<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by LightOn on Medium]]></title>
        <description><![CDATA[Stories by LightOn on Medium]]></description>
        <link>https://medium.com/@LightOnIO?source=rss-1a8df9096570------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*Zon85rbjeupfZta-tqQhGw.png</url>
            <title>Stories by LightOn on Medium</title>
            <link>https://medium.com/@LightOnIO?source=rss-1a8df9096570------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 14:41:39 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@LightOnIO/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[A Paradigm Shift: Reasoning at Enteprise Scale]]></title>
            <link>https://medium.com/@LightOnIO/a-paradigm-shift-reasoning-at-enteprise-scale-0b8ab45d61a7?source=rss-1a8df9096570------2</link>
            <guid isPermaLink="false">https://medium.com/p/0b8ab45d61a7</guid>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[information-retrieval]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[LightOn]]></dc:creator>
            <pubDate>Wed, 01 Oct 2025 02:39:32 GMT</pubDate>
            <atom:updated>2025-10-01T02:39:32.433Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xAEbguMRy2mQuBKOiOIY2Q.png" /></figure><p>When performing retrieval at scale on large sets of enteprise documents, it becomes very clear that current Retrieval Augmented Generation (RAG)-like approaches are not well suited (irrespective to the context windows becoming very large). The “RAG is dead” meme comes out every so often yet willfully ignores that</p><ul><li>most interesting sets of documents are always beyond the latest largest context window that the cool kids talk about</li><li>the reason we want a satisfying RAG is that <em>we do not want to choose</em> the documents that will come into the context window</li><li>the current story is about text, get ready for images, voice and videos</li><li>large context windows do not assure a level of recall quality</li></ul><p>If company documents are the context needed to have a purposeful discussion with LLMs inside a company or if new services or products are built on internal documents, then we need to have new algorithms for an enriched experience with all the company knowledge.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/0*DPL8aspDUKfVs-5m.png" /></figure><p>At <a href="https://www.lighton.ai/">LightOn</a>, we believe the future of AI retrieval lies in reasoning, not just pattern matching. As Antoine Chaffin explained in his<a href="https://maven.com/p/1973fe/going-further-late-interaction-beats-single-vector-limits"> Maven podcast appearance</a>, single-vector embeddings collapse nuance into one dimension, limiting systems to shallow similarity. (Before you read the rest of the blog post, do not hesitate to get in touch if you want to help in building this new stack)</p><p>Late-interaction models take a different approach:</p><ul><li>Every token is preserved as its own vector.</li><li>Matching happens <em>late</em>, at the interaction stage.</li><li>The result: deeper semantic understanding and genuine reasoning.</li></ul><p>This simple but powerful insight has sparked an open-source ecosystem that’s now shaping both academic research and production-scale AI systems.</p><h3>PyLate: From Experimental Code to Peer-Reviewed Paper</h3><p><a href="https://lightonai.github.io/pylate/">PyLate</a> began as an internal experiment to simplify multi-vector training. Today, it’s a full-fledged library with 527 GitHub stars and growing adoption.</p><ul><li>Academic recognition: PyLate’s paper was accepted at CIKM 2025 (see below), becoming the first peer-reviewed library dedicated to training ColBERT-style models.</li><li>Practical impact: Researchers can train a state-of-art retrieval model on MS MARCO in under 2 hours with just ~80 lines of code.</li><li>Real-world benefit: Out-of-domain search, reasoning-heavy tasks, and long-context retrieval become accessible to any team.</li></ul><p>if you want to learn more about the library:<a href="https://lightonai.github.io/pylate/"> PyLate documentation</a></p><h3><a href="https://nuit-blanche.blogspot.com/2024/12/modernbert-smarter-better-faster-and.html">ModernBERT: Re-Imagining the Encoder</a></h3><p>In partnership with <a href="https://answer.ai/">Answer.AI</a>, <a href="https://www.lighton.ai/">LightOn</a> co-developed <a href="https://github.com/AnswerDotAI/ModernBERT">ModernBERT</a>, a model that fundamentally rethinks encoder architecture.