CINXE.COM

<!DOCTYPE html> <html lang="en"> <head> <meta http-equiv="content-type" content="text/html;charset=UTF-8"> <meta http-equiv="pragma" content="no-cache"> <meta name="robots" content="noarchive"> <meta name="viewport" content="width=device-width, initial-scale=1.0"> <link rel="canonical" href="https://www.universityworldnews.com/post.php?story=20240510092346341" /> <title>Open-source LLMs – AI researchers are spoiled for choice</title> <meta property="twitter:site" content="@uniworldnews" /> <meta property="twitter:card" content="summary_large_image" /> <meta property="og:site_name" content="University World News"> <meta property="og:title" content="Open-source LLMs – AI researchers are spoiled for choice"> <meta property="og:description" content="AI researchers are spoiled for choice in large language models or LLMs – something that was almost inconceivable a year ago. There is both space and..."> <meta property="og:type" content="article"> <meta name="keywords" content="university, news, academic, higher education, education, transformative, leadership, academic freedom"> <meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1" /> <link rel="search" href="https://www.universityworldnews.com/fullsearch.php" title="Advanced Search"> <link rel="stylesheet" href="https://www.universityworldnews.com/mobile/css_reset.css" /> <link rel="stylesheet" href="https://www.universityworldnews.com/mobile/css_main_desktop.css" /> <link rel="stylesheet" href="https://www.universityworldnews.com/mobile/bulma-uwn.css"> <link rel="stylesheet" href="https://www.universityworldnews.com/mobile/superfish.css"> <script src="https://www.universityworldnews.com/mobile/bankersbox.js"></script> <script src="https://www.universityworldnews.com/mobile/jquery-3.3.1.min.js"></script> <script defer data-domain="universityworldnews.com" src="https://plausible.io/js/plausible.file-downloads.outbound-links.pageview-props.js" event-pageid="article-global"></script> <style> .container { max-width: 960px; width: 960px } .container.is-fluid { margin-left: 64px; margin-right: 64px; max-width: none; width: auto } .mobile-hidden { display: none; } .desktop-hidden { display: none; } .always-hidden { display: none; } .always-visible { display: block; } @media (min-width: 960px) { .mobile-hidden { display: block; } } @media (max-width: 960px) { .desktop-hidden { display: block; } } .navbar.has-shadow { box-shadow: 0 2px 0 0 #000; } </style> <meta property="og:image" content="https://www.universityworldnews.com/images/articles/20240510092346341_5_1200.jpg?cb=12866"><meta property="twitter:image" content="https://www.universityworldnews.com/images/articles/20240510092346341_5_1200.jpg?cb=12866"> <script>dataLayer = [{ "page_type": "article", "device_type": "desktop", "pub": "global", "publish_year": "2024", "publish_month": "2024-05", "authors": [ "Jean-Paul Van Belle" ], "writers": "Jean-Paul Van Belle", "section": "special-reports", "special_report": "AI and Research", "categories": "EdTech\/AI, Research and Innovation", "pageCategory": [ "EdTech\/AI", "Research and Innovation" ], "global_edition": "785", "africa_edition": "450" }];</script> <script>(function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start': new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0], j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src= 'https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f); })(window,document,'script','dataLayer','GTM-MWC5B4F');</script> </head> <body class="layout-documentation page-layout"> <div id="wrapper"><div id="mainsection"> <div id="the-top"></div> <div id="ad1"> <div style="padding-top:12px;padding-bottom:6px;text-align:center;border-bottom:1px solid #999"><a href="https://ads.universityworldnews.com/bannerclick.php?id=nafsanov2024ldr" id="zone-leaderboard-1-link"><img src="https://ads.universityworldnews.com/bannerview.php?id=nafsanov2024ldr" alt="Leaderboard" id="zone-leaderboard-1-image" style="width:970px;height:90px"></a></div> </div> <div style="float:right;padding-top:36px;padding-right:12px"><form name="searchbox" action="fullsearch.php" method="get"><input style="width:190px" class="input is-rounded" type="text" placeholder="Search" name="query"><input type="submit" style="height: 0px; width: 0px; border: none; padding: 0px;" hidefocus="true" value="Search"/><input type="hidden" name="mode" value="search"></form></div> <div id="header" style="border-bottom:2px solid #000;padding-top:8px;height:112px"><a href="https://www.universityworldnews.com/page.php?page=UW_Main"><img src="https://www.universityworldnews.com/mobile/uwn_logo_oldUWorld.png" style="" alt=""></a> </div> <div style="border-bottom:2px solid #000;"> <ul class="sf-menu" id="main_menu" style="font-size:14.5px"> <li> <a href="page.php?page=UW_Main">Global</a> <ul> <li><a href="page.php?page=UW_Main">Home</a></li> <li><a href="topic-page.php?topic=News">News</a></li> <li><a href="page.php?page=Coronavirus_Hub">Coronavirus Crisis</a></li> <li><a href="topic-page.php?topic=WorldRoundup">World round-up</a></li> <li><a href="topic-page.php?topic=Commentary">Commentary</a></li> <li><a href="topic-page.php?topic=World_Blog">World Blog</a></li> <li><a href="topic-page.php?topic=Feature">Features</a></li> <li><a href="page.php?page=Trans_Lead">Transformative