{"id":1790,"date":"2024-04-18T08:11:43","date_gmt":"2024-04-18T08:11:43","guid":{"rendered":"https:\/\/geneea.com\/news\/?p=1790"},"modified":"2026-01-27T20:25:31","modified_gmt":"2026-01-27T20:25:31","slug":"geneeas-ai-spotlight-10","status":"publish","type":"post","link":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10","title":{"rendered":"Geneea&#8217;s AI Spotlight #10"},"content":{"rendered":"\n<p id=\"ember6215\">The tenth edition of our newsletter on Large Language Models is here.<\/p>\n\n\n\n<p id=\"ember6216\">In this edition, we explore<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>how generative AI is used by companies and people,<\/li>\n\n\n\n<li>new models and infrastructure,<\/li>\n\n\n\n<li>experience from real deployment,<\/li>\n\n\n\n<li>challenges, misuse, and<\/li>\n\n\n\n<li>ColBERT, an efficient approach to search.<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ember6218\">The Business Side of Things<\/h2>\n\n\n\n<p id=\"ember6219\"><strong>GenAI in Fortune 500 companies<\/strong><\/p>\n\n\n\n<p id=\"ember6220\"><a href=\"https:\/\/www.linkedin.com\/in\/sarah-wang-59b96a7\/\">Sarah Wang<\/a> and <a href=\"https:\/\/www.linkedin.com\/in\/shangdaxu\/\">Shangda Xu<\/a> from Andreessen Horowitz, a VC fund, released a <a href=\"https:\/\/a16z.com\/generative-ai-enterprise-2024\/\">report<\/a> on how companies build and buy generative AI. The results are based on responses of ~100 executives from Fortune 500 companies.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Budgets<\/strong> for GenAI are <strong>growing fast<\/strong>, 2x\u20135x per respondent in 2024. And unlike in 2023, most of this will be in regular budget categories, not innovation. <strong>ROI<\/strong> is still <strong>unclear<\/strong> and hard to measure.<\/li>\n\n\n\n<li>Many companies lack the necessary <strong>technical experts<\/strong>. They are often paying the LLM API providers for support.<\/li>\n\n\n\n<li>In 2023, most companies experimented with one model (usually OpenAI). In 2024, more than 90% plan to use at least three models. This will allow them to choose the best model for the job and <strong>avoid vendor lock-in<\/strong>. Many applications are designed to make switching the LLM provider easy.<\/li>\n\n\n\n<li>In 2023, 80\u201390% of respondents used closed models. Now, many want to use open-source models instead. In many use cases and\/or with sufficient fine-tuning, an open-source model can match closed-source models.<\/li>\n\n\n\n<li><strong>Control<\/strong> (security of proprietary data and understanding the logic behind outputs) and <strong>customization<\/strong> are <strong>more important than cost<\/strong>.<\/li>\n\n\n\n<li>Customization via <strong>fine-tuning<\/strong> and <strong>RAG<\/strong> are far more common than pre-training.<\/li>\n\n\n\n<li>Companies use GenAI internally but are <strong>cautious about public-facing use cases<\/strong>.<\/li>\n<\/ul>\n\n\n\n<p id=\"ember6224\"><strong>Popular AI uses<\/strong><\/p>\n\n\n\n<p id=\"ember6225\"><a href=\"https:\/\/www.linkedin.com\/in\/marczaosanders\/\">Marc Zao-Sanders<\/a>, the founder of Filtered, <a href=\"https:\/\/learn.filtered.com\/thoughts\/ai-now-report\">explored GenAI&#8217;s top uses<\/a>. The summary was published in <a href=\"https:\/\/hbr.org\/2024\/03\/how-people-are-really-using-genai\">Harvard Business Review<\/a>. The winner is <strong>technical assistance<\/strong> (think of RAG-based chatbots over documentation, for example). <strong>Content creation and editing tasks<\/strong> follow closely.<strong> <\/strong>This is the area Geneea is most familiar with. Let us know in the comments what you (would) like to use GenAI for!