What You Need to Know About AI Today

Introduction:

Artificial intelligence (AI) is undergoing rapid transformation, largely due to the development of large language models (LLMs). These deep learning algorithms utilize extensive amounts of training data in natural language processing (NLP), encompassing text, images, sound, and video. Supervised and unsupervised learning techniques are employed to process and generate natural language and other structured outputs. In a previous post, I covered the exponential pace of AI and LLM development, highlighting the unanswered question of how society should adapt. In this article, we will explore the latest updates and announcements in the field of AI, focusing on LLMs, their applications, and the associated business implications. We will also discuss strategies to improve accuracy and reduce misuse and risks with these models.

We'll delve into some of their capabilities and reference the information which I have found useful, discuss their applications that broadly range in two key categories, in my opinion, so far, of – a) productivity and efficiency improvements with augmentation, to b) business models – both disruptions and new, and then highlight potential pitfalls that come with their use.

The topic of #generativeai AI is top of the mind for every company and its management and leadership, notwithstanding the obvious hype that accompanies such profound developments. In various conversations, podcasts and discussions on this topic, parallels are obviously drawn with examples in history from the agricultural and industrial revolution to the introduction of personal computers and internet in early 90’s, to the introduction of mobile phones earlier in the first decade of this century for example.

Foundation LLMs:

Major corporations such as 微软 and OpenAI, Google, Meta, and others develop foundation LLMs with access to vast resources and training data. Examples of these LLMs include #gpt3, GPT-3.5, #gpt4 by OpenAI , PaLM and #palm2 by 谷歌 , #llama, and SAM by Meta . Specialized models like DALL.E by OpenAI, one from Midjourney, and Stable Diffusion by Stability AI focus on generating images from text or natural language inputs. These models have been introduced in the past year and continue to make significant improvements through reinforcement learning for all the ones based on #gpt. Each model has unique features, such as the number of parameters, corpus size, training data types, license, context size, and performance.

It's worth noting that LLMs vary in openness and accessibility. Some models have become closed, like those from OpenAI and 谷歌 , while others follow an open-source strategy, such as Meta , EleutherAI , and Hugging Face . While the specifics of closed vs. open models are beyond the scope of this write-up, you can read a summary about this battle at the MIT Technology review here. To better understand the technical features and differences, I have provided some references below.

Domain and Vertical Applications and Use Cases:

The development of LLMs has led to intriguing and exciting applications across various domains. APIs and private instances are now available for different task types and use cases, catering to both enterprise-level B2B and consumer-level B2C applications. Mustafa Suleyman, co-founder of DeepMind and now at Inflection AI, aptly describes this as a paradigm shift, a flip, where machines have learned to interact with humans effectively in natural language and conversation, encompassing text, images, video, and sound. This as opposed to us learning low-level and high-level programming languages to make machines and computers do what we need them to.

These advancements have far-reaching implications for work, business, and society. Firstly, there are significant improvements in productivity and efficiency across various domains. For example, ChatGPT, powered by GPT3.5 from OpenAI, has democratized the creation of chatbots and specialized agents for different domains. This has the potential to revolutionize call centers and customer service-focused businesses. In my own experience, using a general-purpose storytelling and presentation generation tool like Tome (reference below) reduced the time needed to create a basic deck and collaborate with colleagues, resulting in a 50% improvement in review time.

Second, as we look deeper, similar productivity and business model disruptions are observed in digital advertising and content publishing, where personalized, real-time, multi-modal content generation is transforming the industry. Case in point here – recent announcements by Microsoft on Ads for chat API and also Amazon’s plans to enable merchants to generate photos, videos and content for Ads using AI, here.

In the B2B space, various LLMs and derivative applications are enabling content automation, app generation, website and app creation, conversational agents, low-code automation, and more. These applications have the potential to reshape user experience, emphasizing natural language and prompt engineering. Some of them are outlined below -

·??????MetaGPT/Pico enables content automation and app generation albeit very simple at this point (one or two screen inputs type apps now – (#nocode ).

·??????Build.ai facilitates website and app creation – (#lowcode ). One of the implications with these examples is that the paradigm of User Experience (design, points and clicks) as we know it, is going to go through a profound inflection that is will, in the future, be more focused on natural language and prompt engineering.

·??????#chatgpt offers conversational agents and other variants around education and tutoring domains. Also, with API’s to enable companies to create their own Apps.