</p><ul><li>8192-token context with Flash Attention, running efficiently on consumer GPUs.</li><li>1,500 GitHub stars and 27M+ downloads on HuggingFace.</li><li>Poster presentation at ACL 2025 (Vienna): validation from one of NLP’s most competitive venues.</li></ul><p>ModernBERT has already been <a href="https://scholar.google.com/scholar?cites=2135362229494833314&amp;as_sdt=2005&amp;sciodt=0,5&amp;hl=fr">cited 305+ times</a>, with variants like <em>BioClinical ModernBERT</em> emerging for healthcare applications.</p><p>👉 Explore:<a href="https://www.lighton.ai/lighton-blogs/finally-a-replacement-for-bert"> ModernBERT LightOn blog post</a></p><h3>FastPlaid: Performance That Scales</h3><p>Building great models is only half the challenge, making them work in production is the other. That’s where<a href="https://github.com/lightonai/fast-plaid"> FastPlaid</a> comes in.</p><ul><li>A Rust + CUDA engine for multi-vector search.</li><li>Delivers +554% throughput improvements for multi-vector search compared to Stanford’s PLAID baseline.</li><li>Designed for scalability: powering recommendation engines, retrieval-augmented generation (RAG), and real-time search.</li></ul><p>As Raphael Sourty explains, static indexes solve many use cases, but mutable indexes (new in v1.10.0) unlock real-world applications where data evolves continuously.</p><p>👉 Read more: <a href="https://www.lighton.ai/lighton-blogs/fastplaid">FastPlaid LightOn blogpost</a></p><h3>PyLate-rs: Retrieval in the Browser</h3><p>Finally, to push accessibility even further, PyLate-rs compiles late-interaction inference to WebAssembly (WASM).</p><p>That means:</p><ul><li>Run a state-of-the-art retriever directly in the browser.</li><li>Achieve 97% faster cold-start performance on CPU.</li><li>Remove server dependencies entirely.</li></ul><p>This lowers the barrier for demos, education, and lightweight deployments, proving late-interaction isn’t just powerful, it’s portable.</p><h3>From Theory to Production: A Movement</h3><p>Taken together, these projects form a technical symphony:</p><ul><li><a href="https://www.lighton.ai/lighton-blogs/finally-a-replacement-for-bert">ModernBERT</a> provides the backbone.</li><li><a href="https://github.com/lightonai/pylate">PyLate</a> enables fast and easy training of SOTA models.</li><li><a href="https://www.lighton.ai/lighton-blogs/fastplaid">FastPlaid</a> ensures scalable search performance.</li><li><a href="https://www.lighton.ai/lighton-blogs/pylate-rs">PyLate-rs</a> brings inference to any environment.</li></ul><p>The ecosystem has grown from an academic curiosity into a reasoning-first retrieval stack. With recognition at CIKM and ACL, adoption across GitHub and HuggingFace, and practical tools for real-world workflows, LightOn is helping shape the next era of AI search.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/0*qChhbLNfp-mrfQKG.png" /></figure><p>📖 Explore <a href="https://www.lighton.ai/">LightOn</a>’s open-source ecosystem:</p><ul><li><a href="https://lightonai.github.io/pylate/">PyLate</a></li><li><a href="https://www.lighton.ai/lighton-blogs/finally-a-replacement-for-bert">ModernBERT</a></li><li><a href="https://github.com/lightonai/fast-plaid">FastPlaid</a></li><li>PyLate has already enabled the development of state-of-the-art models, such as:</li><li><a href="https://www.lighton.ai/lighton-blogs/lighton-releases-gte-moderncolbert-first-state-of-the-art-late-interaction-model-trained-on-pylate">GTE-ModernColBERT</a> and</li><li><a href="https://www.lighton.ai/lighton-blogs/lighton-releases-reason-colbert">Reason-ModernColBERT</a></li><li>Other models:</li><li><a href="https://www.lighton.ai/lighton-blogs/introducing-ettin-suite-the-sota-open-recipe-to-outperform-existing-generative-retrieval-models">Ettin Suite: SoTA Paired Encoders and Decoders</a></li><li><a href="https://www.lighton.ai/lighton-blogs/announcing-bioclinical-modernbert-a-new-sota-encoder-model-for-medical-nlp">BioClinical ModernBERT: a new SOTA encoder model for Medical NLP</a></li><li><a href="https://www.lighton.ai/lighton-blogs/monoqwen-vision">MonoQwen-Vision, the first visual document reranker</a></li></ul><p>Dataset</p><ul><li><a href="https://www.lighton.ai/lighton-blogs/fc-amf-ocr-dataset">FC-AMF-OCR Dataset : a 9.3 million