<\/p>\n\n\n\n<p id=\"ember6227\"><strong>Automatic programmers<\/strong><\/p>\n\n\n\n<p id=\"ember6228\">Recently, the <a href=\"https:\/\/generativeai.pub\/the-first-ai-software-engineer-is-here-dfa0d562bace\">spotlight has been<\/a> on <strong>Devin<\/strong>, Cognition&#8217;s <strong>automatic software engineer<\/strong>. Their demos showcase its abilities to code tasks end-to-end and even search for documentation and bugs. But with access to Devin moving at a snail&#8217;s pace, <a href=\"https:\/\/github.com\/OpenDevin\/OpenDevin\">OpenDevin<\/a><strong> and Devika<\/strong> <a href=\"https:\/\/levelup.gitconnected.com\/meet-with-devika-is-this-the-end-of-coding-free-alternative-of-devin-53245152c642\">stepped in<\/a> as <strong>open-source alternatives<\/strong>. Obviously, some <a href=\"https:\/\/medium.com\/@machine-learning-made-simple\/did-the-makers-of-devin-ai-lie-about-their-capabilities-cdfa818d5fc2\">criticism appeared<\/a>, arguing Devin&#8217;s capabilities are exaggerated. <strong>Princeton<\/strong> released their own <a href=\"https:\/\/swe-agent.com\/\">SWE-agent<\/a>. It is able to fix 12.3% of issues in <a href=\"https:\/\/www.swebench.com\/\">SWE-bench<\/a>, Princeton\u2019s own bug-fixing benchmark. Devin can fix 1.5 percentage points more, but only on a subset of the benchmark. <strong>GitHub<\/strong> <a href=\"https:\/\/techcrunch.com\/2024\/03\/20\/githubs-latest-ai-tool-that-can-automatically-fix-code-vulnerabilities\/\">didn&#8217;t stay behind<\/a> either and added its own <strong>Autofix agent<\/strong> for fixing code vulnerabilities. <strong>Microsoft<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2403.08299.pdf\">introduced its AI coding agent<\/a> <strong>AutoDev<\/strong>, which orchestrates various agents with diverse tools to build, test, and version control projects in addition to coding.<\/p>\n\n\n\n<p id=\"ember6229\"><strong>In short<\/strong><\/p>\n\n\n\n<p id=\"ember6230\">Despite revealing the MM1 model, <strong>Apple<\/strong> is <a href=\"https:\/\/www.bloomberg.com\/news\/articles\/2024-03-18\/apple-in-talks-to-license-google-gemini-for-iphone-ios-18-generative-ai-tools\">in discussions with Google<\/a> to <strong>integrate Gemini<\/strong> into iPhones and <a href=\"https:\/\/www.wsj.com\/tech\/ai\/baidu-shares-rise-on-news-that-apple-will-use-its-ai-services-in-china-products-fede5a4f\">with Baidu in China<\/a>. <strong>Microsoft<\/strong> <a href=\"https:\/\/medium.com\/@ignacio.de.gregorio.noblejas\/the-first-big-ai-failure-just-took-place-about-time-0ef53fe0c941\">turned Inflection AI<\/a> into Microsoft&#8217;s AI Studio, <strong>continuing<\/strong> its <strong>AI investments<\/strong>. And <strong>Mistral&#8217;s<\/strong> future looks bright <strong>with<\/strong> the backing of a <a href=\"https:\/\/venturebeat.com\/ai\/snowflake-partners-with-mistral-ai-taking-its-open-llms-to-the-data-cloud\/\">new partner<\/a>, <strong>Snowflake<\/strong>.<\/p>\n\n\n\n<p id=\"ember6231\">At the <strong>GTC conference<\/strong>, Nvidia <a href=\"https:\/\/generativeai.pub\/nvidia-introduces-blackwell-a-40-000-gpu-with-208-billion-transistors-7d5d7648ca53\">announced the Blackwell<\/a> <strong>B100 AI GPU<\/strong>. It should have <strong>3x more transistors<\/strong> than the H100 and consume <strong>25x less power<\/strong>. But they better keep an eye on the University of Pennsylvania&#8217;s <a href=\"https:\/\/blog.seas.upenn.edu\/new-chip-opens-door-to-ai-computing-at-light-speed\/\">silicon-photonic chip<\/a>, which promises <strong>speed-of-light computation<\/strong> and minimal power usage. The chip has a variable height that scatters light in specific patterns, allowing it to perform vector-matrix multiplications.<\/p>\n\n\n\n<p id=\"ember6232\"><strong>Cloudflare<\/strong> is <a href=\"https:\/\/blog.cloudflare.com\/firewall-for-ai\/\">developing a novel firewall<\/a> designed to <strong>safeguard LLMs<\/strong> and prevent exploitation attempts such as prompt injection. It includes a layer for <strong>prompt and response validation<\/strong> alongside features for sensitive data detection and rate limiting.