to business-critical needs.
• Remain the best option in strategic environments where sovereignty is at risk and for customers with significant responsibilities.

The investment and prior collaboration between Aleph Alpha, primarily funded by European companies, and the U.S.-based corporate IT solutions provider Hewlett Packard Enterprise (HPE) is particularly noteworthy.

HPE has linked its investment in Aleph Alpha to specific qualities the company possesses. According to HPE, Aleph Alpha:

• Develops enterprise-grade capabilities to ensure Data Sovereignty, encompassing aspects such as data security, regulatory compliance, and the reliability of model outcomes.
• Holds a significant position in offering on-premise solutions to enterprise customers at a time when most competitors are focused on public cloud or solving consumer problems.
• The participation of major corporate companies like SAP and Schwarz Group, which are Aleph Alpha’s customers, in the funding round has provided a potential ecosystem advantage around Aleph Alpha.
• Plays a central role in EU AI regulations and participates in these regulations.

Aleph Alpha’s CEO, Jonas Andrulis, has stated that they do not want to build their technology entirely on a cloud provider or specialized accelerator chips. The LLM model family “Luminous” was created in Aleph Alpha’s own data center, purpose-built for this task. In this regard, Andrulis emphasized that the last thing they would want is to force those who do not want to be in the cloud into the cloud. This special data center, which includes various hardware and software from HPE and is owned by Aleph Alpha, is considered a crucial factor that differentiates the company from its competitors. Andrulis also underscored that they can perform their customers’ use cases in Aleph Alpha’s data centers without forcing them to send their data to the cloud or any third party.

Another outcome of the ongoing collaboration between HPE and Aleph Alpha has manifested in the context of cloud services. It is now possible to access Aleph Alpha’s models through HPE’s cloud service, GreenLake for Large Language Models, which enables corporate users to train, fine-tune, and deploy large-scale AI models.

According to HPE, this collaboration, which focuses on mutual benefits for both companies, goes beyond technical capabilities and commercial interests. The two companies also share a strong alignment in values and beliefs. HPE emphasizes the need to find solutions to operate generative AI technology in a reliable, sovereign, and sustainable manner.

So, where do we stand now?

We are in a new phase where generative AI can produce novel outcomes for humanity. While the outputs of generative AI are based on cumulative intelligence and the data sets created by humanity, it is crucial to consider who uses this cumulative intelligence to train AI models.

At the core, most generative AI chatbots, which are predominantly based in the U.S., perform best in English, posing a significant risk for other societies. This could exacerbate the digital divide. The digital divide, which refers to inequalities in access to technological devices, the internet, digital literacy, and technology skills, goes beyond differences in individual income and capabilities. For example, a native English speaker may achieve better performance from a chatbot, for which they pay the same access fee, compared to someone whose native language is Turkish, Bosnian, Albanian, or Arabic.

Generative AI, which is rapidly adopted for the efficiency and effectiveness it enables in all components of global society, is at the center of concerns about monopolies.

Understanding the risks posed by global technology giants that can invest billions of dollars, have access to diverse data sets, and can easily attract human resources, as well as the chaotic structure they create in the sector, is a prerequisite for offering alternatives.

On the other hand, exaggerating the risks and threats in this process can lead to the mistaken belief that AI researchers cannot generate value unless they have power similar to that of monopolies. Despite their vast resources, it is significant that technology monopolies are increasingly acquiring innovative technologies through investments in start-ups rather than their internal capabilities. However, it is a grave mistake for a company that has not yet achieved any sales success to compare itself to giants like OpenAI or smaller players that have managed to secure millions of dollars in investment.

By pulling back from engineering ambitions and focusing on market needs, gaps, and existing in-house capabilities, every player can generate value in any critical area, including generative AI, through specialization in vertical markets.

However, for small companies to succeed in this field, conditions and decisions are required, such as offering computing power at fair prices and ensuring that cloud-based service providers do not differentiate certain AI APIs from their competitors. Unless the commercial ecosystem in Germany that supports Aleph Alpha is replicated in different versions for other companies, the pace of development will remain limited.

It is unrealistic to expect actors who tend to fill their inventories with products offered only by monopolies, even in low-cost software and hardware, to make significant decisions in areas requiring strategic investments like generative AI.

Therefore, it is clear that alternative players who:

• Have the ability to generate value in a defined area,
• Understand the power of global technology giants,
• Can make fair agreements with these giants when necessary,
• And receive support from regulatory mechanisms in their country and from partners in various sectors globally, can offer significant value.

It should not be overlooked that AI is on the path toward artificial general intelligence (AGI), a capacity to perform any intellectual task that a human can accomplish.

Countries with various risks concerning technological sovereignty have a very short time to take the necessary steps to strengthen cloud infrastructure, provide resources to AI experts, and become part of the global ecosystem.

In the near future, the difference between lagging behind and not participating at all will become negligible.

Source: “Egemen” Olmayı Sürdürmek: Üretken Yapay Zeka – SavunmaTR