images OCR dataset to improve real world document parsing</a></li></ul><p>Pre-training libraries</p><ul><li><a href="https://www.lighton.ai/lighton-blogs/passing-the-torch-training-a-mamba-model-for-smooth-handover">Passing the Torch: Training a Mamba Model for Smooth Handover</a></li><li><a href="https://www.lighton.ai/lighton-blogs/training-mamba-models-on-amd-mi250-mi250x-gpus-with-custom-kernels">Training Mamba Models on AMD MI250/MI250X GPUs with Custom Kernels</a></li></ul><p>The preprint on PyLate:</p><p><a href="https://arxiv.org/abs/2508.03555">PyLate: Flexible Training and Retrieval for Late Interaction Models</a> by <a href="https://arxiv.org/search/cs?searchtype=author&amp;query=Chaffin,+A">Antoine Chaffin</a>, <a href="https://arxiv.org/search/cs?searchtype=author&amp;query=Sourty,+R">Raphaël Sourty</a></p><blockquote><em>Neural ranking has become a cornerstone of modern information retrieval. While single vector search remains the dominant paradigm, it suffers from the shortcoming of compressing all the information into a single vector. This compression leads to notable performance degradation in out-of-domain, long-context, and reasoning-intensive retrieval tasks. Multi-vector approaches pioneered by ColBERT aim to address these limitations by preserving individual token embeddings and computing similarity via the MaxSim operator. This architecture has demonstrated superior empirical advantages, including enhanced out-of-domain generalization, long-context handling, and performance in complex retrieval scenarios. Despite these compelling empirical results and clear theoretical advantages, the practical adoption and public availability of late interaction models remain low compared to their single-vector counterparts, primarily due to a lack of accessible and modular tools for training and experimenting with such models. To bridge this gap, we introduce PyLate, a streamlined library built on top of Sentence Transformers to support multi-vector architectures natively, inheriting its efficient training, advanced logging, and automated model card generation while requiring minimal code changes to code templates users are already familiar with. By offering multi-vector-specific features such as efficient indexes, PyLate aims to accelerate research and real-world application of late interaction models, thereby unlocking their full potential in modern IR systems. Finally, PyLate has already enabled the development of state-of-the-art models, including </em><a href="https://www.lighton.ai/lighton-blogs/lighton-releases-gte-moderncolbert-first-state-of-the-art-late-interaction-model-trained-on-pylate"><em>GTE-ModernColBERT</em></a><em> and </em><a href="https://www.lighton.ai/lighton-blogs/lighton-releases-reason-colbert"><em>Reason-ModernColBERT</em></a><em>, demonstrating its practical utility for both research and production environments.</em></blockquote><p>🌐 Learn more about<a href="https://www.lighton.ai/"> lighton.ai</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0b8ab45d61a7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LightOn delivers an Artificial Intelligence solution with support from Hewlett Packard Enterprise]]></title>
            <link>https://medium.com/@LightOnIO/lighton-delivers-an-artificial-intelligence-solution-with-support-from-hewlett-packard-enterprise-95cfc5ecb76f?source=rss-1a8df9096570------2</link>
            <guid isPermaLink="false">https://medium.com/p/95cfc5ecb76f</guid>
            <category><![CDATA[hewlett-packard]]></category>
            <category><![CDATA[generative-ai-market]]></category>
            <category><![CDATA[partnerships]]></category>
            <dc:creator><![CDATA[LightOn]]></dc:creator>
            <pubDate>Thu, 08 Aug 2024 11:49:54 GMT</pubDate>