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ember6233\">Model Zoo<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>For a short time, <strong>Claude<\/strong> 3 Opus from Anthropic <strong>outperformed GPT-4<\/strong> in <a href=\"https:\/\/huggingface.co\/spaces\/lmsys\/chatbot-arena-leaderboard\">LMSYS Chatbot Arena<\/a>. From our experience, its <strong>answers<\/strong> are better readable, pleasant, and <strong>more human-like<\/strong>. OpenAI released a <strong>new GPT-4 version<\/strong> to reclaim the first spot, and <a href=\"https:\/\/arstechnica.com\/information-technology\/2024\/03\/openais-gpt-5-may-launch-this-summer-upgrading-chatgpt-along-the-way\/\">rumors are circulating<\/a> about the GPT-4 successor, expected in mid-2024. Anyway, Anthropic seems like a <a href=\"https:\/\/www.aboutamazon.com\/news\/company-news\/amazon-anthropic-ai-investment\">good investment<\/a> for Amazon.<\/li>\n\n\n\n<li><a href=\"https:\/\/txt.cohere.com\/command-r\/\">Cohere released<\/a> <strong>Command-R<\/strong>, securing a great 7th place in Chatbot Arena. It focuses on <strong>enterprise use<\/strong> (low latency, high throughput) and <strong>Retrieval Augmented Generation <\/strong>(RAG).<\/li>\n\n\n\n<li><strong>01.AI<\/strong> <a href=\"https:\/\/www.marktechpost.com\/2024\/03\/13\/01-ai-introduces-the-yi-model-family-a-series-of-language-and-multimodal-models-that-demonstrate-strong-multi-dimensional-capabilities\/\">released the open foundation<\/a> <strong>model family Yi<\/strong>, focusing on multimodality, emphasizing <strong>visual comprehension<\/strong>. Now, 25th in Chatbot Arena.<\/li>\n\n\n\n<li><strong>273 Ventures<\/strong> <a href=\"https:\/\/venturebeat.com\/ai\/the-first-fairly-trained-ai-large-language-model-is-here\/\">released a very small<\/a> but indisputably <strong>legal Kelvin<\/strong> Legal Large Language Model (KL3M) for the <strong>legal industry<\/strong>. They got a <a href=\"https:\/\/www.fairlytrained.org\/blog\/fairly-trained-launches-certification-for-generative-ai-models-that-respect-creators-rights\">fairly trained certification label<\/a> for training only on public domain data, mostly government and legal documents.<\/li>\n\n\n\n<li><a href=\"https:\/\/generativeai.pub\/dbrx-the-game-changer-in-large-language-models-1ad0f5769ae2\">Databricks released<\/a> <strong>DBRX<\/strong>, a strong <strong>open-source<\/strong> model with <strong>132 billion<\/strong> parameters and a Mixture of Experts (MoE) architecture, excelling particularly in <strong>coding tasks<\/strong>.<\/li>\n\n\n\n<li>After Elon <a href=\"https:\/\/www.theatlantic.com\/technology\/archive\/2024\/03\/xai-grok-open-source-ai\/677795\/\">Musk accused OpenAI<\/a> of being ClosedAI, xAI release <a href=\"https:\/\/generativeai.pub\/xai-releases-grok-1-the-biggest-open-source-llm-28fe8ab84575\">Grok-1<\/a> (see also <a href=\"https:\/\/pub.towardsai.net\/opensource-grok-1-a-new-frontier-in-ai-by-xai-2567f01700f1\">here<\/a>), \u201cthe biggest LLM\u201d, with <strong>314 billion<\/strong> parameters. <strong>Grok<\/strong> incorporates MoE with two active experts, making it an 86B model at inference with <strong>massive<\/strong> memory demands <strong>even<\/strong> when <strong>quantized<\/strong>.<\/li>\n\n\n\n<li>Yet, quantization techniques are advancing. Microsoft&#8217;s researchers have <a href=\"https:\/\/arxiv.org\/abs\/2402.17764\">introduced ternary quantization<\/a>, employing <strong>solely -1\/0\/1 values<\/strong>. This significantly lowers cost while preserving the performance of the 16-bit model LLaMA 3B.