Leadership</a></li> <li><a href="topic-page.php?topic=Academic_Freedom">Academic Freedom</a></li> <li><a href="special-report.php?publication=global">Special Reports</a></li> <li><a href="topic-page.php?topic=StudentView">Student View</a></li> <li><a href="topic-page.php?topic=U-Say">Letters</a></li> <li><a href="page.php?page=subscribe">Subscribe Newsletter</a></li> </ul> </li> <li> <a href="page.php?page=Africa_Edition">Africa</a> <ul> <li><a href="page.php?page=Africa_Edition">Home</a></li> <li><a href="topic-page.php?topic=AfrNews">Africa News</a></li> <li><a href="page.php?page=Coronavirus_Africa_Hub">Coronavirus Africa</a></li> <li><a href="topic-page.php?topic=AfrCommentary">Africa Commentary</a></li> <li><a href="topic-page.php?topic=AfrFeatures">Africa Features</a></li> <li><a href="page.php?page=Student_Hub_Africa">Africa Student View</a></li> <li><a href="page.php?page=subscribe">Subscribe Newsletter</a></li> <li><a href="special-report.php?publication=africa">Special Reports</a></li> </ul> </li> <li><a href="page.php?page=Asia_Hub">Asia</a></li> <li><a href="page.php?page=SDGs_Hub">SDGs</a></li> <li><a href="page.php?page=Edtech_AI_Hub">Edtech/AI</a></li> <li><a href="page.php?page=Commentary_Hub">Commentary</a></li> <li> <a href="special-report.php?publication=global">Special Reports</a> <ul> <li><a href="special-report.php?publication=global">Global Special Reports</a></li> <li><a href="special-report.php?publication=africa">Africa Special Reports</a></li> <li><a href="page.php?page=Coronavirus_Hub">Coronavirus Crisis</a></li> <li><a href="page.php?page=Coronavirus_Africa_Hub">Coronavirus Africa</a></li> </ul> </li> <li> <a href="https://www.universityworldnews.com/page.php?page=sponsor-partner-profservices">Partnerships</a> <ul> <li><a href="https://www.universityworldnews.com/page.php?page=sponsor-partner-profservices">Creative Services</a></li> <li><a href="https://www.universityworldnews.com/page.php?page=Advertising">Advertising</a></li> </ul> </li> <li> <a href="fullsearch.php">Search</a> <ul> <li><a href="fullsearch.php">Advanced Search</a></li> <li><a href="fullsearch.php?mode=country_list">Articles by Country</a></li> <li><a href="newsletter-archives.php?pub=UWorld">Global Newsletter Archives</a></li> <li><a href="newsletter-archives.php?pub=Africa_Edition">Africa Newsletter Archives</a></li> </ul> </li> <li> <a href="page.php?page=About_Us">About Us</a> <ul> <li><a href="page.php?page=About_Us">About UWN</a></li> <li><a href="page.php?page=subscribe">Subscribe Newsletter</a></li> <li><a href="page.php?page=Advertising">Advertising</a></li> <li><a href="page.php?page=sponsor-partner-profservices">Sponsorship</a></li> <li><a href="page.php?page=Terms_Conditions">Terms & Conditions</a></li> <li><a href="page.php?page=Contact_Us_Intro">Contact Us</a></li> </ul> </li> </ul> </div> <div class="wdiv p8px" id="content"> <table><tr> <td style="width:235px;padding-right:10px;padding-top:6px" valign="top"><div style="background:#eee;margin-top:3px"><img src="https://www.universityworldnews.com/images/articles/20240510092346341_5.jpg" style="" alt="Image: iStock"><div style="padding:4px 8px;font-size:12px">Image: iStock</div></div><div style="border-top:2px solid #000;;border-bottom:2px solid #000;margin-top:20px;margin-bottom:18px;padding:8px;">Join us on <a href="https://www.facebook.com/UniversityWorldNews/" target="_fb"><b>Facebook</b></a><br>Follow us on <a href="https://twitter.com/uniworldnews" target="_tw"><b>Twitter</b></a></div><div class="section-head-white-on-black botmar8">Related Links</div><span class="story-country">GLOBAL</span><br><a href="post.php?story=20240424125411397" style="font-weight:bold">Integrating AI in academic research – Changing the paradigm</a><br><span class="story-country">GLOBAL</span><br><a href="post.php?story=20240313142939926" style="font-weight:bold">Small language models: A cheaper, greener route into AI</a><br><span class="story-country">UNITED KINGDOM</span><br><a href="post.php?story=20240508134752792" style="font-weight:bold">AI levels the playing field for researchers applying for grants</a><br><span class="story-country">GLOBAL</span><br><a href="post.php?story=20240405145123397" style="font-weight:bold">Report highlights strategies to accelerate AI in research</a><br><span class="story-country">SOUTH AFRICA</span><br><a href="post.php?story=20240503111715657" style="font-weight:bold">AI revolutionises astronomy – and understanding of the universe</a><br><span class="story-country">DENMARK</span><br><a href="post.php?story=20240319144103814" style="font-weight:bold">Research explores how AI might change knowledge production</a><br><br><div class="section-head-white-on-dark botmar8">Featured Employers</div><div style="border:1px solid #fff; margin-bottom: 4px"><a href="https://ads.universityworldnews.com/bannerclick.php?id=uaeulogo223"><img src="https://ads.universityworldnews.com/bannerview.php?id=uaeulogo223" alt="UAEU Logo" width="223" height="74" border="0"></a></div> </td> <td style="width:489px;margin-left:10px;border-left:1px solid #999;padding-left:10px;padding-right:10px" valign="top"> <div><div class="article-body"><div class="story-country">GLOBAL</div> <img