Generative AI in Business: ChatGPT is a “jack of all trades, master of none”

Read the interview with Piotr Bombol, Founder of adaily—a platform offering AI tools for marketers—about how Generative AI will change businesses across industries and why what we see now is just the beginning.”

The key insights

• ChatGPT revolutionized conversational interfaces, showing AI can sustain more dynamic and natural interactions. This is just the start of broader AI advancements.
• While there’s high consumer interest, business adoption has been slower. Many employees use generative AI without formal approval, risking sensitive data leaks.
• Outputs from generative AI aren’t copyright-protected, posing challenges in creative industries. Additionally, models like ChatGPT often produce “hallucinations”—convincing but incorrect responses.
• The initial phase focused on tweaking AI parameters, but the future lies in vertical applications tailored to specific industries, addressing inherent AI risks. Notable examples include Adobe and Notion’s AI integrations.
• Fine-tuning adds specific knowledge to models but is resource-intensive. Vector databases improve search accuracy and reduce hallucinations by managing unstructured data effectively.
• APIs are crucial for leveraging AI capabilities without needing dedicated apps.

Addepto: How has ChatGPT changed the world?

Piotr Bombol, CEO of Adaily: It was the beginning of a revolution, without a doubt. ChatGPT demonstrated that conversational interfaces don’t have to be as limited as the bots we were accustomed to in the past. Today, we observe that AI can initiate and sustain a regular conversation. However, this is merely the beginning. Conversational AI is rapidly evolving into a precursor to something far more significant.

Isn’t the adoption of Gen AI-based solutions lower than one might anticipate?

Among consumers, there’s significant hype surrounding Gen AI, often perceived as a universal tool. It was only later that businesses began exploring ways to harness Gen AI for their distinct needs. GPT, in its essence, is a fundamental, general-purpose tool beneficial to companies. However, it’s not without its risks.

Over a quarter (28%) of workers are currently using generative AI at work, and over half without the formal approval of their employers (Source: Salesforce.com)

Employees with unrestricted access to GPT might inadvertently disclose sensitive business-related information. This is primarily because they often find it challenging to discern between confidential and non-confidential information. Regrettably, ChatGPT in its free version, as opposed to the enterprise version, utilizes all provided information for its training. Data coming to the API is also not used, which represents a missed business opportunity.

Users, yielding to the innate temptation to reduce their workload, might delegate their entire tasks to AI and subsequently replicate the results. This behavior is colloquially termed as “falling asleep at the wheel.”

An experiment by the Boston Consulting Group revealed that ChatGPT typically accelerates and enhances the quality of consultants’ work. However, it can also induce a sense of complacency in them, given that their responses are consistently convincing (Source: Papers SSRN)

Second, there are legal issues. The output of language models isn’t protected by copyright. Therefore, in the creative industry, for example, you can’t transfer its rights to clients. Not to mention the legal implications surrounding the datasets on which Gen AI models have been trained.

So, is ChatGPT not suitable for companies?

ChatGPT is a “jack of all trades, master of none.” Initially, we were dazzled by its conversational capabilities. However, when we attempt more specific tasks, its limitations become evident. In my view, these limitations stem from two main factors:

GPT is a predefined interface, meaning it has set conversation parameters designed to meet general needs. This, however, does not imply that GPT as a model lacks other capabilities. OpenAI has provided users with a ‘Playground’ mode where all parameters, including those affecting the randomness of responses, can be modified.

It experiences ‘hallucinations,’ which are not bugs but rather features. GPT is engineered to consistently provide a convincing response rather than admitting its lack of knowledge in certain areas.

But even with its predefined behavior and hallucinations, ChatGPT finds widespread use.

It is a constant evolution. Initially, the ability to tweak parameters was capitalized upon by ‘wrappers’—products that essentially monetized what was available via API even before ChatGPT, and with the ChatGPT launch became free. A prime example is JasperAI. Its basic plan, priced at $49 a month, costs two and a half times more than ChatGPT Plus.

This trend, which I consider the ‘first wave’, seems to be waning. Merely adjusting parameters or adding specific instructions to GPT doesn’t resolve inherent issues, such as hallucinations.

A notable incident involved a US lawyer who, relying on a ChatGPT-based program, drafted a legal argument. This document contained entirely fake court citations, leading the lawyer into significant trouble due to breaches in professional ethics (read full story: The Guardian).

The ‘next wave’ is characterized by vertical applications tailored to specialized knowledge areas. Numerous businesses are emerging, aiming to harness the power of Gen AI while mitigating its associated risks. Major corporations are successfully integrating AI components into their software to boost user-friendliness, with Adobe and Notion being notable examples. Concurrently, standalone applications like Perplexity are being developed to tackle specific Gen AI challenges. Perplexity tries to collect information from the Internet, add an appropriate filter, and build a tailored answer.

Furthermore, there are AI applications that synergize specific industry data with the GPT interface. Take HubSpot, for instance—a standard CRM that, with the aid of ChatGPT, has significantly streamlined data access and utilization. Users can effortlessly ask a question and receive an answer via a so-called ChatSpot.

In my perspective, these trends will shape the future of Gen AI in the business realm, catering to precise, niche requirements and delivering tangible business value.

So, it’s not merely about changing parameters but more about fine-tuning?

Not exactly. In the fine-tuning process, we introduce additional knowledge packets to train the model. While this doesn’t eradicate hallucinations entirely, it equips the model with supplementary information. However, this method demands extensive, meticulously detailed knowledge bases and substantial computational resources, translating to increased costs.

Another strategy, which we at Adaily employ and has also emerged as an industry norm, revolves around vector databases. Vector databases are intended to provide the ability to search for information contextually, which is not possible with standard databases that rely on keywords (…)

Read more in Gen AI in Business: Global Trends Report 2024