<\/li>\n\n\n\n<li><strong>Apple&#8217;s<\/strong> researchers <a href=\"https:\/\/arxiv.org\/pdf\/2403.09611.pdf\">developed a multimodal model family<\/a> <strong>MM1<\/strong>, varying in size and architecture. They were especially careful with composing their pre-training data, comprising image-text docs and image-caption pairs, and even studied the impact of image resolution. This helped them outperform the top models, such as Emu2 and Flamingo.<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ember6235\">Practical Advice<\/h2>\n\n\n\n<p id=\"ember6236\">Check out this excellent post by <a href=\"https:\/\/www.linkedin.com\/in\/ken-kantzer-222b7a20\/\">Ken Kantzer<\/a>, CTO at Truss, who summarizes their experiences with LLMs after processing 500 million tokens. Their conclusions are very similar to ours. Few highlights:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Short prompts<\/strong> are often better than detailed instructions.<\/li>\n\n\n\n<li>Typically, <strong>LangChain<\/strong> is <strong>overkill<\/strong>.<\/li>\n\n\n\n<li>GPT is really <strong>bad at saying nothing<\/strong>. It prefers to hallucinate instead of acknowledging the absence of requested information.<\/li>\n\n\n\n<li>Despite ever-increasing input contexts, the <strong>output<\/strong> length is still quite <strong>limited<\/strong>. They encounter problems when the output should contain ten or more items.<\/li>\n\n\n\n<li><strong>RAG<\/strong> (Retrieval-Augmented Generation) is <strong>hard<\/strong>. There are no ideal similarity thresholds. Semantic searches are confusing for users. The old-fashioned faceted search might be better in many scenarios.<\/li>\n\n\n\n<li>Hallucinations are rare in information extraction. (With the exception of making up results when there are none mentioned above.)<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ember6239\">Challenges<\/h2>\n\n\n\n<p id=\"ember6240\"><strong>Pledge to mitigate GenAI misuse<\/strong><\/p>\n\n\n\n<p id=\"ember6241\">At the <strong>Munich Security Conference,<\/strong> the big tech companies <a href=\"https:\/\/www.ft.com\/content\/8dcbc162-a0f5-47ce-a2ea-5fb25cb160c5\">pledged to mitigate<\/a> the misuse of GenAI as its <strong>use for misinformation<\/strong> is a big <strong>challenge<\/strong>. This can be tackled with several strategies such as <strong>red teaming<\/strong> (humans finding weaknesses), <strong>safety guardrails<\/strong> (designing protocols to steer away from harmful outcomes), <strong>labeling<\/strong> the creations with <a href=\"https:\/\/www.theverge.com\/2024\/2\/6\/24063954\/ai-watermarks-dalle3-openai-content-credentials\">technical standards like C2PA<\/a>, as OpenAI did with DALL-E, or <strong>detecting<\/strong> them, as Meta <a href=\"https:\/\/www.reuters.com\/technology\/meta-start-labeling-ai-generated-images-companies-like-openai-google-2024-02-06\/\">wants to do<\/a> on their platforms with images.<\/p>\n\n\n\n<p id=\"ember6242\"><strong>Detecting AI-generated texts<\/strong><\/p>\n\n\n\n<p id=\"ember6243\">Detecting AI-generated <strong>text is trickier<\/strong>, especially if we want to <strong>avoid mislabeling human texts as AI-generated<\/strong>. Fortunately, researchers from the University of Maryland and Carnegie Mellon University devised a <a href=\"https:\/\/arxiv.org\/pdf\/2401.12070.pdf\">method with much fewer<\/a> false-positive results than others, <a href=\"https:\/\/www.theguardian.com\/technology\/2023\/jul\/10\/programs-to-detect-ai-discriminate-against-non-native-english-speakers-shows-study\">even for non-native speakers<\/a>, which is a big challenge. It <strong>detects over 90%<\/strong> of GPT-generated text, and it <strong>does not need modifications <\/strong>for different language models. Usually, this classification relies on <strong>perplexity<\/strong>, a measure of how well a language model predicts the next word, which is reliably different for AI and human-generated text. The researchers add another measure, <strong>cross-perplexity<\/strong>, that measures how surprising the predictions of one language model are to another model.