src="https://www.universityworldnews.com/mobile/bookmark_icon44.png" alt="bookmark" id="bmi" style="float:right" onclick="bookmark('20240510092346341','Open-source LLMs – AI researchers are spoiled for choice');"> <h1 class="botmar4 biggerheading">Open-source LLMs – AI researchers are spoiled for choice</h1> <div class="byline-country botmar12"><b><a href="https://www.universityworldnews.com/fullsearch.php?mode=search&writer=Jean-Paul+Van+Belle" style="white-space: nowrap">Jean-Paul Van Belle</a></b>  <span style="white-space: nowrap">10 May 2024</span></div> <div style="height:39px"><script src="https://platform.linkedin.com/in.js" type="text/javascript">lang: en_US</script> <div style="float:left;margin-right:8px"><script type="IN/Share" data-url="https://www.universityworldnews.com/post.php?story=20240510092346341"></script></div><div style="float:left;margin-right:8px"><a href="https://twitter.com/share?ref_src=twsrc%5Etfw" class="twitter-share-button" data-show-count="true">Tweet</a><script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script></div><iframe src="https://www.facebook.com/plugins/share_button.php?href=https%3A%2F%2Fwww.universityworldnews.com%2Fpost.php%3Fstory%3D20240510092346341&layout=button_count&size=small&mobile_iframe=true&width=106&height=28&appId" width="106" height="28" style="border:none;overflow:hidden" scrolling="no" frameborder="0" allowTransparency="true" allow="encrypted-media"></iframe></div> It is hard to believe that it is not even a year-and-a-half ago that ChatGPT propelled AI into the global mindset, unleashing a wave of excitement, hope, hype and investment as well as dire warnings. ChatGPT is based on the GPT large language model (LLM), which is itself an application of a deep learning neural network.<br /> <br /> Combining massive datasets (many terabytes of tokens), advances in algorithms and optimised hardware (GPUs) parallelised across a cloud computing infrastructure enabled researchers to use fairly classic neural network architectures with some added recent developments such as generative adversarial networks (GANs) and transformer models to build very deep models (for example, more than 100 hidden layers) comprising many billions or even over one trillion parameters.<br /> <br /> The capabilities of this current generation of LLMs – as a generative AI system – surprised even some seasoned AI experts.<br /> <br /> Simply running these LLMs to generate tokens in real time already requires dedicated GPU hardware; but building – that is, training – the more well-known LLMs took massive investments – often tens of millions of US dollars.<br /> <br /> It comes therefore perhaps as a surprise that many well-known LLMs are available as open-source. This article aims to explain what it means for an LLM to be ‘open-source’; why and which LLMs are being made open; the benefits and issues of open LLMs; some recent trends; and finally what some opportunities of open-source LLMs present to researchers in the developing world.<br /> <br /> <b>What qualifies as an open-source LLM?</b><br /> <br /> There are different degrees to which LLMs are ‘opened up’. The most basic approach is to simply publish the module structure and the trained parameters – that is, the weights and biases of all layers, for other researchers’ ‘fair use’.<br /> <br /> However, this cannot really be seen as open-source. As <a href="https://twitter.com/percyliang/status/1708560401754202621" target="_new">Percy Liang</a> states: “In software, this is analogous to releasing a binary without code (you wouldn’t call this open-source). To get the full benefits of transparency, you need the training data.”<br /> <br /> In fact, there is considerable nuance along the open-source ‘continuum’, ranging from ‘just publishing the weights’ to detailing the entire pipeline or process of training the model, the various hyper-parameters (perhaps also specifying the ranges of these values that were explored), seeds for various random initialisations, the particular software libraries and hardware specifications used for training, and the (usually massive) data set.<br /> <br /> The most onerous requirement is usually the data set since it has immense storage and bandwidth requirements in addition to potential governance and legal issues relating to data that may have restrictive or unclear copyright.<br /> <br /> This also raises a final requirement: a true open-source LLM should have a generous use licence: if its licence states that the published model can only be used for personal research purposes, then its value is mainly educational or for validating the research. Ideally, an open-source LLM model licence allows unlimited commercial use as well as the creation of derivative models.<br /> <br /> <b>What just happened? Aka you could not made this up if you tried</b><br /> <br /> GPT was developed by OpenAI in the United States. OpenAI was founded in 2015 specifically as a non-profit with the <a href="https://openai.com/about/" target="_new">aim</a> of attracting the top AI researchers to develop AI and make its research, products and patents open to the public in order to ensure that the future of AI would not be beholden to for-profit companies.