            <atom:updated>2024-08-08T11:49:54.557Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GO7AyzqZaNj9Y2W8m6Vu0A.png" /></figure><p>LightOn, a French deeptech company specializing in generative AI for enterprises, Today announced that it is collaborating with Hewlett Packard Enterprise (HPE)</p><p>LightOn, a French deeptech company specializing in generative AI for enterprises, Today announced that it is collaborating with Hewlett Packard Enterprise (HPE) to develop a joint artificial intelligence offering with the Paradigm platform and HPE ProLiant DL380a Gen11 servers accelerated by NVIDIA H100 NVL 94GB GPUs with NVLink.</p><p>LightOn’s Paradigm platform integrates tools that enable business teams to incorporate generative AI into their business cases without technical team intervention. The Paradigm platform is able to plug into all open-source language models, offering maximum flexibility for enterprises.</p><blockquote>“Our commitment is to provide companies with generative AI tools that will revolutionize their operations, services, and products without compromising the security of their data. Our goal is to offer tools that are easy to deploy in their infrastructure, customizable to their needs, and capable of putting cost-effective solutions into production that deliver productivity gains” said<br>‍<strong>Igor Carron, President and Co-Founder of LightOn.</strong></blockquote><h3>A Complete Generative AI Solution for Businesses</h3><p>This fully integrated offering is designed for businesses of all sizes looking to implement a ready-to-use solution that includes a high-performance LLM and an intuitive business interface that makes usage and optimization extremely easy. Paradigm allows for the integration of internal document databases, which become queryable by a conversational agent. This joint offering reconciles confidentiality by design and data security with cost control.</p><blockquote>“This new era of generative AI enables businesses to transform data, providing new ways to uncover innovation and value. In order to meet the full potential of AI models, companies need powerful systems. By combining HPE’s compute-dense servers with LightOn’s learning models and intuitive user interface, we’re giving companies the means to unlock the value of their data, wherever it resides” said<br>‍<strong>Alain Melon, Managing Director, HPE France.</strong></blockquote><h3>Get the Most Out of AI with a Customized Solution</h3><p>To make the most out of what AI can offer businesses, the two companies also engage in joint research into new opportunities for innovation in generative AI to develop customized solutions for joint customers. This collaboration offers the opportunity to customize models to meet specific customer needs and objectives.</p><h3>About LightOn</h3><p><strong>‍</strong>LightOn is a French company and a pioneer in generative AI. Its Paradigm platform enables large-scale integration of generative AI into the processes of major corporations, while guaranteeing the confidentiality of data and know-how.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=95cfc5ecb76f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2023, The Rise of On-Prem LLMs]]></title>
            <link>https://medium.com/@LightOnIO/2023-the-rise-of-on-prem-llms-21ef0694bc98?source=rss-1a8df9096570------2</link>
            <guid isPermaLink="false">https://medium.com/p/21ef0694bc98</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[LightOn]]></dc:creator>
            <pubDate>Mon, 02 Jan 2023 14:41:44 GMT</pubDate>
            <atom:updated>2023-01-02T15:25:06.451Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*T2r6mrKL-kLq5YbIMC8YzA.png" /></figure><p>In 2021, AI ate HPC, but that news affected only the AI community and went mostly unnoticed by the general public. 2022 saw many upheavals, not least the release of ChatGPT. Unlike GPT-3 in 2020, <a href="https://trends.google.com/trends/explore?q=GPT-3,Chatgpt,Ukraine,BTS,Messi">the release of ChatGPT registered at the level of a world event</a>. That event showed that a powerful Large Language Model (LLM) tuned to human feedback and an intuitive UI changed the game in AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SlekwgqHulBRoNHTlWf-6w.png" /></figure><p>And here come the new interesting opportunities and new challenges:</p><p>For one, knowledge workers in all verticals are being served more uniformly. This is new. What traditionally was considered compartmentalized internal company know-how is now being uniformly funneled over to a partner API.</p><p>More importantly, data fed into the partner API may produce more powerful and capable partner LLMs. And then, the ease by which prompt retro-engineering can be undertaken sends a clear message: simply running a business on top of a partner LLM cannot be a moat for long unless you own the LLM or the partner!</p><p>Because data is text and everyone in the company can and wants to use it, whoever controls the LLM, controls the know-how of that company. In other words, the LLM has become a core strategic asset of the company using it.