<\/p>\n\n\n\n<p id=\"ember6244\"><strong>Faithful citations<\/strong><\/p>\n\n\n\n<p id=\"ember6245\">Another credibility challenge is generating <strong>faithful citations to support the claims<\/strong> of generated texts. Researchers from the University of Singapore, Washington, and Nanyang Technological University <a href=\"https:\/\/arxiv.org\/pdf\/2402.04315.pdf\">trained models for this purpose<\/a> using fine-grained rewards. They achieved around <strong>60% precision and recall<\/strong> on the ELI5 dataset, 10% higher than ChatGPT&#8217;s performance. Even though GPT-4 would probably do a bit better, out-of-the-box usability <strong>remains a challenge<\/strong>, necessitating <strong>frameworks like RAG<\/strong>.<\/p>\n\n\n\n<p id=\"ember6246\"><strong>No butterfly effect<\/strong><\/p>\n\n\n\n<p id=\"ember6247\">Contrary to the title of their paper, <a href=\"https:\/\/arxiv.org\/pdf\/2401.03729.pdf\">Butterfly Effect of Altering Prompts<\/a>, researchers from the University of Southern California showed that the <strong>performance of larger models<\/strong> is <strong>relatively unaffected by minor, irrelevant additions <\/strong>to prompts such as \u201chello,\u201d \u201cthank you,\u201d or offering a tip. However, prompting for a <strong>specific format<\/strong>, such as JSON or CSV, <strong>may lower<\/strong> the performance <strong>differently for specific models<\/strong>, depending on their training data. Not surprisingly, <strong>jailbreak<\/strong> prompts <strong>hurt<\/strong> the performance <strong>a lot<\/strong>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ember6248\">ColBERT: Efficient and Effective Passage Search<\/h2>\n\n\n\n<p id=\"ember6249\">Four years ago, <a href=\"https:\/\/www.linkedin.com\/in\/omar-k-09747b188\/\">Omar K.<\/a> and <a href=\"https:\/\/www.linkedin.com\/in\/mateizaharia\/\">Matei Zaharia<\/a> from Stanford published <a href=\"https:\/\/arxiv.org\/abs\/2004.12832\">a great article describing ColBERT<\/a>, a retrieval model that attempts to strike a middle ground between fast yet less accurate search approaches (keywords, embedding similarity) and slow but more accurate approaches that use LLMs.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>In ColBERT, documents and queries are represented by matrices with contextualized token embeddings (computed by BERT passed through a linear layer, which reduces dimension).<\/li>\n\n\n\n<li>The score is computed as a sum of MaxSim through all query tokens. MaxSim for a query token is the maximum similarity through all document tokens.<\/li>\n\n\n\n<li>BERT is fine-tuned, and linear layers are trained using triples (query, positive document, negative document).<\/li>\n\n\n\n<li>ColBERT can be used either for re-ranking pre-selected results or for full retrieval (optimized using search indexes like Faiss). For re-ranking, ColBERT is competitive with BERT-based approaches in quality but achieves more than 100x shorter latency. For full retrieval, ColBERT is ~5x slower than traditional approaches but achieves significantly higher retrieval quality.<\/li>\n\n\n\n<li>Research continued with <a href=\"https:\/\/arxiv.org\/abs\/2112.01488\">ColBERTv2<\/a> and made its way into a RAG implementation called <a href=\"https:\/\/github.com\/bclavie\/ragatouille\">Ragatouille<\/a>.<\/li>\n<\/ul>\n\n\n\n<p>Please <a href=\"https:\/\/www.linkedin.com\/pulse\/geneeas-ai-spotlight-10-geneea-bp99e\/?trackingId=n4c9HE5xoBqOvxFvQsPv%2Bw%3D%3D\">subscribe<\/a> and stay tuned for the next issue of Geneea\u2019s AI Spotlight newsletter!<\/p>\n","protected":false},"excerpt":{"rendered":"<p>The tenth edition of our newsletter on Large Language Models is here.