<br /> <br /> But in 2019 OpenAI transitioned to a ‘capped for-profit’ model and ‘simultaneously’ attracted a one-billion dollar investment from Microsoft. Just the following year they announced GPT-3, followed soon by the release of DALL-E (AI image generation from pictures) and then ChatGPT, the revolutionary chatbot based on GPT-3.5.<br /> <br /> But, whilst offering a limited free use option, none of these models were opened up. Although this closing off went arguably directly <I>against</I> the founding philosophy of OpenAI, its executive stated that this was motivated by the risk of abuse of its AI technology.<br /> <br /> Thus, just last year in early 2023, the future for independent and non-profit AI research looked bleak: the most promising AI technology had just closed the door on open access, and all other state-of-the-art AI technology players were huge corporates with a historic tradition of keeping their competitive technology proprietary.<br /> <br /> There was simply no hope that smaller pure research players would have the deep pockets (US$10 million to US$100 million) required to train LLMs.<br /> <br /> But then, rather unexpectedly and quite counter-intuitively, several of the other large <I>commercial</I> players started releasing some versions of <I>their</I> LLMs as fully or partially open. Surprisingly, the arguably strongest open-source product, the series of <a href="https://llama.meta.com/" target="_new">Llama models</a>, was released by Meta AI, which is owned by perhaps the least sharing-oriented company of them all: Facebook.<br /> <br /> Apple, another notoriously open-source eschewing company, has <I>also</I> released a series of Open LLMs, OpenELM, even though these are ‘small’ versions intended to run on edge devices such as mobile phones, tablets and lower-end laptops. Equally, Google and Microsoft also released Open LLMs.<br /> <br /> So what just happened?<br /> <br /> Some research units have released open-source LLMs in the spirit of open, collaborative research into what is potentially one of the most impactful developments in history.<br /> <br /> But the larger commercial players are likely to have another motive: they have lost both technical leadership and first-mover advantage to OpenAI and are keen to re-establish their presence and reputation in order to develop an ecosystem of derivative applications, services and research communities around <I>their</I> LLMs.<br /> <br /> This is quite reminiscent of the open-source (Linux) adoption by, for example, IBM and Novell when they lost the operating system initiative; or perhaps even the release of Android by Google in their farsighted strategy to seek to capture a significant slice of the mobile app market from Apple.<br /> <br /> Motivations of these players aside, the net outcome of this evolution is that AI researchers are currently spoiled for choice in LLMs, something that was perhaps inconceivable but certainly unexpected in the first few months after the release of OpenAI’s flagship products in December 2022.<br /> <br /> <b>Some open LLMs of note</b><br /> <br /> There are many <I>leader boards</I> that rank the performance of open and selected proprietary LLMs on a wide range of tests and criteria; Hugging Face’s leaderboard is perhaps the best-known. The ‘gold standard’ currently typically remains the proprietary GPT-4.<br /> <br /> Researchers shopping around for an open-source LLM to adopt should be aware that, given that some of the tests are in the public domain, it is easy to train (that is, tweak) a candidate LLM to perform particularly well on a particular test; although even LLMs that weren’t trained by ‘cheating’ may improve their performance through accidental data leakage.<br /> <br /> Given that there are now literally hundreds of thousands of models listed on the Hugging Face repository, a short overview of some open LLMs is given to demonstrate the variety in models out there.<br /> <br /> Arguably the best known ‘open’ LLM is the Llama family, from Hugging Face. It is a family of models with <a href="https://llama.meta.com/llama3" target="_new">Llama3</a> available with a parameter size from 8B to 70B. The smaller models can run on single GPU and, with heavy quantisation, could even run on a smartphone or Raspberry Pi. The training code was publicly released using the GPL3 licence, and the model weights could be obtained by application, although they also leaked elsewhere on the internet.<br /> <br /> Since Llama was the first openly released LLM with performance that was anywhere near to GPT-3’s, it gained quick prominence. Since then, other researchers have released guides on how to train your own Llama ‘look-alike’ for as little as a few hundreds of dollars in cloud computing cost; the RedPajama project released the same or at least very similar open-source datasets; and an exciting project called llamafile allows one to download <I>one single file</I> which contains the entire model, the executable and necessary dependencies for different operating systems.<br /> <br /> Phi LLM was announced in the article “Textbooks Is All You Need” – a play of words on the original “Attention Is All You Need” pioneering article that introduced the transformer architecture with its novel attention mechanism.