</p><p>At LightOn, we provide moats to companies by offering them the ability to host powerful LLMs on their infrastructure so they can retain their know-how and remain competitive. If you are interested in hearing more about <a href="https://lighton.ai/transform-your-enterprise-with-llms/">this opportunity, get in touch!</a></p><p>Happy New Year 2023, and Good Luck!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=21ef0694bc98" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LightOn publicly launches Muse, an API to VLM-4 Large Language Models trained natively in five…]]></title>
            <link>https://medium.com/@LightOnIO/lighton-publicly-launches-muse-an-api-to-vlm-4-large-language-models-trained-natively-in-five-5696b363929f?source=rss-1a8df9096570------2</link>
            <guid isPermaLink="false">https://medium.com/p/5696b363929f</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[gpt-3]]></category>
            <dc:creator><![CDATA[LightOn]]></dc:creator>
            <pubDate>Tue, 12 Apr 2022 20:01:46 GMT</pubDate>
            <atom:updated>2022-04-12T20:01:46.950Z</atom:updated>
            <content:encoded><![CDATA[<h3>LightOn publicly launches Muse, an API to VLM-4 Large Language Models trained natively in five European Languages</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/266/1*ZnnkmalSu_GqkIBu2YV8mg.png" /></figure><h3>Create, Process, Understand and Learn with LightOn Muse</h3><p><strong>Paris, April 12th, 2022.</strong> Paris-based <a href="https://www.lighton.ai/">LightOn</a> today announced the public release of <a href="https://muse.lighton.ai/">Muse</a>, an API that leverages state-of-the-art AI to democratize Natural Language Processing (NLP) for everyone, everywhere. Everyday workflows that include answering endless emails, searching and analyzing large amounts of documents or review answering can now be made easier and faster thanks to the use of <a href="https://muse.lighton.ai/">Muse</a>. Companies that seek to enhance their SaaS products can now improve the intelligence of their own products with the <a href="https://muse.lighton.ai/">Muse</a> API, a new one-of-a-kind powerful NLP engine, in several major European languages.</p><p>Like all Large Language Models, LightOn’s VLM-4 models are part of a new generation of AI for NLP that is revolutionizing the way business and science are being conducted. LightOn goes a step further by being the first company to publicly commercialize an API that accesses Large Language Models natively trained in European languages other than English. Muse users can now perform a large variety of text-related tasks directly in the language of their choice and avoid the hassle and complexity of a combination of English-only Large Language Models and costly translation APIs.</p><p>Key features of the LightOn <a href="https://muse.lighton.ai/">Muse</a> API include:</p><ul><li>business-ready intelligence primitives such as best in class summarization, steerable generation, text classification and search are accessible in just a few lines of code,</li><li>access to VLM-4, a suite of powerful and unique foundation language models trained on civilization scale high-quality curated data,</li><li>access to natively trained in many languages starting with French, English, German, Spanish and Italian — with more languages coming soon.</li><li>simplicity of use: Muse makes state-of-the-art language models directly actionable to any developers who are not specialists in AI or NLP.</li></ul><p>Muse customers include startups and large enterprises, whether they build innovative services and products or boost their digital transformation through AI. Every user can easily adapt Muse to perform the task they need through the use of skills currently available through the API or by creating new ones. Users of the Muse API can also exchange and share skills through the Muse community site.</p><p>Igor Carron, CEO and co-founder at LightOn, says about Muse and VLM-4:</p><p><em>“The Large Language Model paradox is that, while very hard to build, these models are intuitive and easy to use as they mimic natural human interactions. We believe it is essential to build these models natively, as every language reflects some cultural elements of the people who speak it — a key feature within European diversity. And as in the case of English Large Language Models, we see a large growth potential as the product serves virtually all areas of business. At LightOn, since we provided early beta access to Muse to selected business customers, we have witnessed an exponential increase in its usage.”