<\/p>\n<p>In this edition, we explore<\/p>\n<p>\u2022 how generative AI is used by companies and people,<br \/>\n\u2022 new models and infrastructure,<br \/>\n\u2022 experience from real deployment,<br \/>\n\u2022 challenges, misuse, and<br \/>\n\u2022 ColBERT, an efficient approach to search.<\/p>\n","protected":false},"author":15,"featured_media":1791,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[378,374],"tags":[244,240,242],"class_list":["post-1790","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-large-language-models","category-newsletter","tag-ai","tag-generativeai","tag-newsletter"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.3 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>Geneea&#039;s AI Spotlight #10 - Geneea News<\/title>\n<meta name=\"description\" content=\"LLM newsletter #10: GenAI use cases; new models and infrastructure; real deployment experiences \u2013 challenges and misuse; and a new search model ColBERT.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Geneea&#039;s AI Spotlight #10 - Geneea News\" \/>\n<meta property=\"og:description\" content=\"LLM newsletter #10: GenAI use cases; new models and infrastructure; real deployment experiences \u2013 challenges and misuse; and a new search model ColBERT.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10\" \/>\n<meta property=\"og:site_name\" content=\"Geneea News\" \/>\n<meta property=\"article:published_time\" content=\"2024-04-18T08:11:43+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2026-01-27T20:25:31+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/geneea.com\/news\/wp-content\/uploads\/2024\/04\/Newsletter-10-robot-na-web-1024x577.png\" \/>\n\t<meta property=\"og:image:width\" content=\"1024\" \/>\n\t<meta property=\"og:image:height\" content=\"577\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"Marcela Soukupova\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Marcela Soukupova\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"8 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10\"},\"author\":{\"name\":\"Marcela Soukupova\",\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/#\\\/schema\\\/person\\\/69c8751a4c026723f4bac2e892f52cd8\"},\"headline\":\"Geneea&#8217;s AI Spotlight #10\",\"datePublished\":\"2024-04-18T08:11:43+00:00\",\"dateModified\":\"2026-01-27T20:25:31+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10\"},\"wordCount\":1558,\"publisher\":{\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/geneea.com\\\/news\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/Newsletter-10-robot-na-web.png\",\"keywords\":[\"AI\",\"generativeAI\",\"newsletter\"],\"articleSection\":[\"Large language models\",\"Newsletter\"],\"inLanguage\":\"en-US\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10\",\"url\":\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10\",\"name\":\"Geneea's AI Spotlight #10 - Geneea News\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/geneea.com\\\/news\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/Newsletter-10-robot-na-web.png\",\"datePublished\":\"2024-04-18T08:11:43+00:00\",\"dateModified\":\"2026-01-27T20:25:31+00:00\",\"description\":\"LLM newsletter #10: GenAI use cases; new models and infrastructure; real deployment experiences \u2013 challenges and misuse; and a new search model ColBERT.\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10#primaryimage\",\"url\":\"https:\\\/\\\/geneea.com\\\/news\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/Newsletter-10-robot-na-web.png\",\"contentUrl\":\"https:\\\/\\\/geneea.com\\\/news\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/Newsletter-10-robot-na-web.png\",\"width\":2561,\"height\":1442},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/geneeas-ai-spotlight-10#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/geneea.com\\\/news\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Geneea&#8217;s AI Spotlight #10\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/#website\",\"url\":\"https:\\\/\\\/geneea.com\\\/news\\\/\",\"name\":\"Geneea News\",\"description\":\"Learn more about what&#039;s happening at Geneea: new NLP features, newest case studies, tutoring projects, conferences we attended, etc.