<br /> <br /> Bearing in mind that a huge amount of low quality internet data is often included in LLM training data, the approach taken by Phi is to train on higher quality datasets. The <a href="https://news.microsoft.com/source/features/ai/the-phi-3-small-language-models-with-big-potential/" target="_new">Phi models, released by Microsoft</a>, are trained on a ‘relatively small’ quality dataset which includes textbooks and high quality web-based materials, including code and text from top platforms.<br /> <br /> They are often touted as ‘small’ models although they still are trained on trillions of tokens and have from 1.5 to four billion parameters (although larger versions also exist). The quality of the training datasets is advanced as the reason why they often outperform larger LLMs on such tasks as reasoning or code generation.<br /> <br /> Not surprisingly, all other leading for-profit tech companies have also released LLMs with a degree of openness. For instance, Google released their Gemma family, which was developed using similar techniques and data as their proprietary Gemini models and is claimed to outperform much of the competition. Even Apple released OpenELM, as discussed above.<br /> <br /> Perhaps the most open-source of them all is <a href="https://allenai.org/olmo" target="_new">OLMo</a>, which is developed by the Allen Institute for AI, a genuinely non-profit AI research group. It is fully open-source including all training parameters, logs and code; the evaluation and fine-tuning code and metrics; and even the dataset.<br /> <br /> Another interesting open LLM release is <a href="https://mistral.ai/news/mixtral-of-experts/" target="_new">Mixtral (for example, Mixtral 8x22B)</a> which uses a Mixture of Experts architecture to blend multiple models into one. Bloom is a multilingual model covering some 46 natural languages. And there are a number of ‘domain specific models’ such as FinGPT for financial applications or Health-LLM for diagnostics and treatment optimisation.<br /> <br /> <b>Benefits of Open LLMs</b><br /> <br /> Many of the benefits of, and issues with, Open LLMs are similar to those of open-source software. A major benefit of open LLMs, especially for resource-poor researchers and innovators, is the reduced cost of access.<br /> <br /> In principle, open LLMs can be used free of charge, although even using these at scale for inference purposes requires access to significant hardware infrastructure; in fact some open LLMs are used as loss-leaders to sell cloud compute infrastructure (for example, Google and Amazon Web Services).<br /> <br /> Equally important is that access to LLM weights means that, as a researcher or developer, you have guaranteed future control and stability; if you use a proprietary LLM API – an API is the gateway for integrating LLMs into different apps – you have no guarantee as to the performance or stability of future models or API access. This is similar to the benefit of having access to the software code base in the case of open-source software.<br /> <br /> Another benefit of open LLMs is their inspectability and transparency: researchers can see how they were built, what their architecture is like, and what the specific settings for the hyper-parameters were. At the very least, a detailed description of the source of the training data is also given.<br /> <br /> This, in turn, enhances trust in these models. It also contributes towards another benefit: the replicability of building these LLMs. This leads into perhaps one of the bigger benefits for developing country researchers who are often far removed from social networks and tech hubs where LLM development skill sets diffuse: Open LLMs provide a learning scenario and democratise LLM development skills.<br /> <br /> This is crucial in a field where major advancements are made in the space of months, not years. Open LLMs also incentivise and are arguably easier to use for fine-tuning the base models to specific application domains; indeed, this is part of the motivation behind commercial players opening up their LLMs.<br /> <br /> Along with this, communities of like-minded developers, researchers and innovators tend to grow around the more popular or better open LLMs; which is exactly an explicit goal of the releasing organisations.<br /> <br /> <b>Issues with open LLMs</b><br /> <br /> As mentioned, many ‘openly released’ LLMs are actually not sufficiently open source to yield many of the touted benefits. Apart from model structure and weights, it is important to have access to training and evaluation settings, hyper-parameter choices and a very (ideally repeatable) description of data sources.<br /> <br /> Few open models have released the full training process pipeline and details, although there seems to have been more recent progress in this space, upping the ante for the remaining players.<br /> <br /> Another issue, very reminiscent of what happens in the open-source software world, is that of choice and forking: there is often too much choice and, within families, there are sometimes too many variants which may lead to confusion and decision inability. For instance, there are currently already more than 350,000 models available for download on Hugging Face’s repository!