A GenAI forecast: the most promising players of 2024
Taking a look at the large language models currently making waves in the AI industry
Ever since ChatGPT burst onto the scene in the fall of 2022, generative AI (GenAI) solutions have continued to disrupt industries with their surprising and never-before-seen features.
Many of the generative AI solutions we hear about today are based on large language models (LLM), which allow them to successfully communicate with end users. Every such model has to be pre-trained on vast amounts of data and then fine-tuned using a variety of different methods.
As we’re approaching the end of January, let’s look at the list of the most noteworthy of these models and explore the generative AI solutions they power. Compiled by the AIport newsletter team, this list is by no means exhaustive, but it is a review of what we believe are the LLM-based generative AI solutions and tools that are likely to make headlines this year.
GPT (OpenAI and Microsoft)
Expectedly, the first LLM on our list with some of the highest expectations this year is at the core of ChatGPT from OpenAI. GPT wasn’t the first LLM out there (there had been others like BERT), but it was certainly the first one that resulted in a market-ready GenAI product.
The original GPT model (Generative Pretrained Transformers) had 170 billion parameters, while the latest version, GPT-4, boasts over 1.75 trillion.
The model’s applications are far and wide, being the basis not only for ChatGPT, but also Microsoft’s Bing Chat, which is now part of the Copilot infrastructure. While ChatGPT remains arguably the most impressive text-producing GenAI tool, Bing Chat (unlike ChatGPT) has internet access, making it ideal for web browsing.
In addition, the same core GPT model is being utilized by other entities — from Panasonic Connect in Japan, OCBC Bank’s ChatGPT in Singapore, and KPMG’s KaiChat in Germany to the UK government’s upcoming GOV.UK chat. Notably, GPT is also a vital part of OpenAI’s DALL-E suite — a solution that generates images based on descriptions from text.
PaLM and Gemini (Google)
The next one on our list is Google’s PaLM (Pathways Language Model). The model, boasting 540 billion parameters, was first successfully tested in the spring of 2022. A year later, PaLM 2 was announced and subsequently integrated into Bard — Google’s GenAI chatbot — becoming available to test users around the world.
Bard is considered among the finest solutions for productivity, having internet access like Bing Chat, but being less about navigating the web and more about answering burning questions.
Bard was seriously bolstered at the end of last year with Google’s highly anticipated launch of the multimodal LLM, Gemini, which had been trained on images and audio/video files, not just text. The jury is still out on how far this update will take Bard, but the expectations are very high indeed.
LLaMA (Meta, Microsoft, and Hugging Face)
LLaMA (Large Language Model Meta AI) from Meta was released in February last year. However, the 65-billion-parameter model was leaked online as a downloadable torrent only a week later.
While the model did not lead to any ChatGPT-like products at Meta, another AI company — Hugging Face — picked it up and released Hugging Chat. Offering internet access, this GenAI solution has been praised as “the first open source alternative to ChatGPT” with a snarky sense of humor.
Not to be outdone, Meta partnered with Microsoft to produce LLaMA 2 in July of last year. Following that, the company announced a beta release of Meta AI at the end of September — a GenAI chatbot, currently available in the US, that can “provide real-time information and generate photorealistic images” from text prompts. This solution is expected to reach Meta users across the rest of the globe later in the year.
Claude (Anthropic)
Our next LLM is made by Anthropic, and it’s named Claude. Founded in 2021, the company along with its language model may be a less familiar name to most non-geeks. But this is likely to change in 2024.
That’s the case not only because the latest version of the namesake GenAI chatbot released last year can now process long PDF files of up to 75,000 words, taking moments to summarize entire books. But also because this has already caught the attention of both Google and Amazon who are keen to invest $2 and $4 billion respectively.
While Claude doesn’t offer internet access, its processing power, namely input character memory, outpowers every competitor by a wide margin, including ChatGPT. And the company is said to be working on yet more handy features as we speak.
HyperCLOVA X (Naver)
HyperCLOVA X is both the LLM and the namesake GenAI chatbot released by the South Korean search engine giant, Naver, last August. Boasting over 200 billion parameters, the LLM is being used mainly for AI-assisted web browsing, much like Bing Chat.