\",\"publisher\":{\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/geneea.com\\\/news\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/#organization\",\"name\":\"Geneea News\",\"url\":\"https:\\\/\\\/geneea.com\\\/news\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/geneea.com\\\/news\\\/wp-content\\\/uploads\\\/2022\\\/02\\\/cropped-geneea-logo-50pc.png\",\"contentUrl\":\"https:\\\/\\\/geneea.com\\\/news\\\/wp-content\\\/uploads\\\/2022\\\/02\\\/cropped-geneea-logo-50pc.png\",\"width\":242,\"height\":64,\"caption\":\"Geneea News\"},\"image\":{\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/#\\\/schema\\\/logo\\\/image\\\/\"}},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/geneea.com\\\/news\\\/#\\\/schema\\\/person\\\/69c8751a4c026723f4bac2e892f52cd8\",\"name\":\"Marcela Soukupova\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/44f35824640c6a5b31bfef2f478d704874dc3d81bfad511c158ab12274072e16?s=96&d=mm&r=g\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/44f35824640c6a5b31bfef2f478d704874dc3d81bfad511c158ab12274072e16?s=96&d=mm&r=g\",\"contentUrl\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/44f35824640c6a5b31bfef2f478d704874dc3d81bfad511c158ab12274072e16?s=96&d=mm&r=g\",\"caption\":\"Marcela Soukupova\"},\"sameAs\":[\"http:\\\/\\\/Marcela%20Soukupova\"],\"url\":\"https:\\\/\\\/geneea.com\\\/news\\\/author\\\/marcela-soukupova\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Geneea's AI Spotlight #10 - Geneea News","description":"LLM newsletter #10: GenAI use cases; new models and infrastructure; real deployment experiences \u2013 challenges and misuse; and a new search model ColBERT.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10","og_locale":"en_US","og_type":"article","og_title":"Geneea's AI Spotlight #10 - Geneea News","og_description":"LLM newsletter #10: GenAI use cases; new models and infrastructure; real deployment experiences \u2013 challenges and misuse; and a new search model ColBERT.","og_url":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10","og_site_name":"Geneea News","article_published_time":"2024-04-18T08:11:43+00:00","article_modified_time":"2026-01-27T20:25:31+00:00","og_image":[{"width":1024,"height":577,"url":"https:\/\/geneea.com\/news\/wp-content\/uploads\/2024\/04\/Newsletter-10-robot-na-web-1024x577.png","type":"image\/png"}],"author":"Marcela Soukupova","twitter_card":"summary_large_image","twitter_misc":{"Written by":"Marcela Soukupova","Est. reading time":"8 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10#article","isPartOf":{"@id":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10"},"author":{"name":"Marcela Soukupova","@id":"https:\/\/geneea.com\/news\/#\/schema\/person\/69c8751a4c026723f4bac2e892f52cd8"},"headline":"Geneea&#8217;s AI Spotlight #10","datePublished":"2024-04-18T08:11:43+00:00","dateModified":"2026-01-27T20:25:31+00:00","mainEntityOfPage":{"@id":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10"},"wordCount":1558,"publisher":{"@id":"https:\/\/geneea.com\/news\/#organization"},"image":{"@id":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10#primaryimage"},"thumbnailUrl":"https:\/\/geneea.com\/news\/wp-content\/uploads\/2024\/04\/Newsletter-10-robot-na-web.png","keywords":["AI","generativeAI","newsletter"],"articleSection":["Large language models","Newsletter"],"inLanguage":"en-US"},{"@type":"WebPage","@id":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10","url":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10","name":"Geneea's AI Spotlight #10 - Geneea