<br /> <br /> Although the lack of a single leader or provider may be seen as a drawback, there are some <I>de facto</I> leaders with deep pockets and good reputations. And, unlike the software world where a choice of corporate enterprise resource planning or database system will have a long term (10 or more years) impact, the state of the art in the LLM space moves much more rapidly, so future lock-in is less of an issue.<br /> <br /> Other potential issues revolve around the fact that many for-profit vendors only release ‘handicapped’ or weaker versions of their proprietary LLMs. Finally, security is sometimes mentioned as an issue, although your author sees this as more of an issue with proprietary LLMs following the ‘many eyeballs make bugs shallow’ line of reasoning.<br /> <br /> <b>What this means for researchers with limited digital resources and infrastructure</b><br /> <br /> So how can open-source LLMs – or even just the more limited ‘open’ LLMs – benefit academic researchers in developing countries and other contexts with severe resource-constraints?<br /> <br /> Obviously, having access to a free-to-use LLM as opposed to a proprietary LLM that charges for API access is attractive to finance-strapped researchers and students. The usage intensity is important: there are low-use ‘for-free’ tiers for most proprietary platforms.<br /> <br /> However, serious AI researchers are likely to need more intensive use, for example to train customised or domain-specific models, for development test kits or derivative tools, or they need downstream access to the full model, for example for purposes of fine-tuning.<br /> <br /> It must also be noted that bandwidth in some developing countries is still limited, and high-end graphics processing units (GPUs) are still an expensive commodity. On the other hand, most LLMs have a business model based on cloud computing, so cloud-based GPU resources are becoming cheaper (for example, Fast.ai).<br /> <br /> Another benefit is that resources for LLM fine-tuning, retrieval-augmented generation, prompt engineering, etcetera are all <I>also</I> becoming more democratised. Especially useful is the free availability of many online courses (for example, those offered by Deeplearning.AI) and workshops. These resources, and skill sets, are arguably much more useful to AI researchers in resource-constrained environments than those related to building full-scale base models.<br /> <br /> However, the biggest obstacles for researchers in developing country contexts remain their lack of dedicated research time, having to balance too many other commitments (teaching, administration, fund raising) as well as a lack of a local ecosystem or AI support-expert network. Thus many AI researchers remain relatively isolated from recent happenings, trends and opportunities in this space.<br /> <br /> Luckily for developing country researchers, there are already huge and growing opportunities in initiatives such as AI for Good, AI Ethics, Explainable AI (XAI) and AI Alignment; many of these come with relatively generous researcher support avenues also open to researchers from developing countries.<br /> <br /> Additionally, there should also be plenty of innovation opportunities in localised or context-sensitive AI solutions, for instance in using local data or addressing particular local needs or creating ‘AI for Development’ applications that address the United Nations Sustainable Development Goals.<br /> <br /> There is both space and an urgent need for researchers from a broad variety of backgrounds – that is, specifically not just software engineers and computer scientists – to engage with and help shape the future of AI and our society. Hopefully, this article has convinced some of you of the feasibility and necessity to engage with AI research, especially along the AI for Good and AI for Development lines.<br /> <br /> <a href="https://scholar.google.co.za/citations?user=lePUWT0AAAAJ&hl=en" target="_new"><I>Jean-Paul Van Belle</I></a><I> is professor in the </I><a href="https://sit.uct.ac.za/citanda" target="_new"><I>Centre for IT and National Development in Africa</I></a><I> and the </I><a href="https://sit.uct.ac.za/" target="_new"><I>department of information systems</I></a><I> at the </I><a href="https://www.uct.ac.za/" target="_new"><I>University of Cape Town</I></a><I>. He has published more than 300 peer-reviewed articles. Most of his historical research has been in emerging technologies and ICT-for-development but he has now fully pivoted to focussing on research on AI, with a focus on explainable AI and hybrid AI architectures.</I> </div><br><script src="https://www.universityworldnews.com/subscribe2.js"></script> <div style="background:#ffe57d;padding:8px" id="subform"> <span style="font-family:arial"><b>Receive email updates from UWN</b></span> <br> <input type="text" name="email" id="email" value="" style="width:65%;font-size:16px" placeholder="enter your email address">   <input type="submit" style="background:#000;color:#fff;padding:4px;font-size:20px; border:1px solid #000;font-weight:normal;border-radius:8px;font-family:arial" name="action" id="signupbutton" value="Sign Up" onclick="return formCheck();"> <br> <input type="checkbox" id="globalnl" name="globalnl" value="1"> Global newsletters    <input type="checkbox" id="africanl" name="africanl" value="1"> Africa newsletters    <input type="checkbox" id="otheremails" name="otheremails" value="1"> Other<br>     <i>(other includes related events and webinars)</i> <br><br> <div id="formresponsearea">Data will be processed according to our standard <a style="color:#777;font-size:13px" href="https://www.universityworldnews.com/index.php?page=Terms_Conditions" target="_blank">terms & conditions</a>.