The most prominent feature of HyperCLOVA X is its proficiency in Korean. Having learned 6500 times more Korean words than ChatGPT, HyperCLOVA X is said to be especially useful for localized scenarios, where understanding natural Korean language and the cultural context is particularly important to provide accurate answers.
In addition to answering user queries, HyperCLOVA X has been designed to offer AI assistance to businesses across numerous sectors — from finance to gaming. Naver is also presently working to make HyperCLOVA X multimodal, following in the footsteps of Google’s Gemini, so another big international headline is likely on the way.
Pangu (Huawei)
Last spring, Huawei, one of the global tech leaders, unveiled Pangu — an LLM with 1.085 trillion parameters in over 40 natural and programming languages.
This LLM now powers the Chinese company’s GenAI assistant called Celia, which resides within Huawei’s HarmonyOS 4 operating system. Supposedly, Celia has already outperformed both Siri and Google Assistant as a phone-based virtual assistant in terms of prompt execution accuracy, such as finding specific photos and documents stored on a device. And as more Huawei smartphones are being churned out, Celia is expected to reach more global users this year with further refinements to the Pangu model.
In addition, Pangu is available through Huawei Cloud for commercial use. According to a statement from the company’s leadership, the newest Pangu 3.0 utilizes hierarchical architecture, allowing it to be quickly fine-tuned for a wide range of downstream applications (including Huawei’s autonomous vehicle solutions), which we’ll surely see more of in 2024.
MiLM-6B (Xiaomi)
Next up is another Chinese company, Xiaomi — one of the largest phone manufacturers in the world. The company made an entry into the LLM race last summer with its MiLM-6B, a lightweight model created for mobile devices.
Boasting 6.4 billion parameters, MiLM-6B has reportedly outperformed rivals in its category and achieved notable scores from evaluation platforms like C-Eval and CMMLU.
Xiaomi’s leadership has been vocal for some time about integrating LLMs into its smartphones. It appears that the company’s virtual assistant, Xiao AI, has now finally become a fully fledged GenAI chatbot, generating over 11 words per second. By this metric, despite operating locally on the phone, Xiao AI has matched the performance of some LLMs running on cloud computing services.
It’ll be interesting to see whether Xiao AI gets assimilated into the company’s other products, as the rumors suggest, including the recently announced Xiaomi SU7, an autonomous-enabled electric sedan.
YandexGPT (Yandex)
The final entry on our list is an LLM named YandexGPT from the Russian search engine Yandex. The namesake GenAI solution with 100 billion parameters was beta-released in May last year. The more recent release of YandexGPT2 has demonstrated a 65% improvement in response quality, with a 1.5-fold increase in training data. Interestingly, this GenAI solution was tested and scored enough points to enter a Russian university.
YandexGPT was also integrated into Yandex’s flagship virtual assistant named Alice, as well as the company’s smart speakers. Furthermore, the LLM has allowed the company’s e-shoppers to view summarized customer reviews in one place. With YandexGPT at the core, the company now also offers visual GenAI solutions, YandexART and Shedevrum, that can generate images from text, much like Open AI’s DALL-E.
Currently, the LLM is also being merged with Yandex Search to offer the end user something similar to Bing Chat and HyperCLOVA X. It’s likely that we’ll see a new announcement pertaining to this development later in the year.
New kids on the block
Apart from the above entries on our list, several brand new LLMs along with their GenAI solutions are expected to arrive any day from other big players. Among them are Samsung from South Korea, Tencent from China, as well as Amazon and Apple.
In fact, Apple already has an internal AI chatbot for employees, but its much-talked-about AppleGPT based on the company’s LLM, Ajax, is said to be around the corner. Amazon began to offer a service for building AI chatbots called Lex a while ago, but more notably, the company recently announced its upcoming GenAI solution tailored for commercial use.
Japan doesn’t intend to be left behind either — at least three major LLM initiatives are currently being played out. Among them is a joint project between Fujitsu, RIKEN, Tokyo Tech, and Tohoku University who are developing a series of LLMs using one of the world’s most powerful supercomputers, Fugaku. In addition, both NTT and SoftBank are reportedly in the final stages of developing their “homegrown” LLMs.
While much remains unclear and covered in mist (or is it gray goo 😉), one thing is certain — 2024 is going to be the most AI-intense year we have seen yet!