News","isPartOf":{"@id":"https:\/\/geneea.com\/news\/#website"},"primaryImageOfPage":{"@id":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10#primaryimage"},"image":{"@id":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10#primaryimage"},"thumbnailUrl":"https:\/\/geneea.com\/news\/wp-content\/uploads\/2024\/04\/Newsletter-10-robot-na-web.png","datePublished":"2024-04-18T08:11:43+00:00","dateModified":"2026-01-27T20:25:31+00:00","description":"LLM newsletter #10: GenAI use cases; new models and infrastructure; real deployment experiences \u2013 challenges and misuse; and a new search model ColBERT.","breadcrumb":{"@id":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10#primaryimage","url":"https:\/\/geneea.com\/news\/wp-content\/uploads\/2024\/04\/Newsletter-10-robot-na-web.png","contentUrl":"https:\/\/geneea.com\/news\/wp-content\/uploads\/2024\/04\/Newsletter-10-robot-na-web.png","width":2561,"height":1442},{"@type":"BreadcrumbList","@id":"https:\/\/geneea.com\/news\/geneeas-ai-spotlight-10#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/geneea.com\/news"},{"@type":"ListItem","position":2,"name":"Geneea&#8217;s AI Spotlight #10"}]},{"@type":"WebSite","@id":"https:\/\/geneea.com\/news\/#website","url":"https:\/\/geneea.com\/news\/","name":"Geneea News","description":"Learn more about what&#039;s happening at Geneea: new NLP features, newest case studies, tutoring projects, conferences we attended, etc.","publisher":{"@id":"https:\/\/geneea.com\/news\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/geneea.com\/news\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Organization","@id":"https:\/\/geneea.com\/news\/#organization","name":"Geneea News","url":"https:\/\/geneea.com\/news\/","logo":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/geneea.com\/news\/#\/schema\/logo\/image\/","url":"https:\/\/geneea.com\/news\/wp-content\/uploads\/2022\/02\/cropped-geneea-logo-50pc.png","contentUrl":"https:\/\/geneea.com\/news\/wp-content\/uploads\/2022\/02\/cropped-geneea-logo-50pc.png","width":242,"height":64,"caption":"Geneea News"},"image":{"@id":"https:\/\/geneea.com\/news\/#\/schema\/logo\/image\/"}},{"@type":"Person","@id":"https:\/\/geneea.com\/news\/#\/schema\/person\/69c8751a4c026723f4bac2e892f52cd8","name":"Marcela Soukupova","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/secure.gravatar.com\/avatar\/44f35824640c6a5b31bfef2f478d704874dc3d81bfad511c158ab12274072e16?s=96&d=mm&r=g","url":"https:\/\/secure.gravatar.com\/avatar\/44f35824640c6a5b31bfef2f478d704874dc3d81bfad511c158ab12274072e16?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/44f35824640c6a5b31bfef2f478d704874dc3d81bfad511c158ab12274072e16?s=96&d=mm&r=g","caption":"Marcela Soukupova"},"sameAs":["http:\/\/Marcela%20Soukupova"],"url":"https:\/\/geneea.com\/news\/author\/marcela-soukupova"}]}},"_links":{"self":[{"href":"https:\/\/geneea.com\/news\/wp-json\/wp\/v2\/posts\/1790","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/geneea.com\/news\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/geneea.com\/news\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/geneea.com\/news\/wp-json\/wp\/v2\/users\/15"}],"replies":[{"embeddable":true,"href":"https:\/\/geneea.com\/news\/wp-json\/wp\/v2\/comments?post=1790"}],"version-history":[{"count":2,"href":"https:\/\/geneea.com\/news\/wp-json\/wp\/v2\/posts\/1790\/revisions"}],"predecessor-version":[{"id":1795,"href":"https:\/\/geneea.com\/news\/wp-json\/wp\/v2\/posts\/1790\/revisions\/1795"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/geneea.com\/news\/wp-json\/wp\/v2\/media\/1791"}],"wp:attachment":[{"href":"https:\/\/geneea.com\/news\/wp-json\/wp\/v2\/media?parent=1790"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/geneea.com\/news\/wp-json\/wp\/v2\/categories?post=1790"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/geneea.com\/news\/wp-json\/wp\/v2\/tags?post=1790"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}