The following is English version of Chapter 2 of ‘One Man Conglomerate’, published in August 2023 by Do-Jeon, Marcus, Jeong.
2nd Chat: How will our lives changes?
#01. LLM technology will be more competitive and the pace of development will be faster.
[01] Competition for LLM technology among companies will become fiercer.
In the future, LLMs (Large Language Models) that will compete with ChatGPT will flood the market. This is because OpenAI’s ChatGPT (LLM-based AI chatbot) has already confirmed its infinite marketability. Therefore, LLM development activities of large companies with capital power will be more active.
First of all, Google, which is said to be ChatGPT’s biggest competitor, will be competing with OpenAI with its AI chatbot Bard. Of course, the real players in this race are Microsoft (MS) and Google. MS has been investing heavily in OpenAI and integrating ChatGPT into all of its services, making an aggressive move against Google.
Google had already embarrassed itself twice by releasing an older version of Bard (with LaMDA). Then they released Bard with PaLM2, which somewhat salvaged their bruised ego, and announced some rather aggressive plans to upgrade the non-Multimodal PaLM2 to Gemini (a language model currently under development) in the future. The intention is to go head-to-head with MS (including OpenAI’s ChatGPT). As a result, both companies will be fighting tooth and nail to gain a competitive advantage in the new era of AI. However, there is a high probability that this competitive composition will not last as long as expected.
And global companies that are well known to us, such as Meta, Amazon, Apple, Baidu, Tencent, Hauwei, and Alibaba, have also joined the competition. A closer look reveals that the competition is between the US and China. If you look at the LLMs released in recent years, the majority of the developers are from the US and China (see Table 3. Recent LLM releases).
In other words, the development of LLMs, which are the foundation of AI chatbots and will be the mainstay of future industries, is largely being done by companies from two rival countries that dominate the global economy. As things stand, the governments of these two countries are not just taking a backseat: on the one hand, they are cautiously saying that AI technology can be dangerous, but on the other hand, they are planning to use it to achieve global hegemony. What is dangerous to me is dangerous to enemy. This is why LLM technology in these two countries is likely to develop much faster than in other countries.
In addition, Korean companies such as Naver, Kakao, Samsung, LG, SKT, KT, and others have also entered the fray, but if they focus on local markets (including Korean-centric strategies), they are likely to struggle to capture profitable markets, because sooner or later, language barriers will disappear and AI will surpass human capabilities, which means that the borders of this market will eventually disappear as well, similar to the search engine market.
In any case, with this level of competition, LLM skills will be rapidly up-leveled. So the cost of learning will be significantly reduced and the model will be lightweight.
As this unfolds, players looking to capture market leadership will be interested in differentiation and entry barriers. First of all, differentiation can be implemented in many ways. For example, players can tie LLM to their other products or services, like MS and Google, or differentiate hardware by creating their own chips, like Google and Meta. And an example of a barrier to entry is convincing (or lobbying) a government or Congress to create an institutional (i.e., regulation) hurdle. This could lead to the kind of “conversational AI licensing” that Sam Altman, CEO of OpenAI, has advocated for in the US Congress.
developed by Midjourney
[02] In addition, individuals (including developers) will join the competition.
Individuals (mainly developers) will also be actively involved in this competition. Of course, the above-mentioned barriers to entry are unlikely to completely exclude individual developers, as they will mostly leverage the open source of the major players (e.g., the LLaMA of Meta). And negative regulation is far more favorable to technological progress than positive regulation.
In fact, the OpenAI CEO’s claim above may not be based on “concern for the future of humanity,” as the pace of individual developers now centered on Github poses a significant threat to OpenAI. This was, of course, unthinkable in the past, as the cost of developing a chatbot based on LLM was astronomical (especially in terms of learning costs and time). However, the achievements of individual developers in developing LLMs (and more specifically, AI chatbots using them), spurred on by Meta’s open source LLaMA, have taken everyone by surprise.
As you can see from the graph below, it took only 23 days after LLaMA’s source was released to create a sLLM (Small LLM) based chatbot (ⓒ) that could compete with Bard and ChatGPT. It was also an achievement achieved with just 300 dollars.
Graph 2. Relative Response Quality by Model assessed by GPT-4
Source: https://lmsys.org/blog/2023-03-30-vicuna/
So we’re likely to see more sLLMs like this in the future. Regardless of their commercial viability, they will certainly make the AI giants (e.g., OpenAI, Google, etc.) nervous, because if their products quickly spread around the world via Github or YouTube, which will eventually impact their paid services market.
So, there may be a situation where ‘giant AI companies’ cannot overcome sLLMs even after developing LLMs with huge funds. This is because most organizations called corporations are more rigid than individuals.
And this situation is likely to expand to the entire generative AI market. Because this possibility has already been shown by the Stable Diffusion (text-to-image model).
[03] This will drive LLM technology forward very quickly.