</div> </div> </div> <div></div> </td> <td style="width:261px;margin-left:10px;border-left:1px solid #999;padding-left:10px" valign="top"><div style="text-align:center;"><a href="https://ads.universityworldnews.com/bannerclick.php?id=abetsympoct2024"><img src="https://ads.universityworldnews.com/bannerview.php?id=abetsympoct2024" style="" alt="abetsympoct2024"></a></div><div style="text-align:center;"><a href="https://ads.universityworldnews.com/bannerclick.php?id=stelunijul2023"><img src="https://ads.universityworldnews.com/bannerview.php?id=stelunijul2023" style="width:250px;height:400px" alt="stelunijul2023"></a></div><div style="text-align:center;"><a href="https://ads.universityworldnews.com/bannerclick.php?id=nafsanov2024"><img src="https://ads.universityworldnews.com/bannerview.php?id=nafsanov2024" style="" alt="nafsanov2024"></a></div><div style="text-align:center;"><a href="https://ads.universityworldnews.com/bannerclick.php?id=uaeuoct2024"><img src="https://ads.universityworldnews.com/bannerview.php?id=uaeuoct2024" style="" alt="uaeuoct2024"></a></div><div style="text-align:center;"><a href="https://ads.universityworldnews.com/bannerclick.php?id="><img src="https://ads.universityworldnews.com/bannerview.php?id=" style="" alt=""></a></div><div style="text-align:center;"><a href="https://ads.universityworldnews.com/bannerclick.php?id="><img src="https://ads.universityworldnews.com/bannerview.php?id=" style="" alt=""></a></div><div style="text-align:center;"><a href="https://ads.universityworldnews.com/bannerclick.php?id="><img src="https://ads.universityworldnews.com/bannerview.php?id=" style="" alt=""></a></div><div style="text-align:center;"><a href="https://ads.universityworldnews.com/bannerclick.php?id=hetl"><img src="https://ads.universityworldnews.com/bannerview.php?id=hetl" style="width:250px;height:250px" alt="hetl"></a></div><div style="" class="botmar12 botmar8"> <table><tbody><tr> <td style="width:100%" class="" valign="top"> <div class="section-head-white-on-dark botmar8">Sponsored Article</div> <div class="article-brief"> <div class=""> <img src="https://ads.universityworldnews.com/bannerview.php?id=20241122075114204spon" style="width:250px;padding-bottom:12px" alt="PHOTO"> <div class="botmar4"><h1><a href="https://ads.universityworldnews.com/bannerclick.php?id=20241122075114204spon" style="color:#000;" class="">Unlocking potential – Diversity initiatives in STEM education</a></h1></div> <div class="byline-country botmar12">Dr Jeffrey Fergus and Dr Rochelle Williams, ABET</div> </div> </div> </td> </tr></tbody></table> </div> <div style="" class="botmar12 botmar8"> <table><tbody><tr> <td style="width:100%" class="" valign="top"> <div class="section-head-white-on-dark botmar8">Sponsored Article</div> <div class="article-brief"> <div class=""> <img src="https://ads.universityworldnews.com/bannerview.php?id=2024111309583869spon" style="width:250px;padding-bottom:12px" alt="PHOTO"> <div class="botmar4"><h1><a href="https://ads.universityworldnews.com/bannerclick.php?id=2024111309583869spon" style="color:#000;" class="">How research can support cyber and data security and privacy</a></h1></div> <div class="byline-country botmar12">IEEE Computer Society</div> </div> </div> </td> </tr></tbody></table> </div> <div style="" class="botmar12 botmar8"> <table><tbody><tr> <td style="width:100%" class="" valign="top"> <div class="section-head-white-on-dark botmar8">Sponsored Article</div> <div class="article-brief"> <div class=""> <img src="https://ads.universityworldnews.com/bannerview.php?id=20240716155250119spon" style="width:250px;padding-bottom:12px" alt="PHOTO"> <div class="botmar4"><h1><a href="https://ads.universityworldnews.com/bannerclick.php?id=20240716155250119spon" style="color:#000;" class="">Patent of catalyst for natural gas reforming technologies</a></h1></div> <div class="byline-country botmar12">UAE University Staff</div> </div> </div> </td> </tr></tbody></table> </div> <div style="" class="botmar12 botmar8"> <table><tbody><tr> <td style="width:100%" class="" valign="top"> <div class="section-head-white-on-dark botmar8">Sponsored Article</div> <div class="article-brief"> <div class=""> <img src="https://ads.universityworldnews.com/bannerview.php?id=20240625145758878spon" style="width:250px;padding-bottom:12px" alt="PHOTO"> <div class="botmar4"><h1><a href="https://ads.universityworldnews.com/bannerclick.php?id=20240625145758878spon" style="color:#000;" class="">Sustainable solutions to rising residential energy consumption</a></h1></div> <div class="byline-country botmar12">UAE University staff</div> </div> </div> </td> </tr></tbody></table> </div> </td> </tr></table> </div> <div id="footer" class="wdiv p8px" style="font-size:16px;line-height:24px;background:#000;color:#fff"> <a href="page.php?page=UW_Main">Global</a> | <a href="page.php?page=Africa_Edition">Africa</a> | <a href="page.php?page=Asia_Hub">Asia</a> | <a href="page.php?page=SDGs_Hub">SDGs</a> | <a