·??????Inflection AI/PI founded by Reid Hoffman and Mustafa Suleyman, offers different conversational agents powered by AI.

·??????Tray.io/MerlinAI – low-code automation and integration platform enabling developers to secure automate complex workflows within and across multiple systems.

·??????Anthrop\C Claude – a next generation AI assistant for tasks and workflows, no matter the scale (size and time of the context). Its still limited though firms are pushing the envelope continuously.

·??????There are innumerable examples of Apps in conversational agents, personal tutoring and education out there already. Arguably, these could be B2B or B2C, but I have Chegg business model disruption on my mind at the moment by ChatGPT and have included it in this category.

In the B2C space, a plethora of Apps have been introduced both on web and mobile. Some of them are -

·??????Character AI enables interactive storytelling, all with a disclaimer that it’s all fictional. But the characters can range from personal assistants, cynical assistants (for a bit of humor), Psychologists, SB Mario and more. I tried the cynical assistant for a bit of fun and it was engaging. Compared to the other examples, I began with being a bit recalcitrant here, but in a few one-word interactions it began asking me questions and we went on from there. However, in the intro version, it cuts off after a handful of interactions and leads you to a registration and subscription (quick monetization play here).

·??????Replika provides an emotional support assistant.

·??????Lyrebird specializes in voice cloning and related content generation for podcasts for example.

Policy and Misinformation Potential: While LLMs offer immense potential to benefit society and revolutionize various industries, they also present risks and social challenges. Fake content, bias, privacy concerns, accountability, and regulation are among the issues that need to be addressed. Much like we deal with today on email, websites and phone scams but only more. A quite insightful article on why generative AI is more dangerous than traditional content creation tools is by Louis Rosenberg here. Geoffrey Hinton (ex-Google) made a recent plea about these challenges.

There is no doubt we need guardrails for this. Governing policy kernels and frameworks need to be developed. But as in all technological advances, particularly regarding the rapid transformation and democratization with generative AI, establishment of corporate policies and governance will lead any development of government policy or standards. Executives today are rightly concerned about how to keep their data secure and ensure there are no accidental incidents, not only regarding data leaks to public cloud based LLMs, but also unintended negative brand impacts due to their use. The fact that LLMs hallucinate and present some incorrect aspects emphatically as fact, is concerning, especially due to the pace ubiquity of these LLMs and derivative Apps. Hence, corporate governance is key here – I would argue the need to evolve the CxO – CPO/CDO/CIO role an AI Czar, with responsibility to outline corporate policies, advocate for data transparency, watermarking, safety, and verification regarding the use of LLMs or mini-LLMs in their businesses – especially in customer facing applications and interactions. For the organizations introducing LLMs to the market - there will always be bad actors and misuse out there, however these corporations have a great opportunity AND responsibility to enable the use of LLMs for the improvement of society and humankind overall and improve their brands in the process. They need to develop clear guidelines and regulations for their use. API’s only access for LLMs is one-way to start with having some modulated control here as articulated by Sam Altman, founder of OpenAI. The risks of closed LLMs vs Open-source LLMs is a continuing discussion point in this context. Public education and awareness programs to make people aware of the potential risks associated with this technology so they are aware of the potential dangers and harmful outcomes and can protect themselves.

Conclusion: In this article, we've explored the advancements in #artificialintelligence, focusing on large language models (LLMs). We enumerated some foundation LLMs developed by major corporations and referenced their unique features and comparisons. We also delved into the domain-specific applications and use cases of LLMs, both in B2B and B2C contexts that lead to significant productivity gains and business model implications. Additionally, we addressed and outlined the potential pitfalls and ethical challenges associated with generative AI, emphasizing the importance of responsible use and some mitigation strategies. Overall, LLMs and derivative applications have immense potential to improve productivity at work and revolutionize various industries, but careful consideration is required to ensure their positive impact on society.

Additional References:

1. GPT-4 vs GPT-3.5 Analysis: A comprehensive analysis.
2. PaLM vs PaLM 2 Comparison: A comparative study of two large language models
3. LLama Language Model Overview: An overview of a powerful language model
4. TechRepublic GPT-4 Cheat Sheet: What is GPT-4 & what is it capable of?
5. VentureBeat Generative AI: Why generative AI is more dangerous than you think.
6. Tome - Marketing and general presentations powered by AI.
7. Inflection AI - Conversational agents powered by AI.
8. Character.AI AI - Interactive storytelling powered by AI.
9. BuildAI - Website creation powered by AI.