In general, if a market ❶is profitable enough (both now and in the future), ❷has a leading company, and ❸has enough competitors, the speed at which the market can develop is unimaginable, because the goal becomes clear. So no one can work slowly.
As the market heats up, ①there will be more opportunities for LLM developers to secure funding. Especially in Korea, there are a lot of companies that are trying to make money through investment rather than their main business, because the main business is not as profitable as before due to environmental changes such as demographics.
②And There will be a lot of activity across the industry, in terms of service and technology partnerships and M&A. ③In addition, if the lightweight of LLM technology is commercialized early, technologies in the field of artificial intelligence, as well as all other fields related to or utilizing it, will advance at a much faster pace than today.
And this will accelerate the emergence of AGI and ANI. So It will be able to recognize user commands (prompts) in more ways than today (e.g., text or images) and deliver the desired answer. And, of course, communication and collaboration between LLMs will be possible. Thus, even large-scale projects between different organizations or countries will be carried out with minimal human intervention.
Table 5. Major LLMs already released or to be released in the future (As of Aug. 2023)
| Open Source | Closed Source |
| T5(Oct 2019, Google AI), mT5(Oct 2020, Hugging Face), PanGu-α(Apr 2021, Huawei), CPM-2(Jun, BAAI), T0(Oct, Hugging Face), CodeGen(Mar 2022, Salesforce), GPT-NeoX-20B(Apr, EleutherAI), Tk-Instruct(Apr, AI2), UL2(May, Google AI), OPT(May, Meta), NLLB(Jul, Meta), CodeGeeX(Sep, ZhipuAI), GLM(Oct, Tsinghua Uni.), Flan-T5(Oct, Google AI), BLOOM(Nov, Hugging Face), mT0(Nov, Hugging Face), Galactica(Nov, Meta), BLOOMZ(Nov, Hugging Face), OPT-IML(Dec, Meta), LLaMA(Feb 2023, Meta), Claude(Mar, Anthropic), Dolly(Mar, Databricks), Dolly 2.0(Apr, Databricks), Pythia(Apr, EleutherAI), MPT-7B(May, Mosaic ML), Falcon(Jun, TII), LLaMA2(Jul, Meta), Claude 2(Jul, Anthropic) | GShard(Jun 2020, Google AI), GPT-3(May, OpenAI), HyperCLOVA(May 2021, Naver), Codex(Jul, OpenAI), ERNIE 3.0(Jul, Baidu), Jurassic-1(Aug, AI21), FLAN(Sep, Google AI), Yuan 1.0(Oct, Inspur), Anthropic(Dec, Anthropic), WebGPT(Dec, OpenAI), Gopher(Dec, DeepMind), ERNIE 3.0 Titan(Dec, Baidu), GLaM(Dec, Google AI), LaMDA(Jan 2022, Google AI), MT-NLG(Jan, Microsoft & NVIDIA), AlphaCode(Feb, DeepMind), InstructGPT(Mar, OpenAI), Chinchilla(Mar, DeepMind), PaLM(Apr, Google AI), AlexaTM(Aug, Amazon), Sparrow(Sep, DeepMind), WeLM(Sep, WeChat AI), U-PaLM(Oct, Google AI), Flan-PaLM(Oct, Google AI), Flan-U-PaLM(Oct, Google AI), GPT-4(Mar 2023, OpenAI), PanGu(Mar, Huawei), Tongyi Qianwen (Apr, Alibaba), AppleGPT(Jul, Apple), EXAONE2.0(Jul, LG), Varco(Aug, NCSOFT), Konan(Aug, Konantech), HyperCLOVAX(Aug, Naver), KoGPT(TBU, Kakao), Gemini(TBU, Google) , Samsung Gauus and Gais(TBU, Samsung) |
Table 6. Major LLMs application models(As of Aug. 2023)
| ChatGPT(GPT-3.5 or higher) | LLaMA |
| AutoGPT(Mar 30 2023, Significant Gravitas), AgentGPT(Srijan Subedi, Asim Shrestha, Adam Watkins, Apr 2023), AskUp(Mar 2023, Upstage) |
Alpaca(Mar 13 2023, Stanford University), Vicuna(Mar 19, 2023, LMSYS ORG), GPT4All(A model with a fine tune of LLaMA-7B, Mar 2023, Nomic AI), Koala(Apr 2023, Berkeley Artificial Intelligence Research Lab), WizardLM(Apr 2023, Microsoft), OpenAssistant RLHF (Apr 2023, OpenAssistant), StableVicuna(Apr 2023, Stability AI), upstage/Llama-2-70b-instruct-v2(A model with a fine tune of LLAMA 2, Upstage) |
※ Note that the above are only representative models. This is because there are a lot of models that apply the existing LLM.
[ Attention ]
The above article excludes footnotes and details, unlike the actual book.
Please read the following post for details of “#02”
위 글은 정도전 작가가 2023년 8월에 출간한 ‘1인 대기업(One Man Conglomerate)’의 챕터2 영문 버전입니다.
Sample page: One Man Conglomerate, 80p