href="page.php?page=Edtech_AI_Hub">Edtech/AI</a> | <a href="page.php?page=Commentary_Hub">Commentary</a> | <a href="topic-page.php?topic=World_Blog">World Blog</a> | <a href="special-report.php?publication=global">Special Reports</a> | <a href="page.php?page=subscribe">Subscribe</a> | <a href="page.php?page=Careers_at_UWN">Careers at UWN</a> <br> <a href="bookmarks.php">Bookmarks</a> | <a href="page.php?page=About_Us">About Us</a> | <a href="page.php?page=Advertising">Advertise</a> | <a href="page.php?page=sponsor-partner-profservices">Sponsorship</a> | <a href="page.php?page=Terms_Conditions">Terms & Conditions</a> | <a href="page.php?page=Contact_Us_Intro">Contact</a> | <a href="https://www.universityworldnews.com/fb/">Facebook</a> | <a href="version-force.php?ver=mobile">Mobile Site</a> </div> <div id="footer" class="wdiv p8px" style="font-size:12px;background:#000;color:#fff;border-top:1px solid #444">Copyright 2024 University World News</div> </div> </div> </div> <script> </script> <script> var bb = new BankersBox(1); function bookmark(sid, stitle) { url = 'https://www.universityworldnews.com/post.php?story='+sid; bb.sadd("bookmark", sid+'|'+stitle); alert('Bookmark Saved'); document.getElementById("bmi").style.opacity = "0.5"; logbookmarkclick(sid); } function logbookmarkclick(sid) { ga('send', 'event', 'Bookmark', 'Clicked', 'Bookmark/'+sid); return true; } function logwhatsappclick(sid) { ga('send', 'event', 'Whatsapp', 'Shared', 'Whatsapp/'+sid); return true; } </script> </body> <script> (function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){ (i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o), m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m) })(window,document,'script','https://www.google-analytics.com/analytics.js','ga'); ga('create', 'UA-20469064-1', 'auto'); ga('set', 'page', '/post.php?story=20240510092346341'); ga('send', 'pageview'); </script><script> window.addEventListener('load', () => { ready2go(); }); var bviewnotrackurl = 'https://ads.universityworldnews.com/bannerview2.php?id='; var bviewtrackurl = 'https://ads.universityworldnews.com/bannerview.php?id='; var bclicktrackurl = 'https://ads.universityworldnews.com/bannerclick.php?id='; var tracked = new Set(); var ads = new Array(); ads[1] = new Array('nafsanov2024ldr', 'uaeuoct2024ldr'); ads[2] = new Array('abetsympoct2024', 'stelunijul2023', 'uaeuoct2024'); ads[3] = new Array('stelunijul2023', 'uaeuoct2024', 'abetsympoct2024'); ads[4] = new Array('nafsanov2024'); ads[5] = new Array('nafsanov2024'); ads[6] = new Array('uaeuoct2024', 'abetsympoct2024', 'stelunijul2023'); var zids = new Array(); zids[1] = 'leaderboard-1'; zids[2] = 'rsb1-1'; zids[3] = 'rsb1-2'; zids[4] = 'rsb2-1'; zids[5] = 'rsb2-2'; zids[6] = 'rsb1-3'; var adnos = new Array(); adnos[1] = 0; adnos[2] = 0; adnos[3] = 0; adnos[4] = 0; adnos[5] = 0; adnos[6] = 0; var adcounts = new Array(); adcounts[1] = 1; adcounts[2] = 3; adcounts[3] = 3; adcounts[4] = 2; adcounts[5] = 2; adcounts[6] = 3; var periods = new Array(); periods[1] = 15; periods[2] = 15; periods[3] = 15; periods[4] = 15; periods[5] = 15; periods[6] = 15; var firstperiods = new Array(); firstperiods[1] = 15; firstperiods[2] = 10; firstperiods[3] = 10; firstperiods[4] = 15; firstperiods[5] = 15; firstperiods[6] = 10; var isfirst = new Array(); isfirst[1] = true; isfirst[2] = true; isfirst[3] = true; isfirst[4] = true; isfirst[5] = true; isfirst[6] = true; var rolledover = new Array(); rolledover[1] = false; rolledover[2] = false; rolledover[3] = false; rolledover[4] = false; rolledover[5] = false; rolledover[6] = false; function ready2go() { if (document.getElementById("zone-"+zids[1]+"-image") != null) { console.log('1'); trackedid = ads[1][0]; tracked.add(trackedid); setTimeout(myAdTimer, firstperiods[1] * 1000, 1); } if (document.getElementById("zone-"+zids[2]+"-image") != null) { console.log('2'); trackedid = ads[2][0]; tracked.add(trackedid); setTimeout(myAdTimer, firstperiods[2] * 1000, 2); } if (document.getElementById("zone-"+zids[3]+"-image") != null) { console.log('3'); trackedid = ads[3][0]; tracked.add(trackedid); setTimeout(myAdTimer, firstperiods[3] * 1000, 3); } if (document.getElementById("zone-"+zids[4]+"-image") != null) { console.log('4'); trackedid = ads[4][0]; tracked.add(trackedid); setTimeout(myAdTimer, firstperiods[4] * 1000, 4); } if (document.getElementById("zone-"+zids[5]+"-image") != null) { console.log('5'); trackedid = ads[5][0]; tracked.add(trackedid); setTimeout(myAdTimer, firstperiods[5] * 1000, 5); } if (document.getElementById("zone-"+zids[6]+"-image") != null) { console.log('6'); trackedid = ads[6][0]; tracked.add(trackedid); setTimeout(myAdTimer, firstperiods[6] * 1000, 6); } console.log(tracked); } function myAdTimer(i) { adnos[i]++; if (adnos[i] == adcounts[i]) { adnos[i] = 0; rolledover[i] = true; } adno = adnos[i]; elimage = document.getElementById("zone-"+zids[i]+"-image"); ellink = document.getElementById("zone-"+zids[i]+"-link"); adcode = ads[i][adno]; if (tracked.has(adcode)) newbvurl = bviewnotrackurl + adcode; else newbvurl = bviewtrackurl + adcode; newbcurl = bclicktrackurl + adcode; elimage.src = newbvurl; ellink.href = newbcurl; tracked.add(adcode); if (isfirst[i] == true) { isfirst[i] = false; setInterval(myAdTimer, periods[i] * 1000, i); } } </script> </html>