</em></p><p>Several business customers who were part of the early Muse private beta program have this to say about the technology.</p><blockquote><em>“By combining our unique predictive SEO technology with the guided generation feature of Muse, Babbar is now able to offer its customers a unique tool that generates millions of words every month and is optimized for their specific use cases.”</em> Sylvain Peyronnet, CEO and Co-founder, Babbar.tech</blockquote><blockquote><em>“The ability of VLM-4 to follow instructions given in natural language is truly amazing.</em>” Mehdi Djabri, CEO and Co-founder, Iteration-X</blockquote><blockquote>” <em>We were able to improve our language assessment service and increase the quality and reliability for our customers, thanks to the multiple languages of Muse and the possibility to adapt the VLM-4 model to our specific use case.</em>” Baptiste Derongs, Co-Founder &amp; Director Strategy &amp; Development, Pipplet, an ETS company.</blockquote><p>LightOn’s <a href="https://muse.lighton.ai/">Muse</a> API is now publicly available at muse.lighton.ai with different subscription options available that include a free developer plan. To test the capabilities of VLM-4, the Muse website also features a no-registration required sandbox.</p><p><strong>About LightOn</strong></p><p><a href="https://www.lighton.ai/">LightOn</a> simplifies access to Artificial Intelligence by developing Large Language Models and making these NLP models available to all. Since it was founded in 2016, LightOn unlocks Extreme-Scale Machine Intelligence by re-imagining the whole computational stack and attendant algorithms. LightOn is funded by <a href="https://www.otiumcapital.com/">Otium Capital</a>, <a href="https://www.anorak.vc/">Anorak Ventures</a>, and <a href="https://www.quantonation.com/">Quantonation</a>.</p><p><a href="https://www.lighton.ai/">LightOn.ai</a></p><p>Contact Information: <a href="https://lighton.ai/blog/lighton-publicly-launches-muse-an-api-to-vlm-4-large-language-models-trained-natively-in-five-european-languages/contact@lighton.ai">contact@lighton.ai</a></p><p><em>All references to LightOn trademarks are the property of LightOn SAS. All other trademarks mentioned herein are the property of their respective owner.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5696b363929f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LightOn Newsletter| February 2022]]></title>
            <link>https://medium.com/@LightOnIO/lighton-newsletter-february-2022-296a3985f3ef?source=rss-1a8df9096570------2</link>
            <guid isPermaLink="false">https://medium.com/p/296a3985f3ef</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[vlm4]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[gpt-3]]></category>
            <dc:creator><![CDATA[LightOn]]></dc:creator>
            <pubDate>Thu, 24 Feb 2022 20:25:30 GMT</pubDate>
            <atom:updated>2022-02-24T20:25:30.222Z</atom:updated>
            <content:encoded><![CDATA[<p>LightOn’s Muse API and VLM-4 are coming!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Gr9Y7yM3Mk6lNPShTL1nQA.png" /></figure><p>In the course of the next few weeks, we, at <a href="https://lighton.ai/">LightOn</a>, will publicly release Muse, an API that accesses VLM-4, one of the Largest Language Model available in several European languages. Muse has been in private beta with several customers since October 2021 and has, for the past 7 weeks, featured an impressive <strong>doubling of the traffic every nine days</strong>. In the coming month, we will talk publicly about VLM-4, its unique features, and its potential impact in Europe and beyond. In particular, we will make an appearance at <a href="https://www.nvidia.com/gtc/session-catalog/?search=french#/session/16384400538550010MOV">NVIDIA’s GTC22</a> but also at the European Commission-sponsored <a href="https://lr-coordination.eu/6thELRC">6th European Language Resource Coordination conference</a> as well as at the “<a href="https://errin.eu/events/ai-industry-and-society-trust-and-frugality">AI for industry and society</a>” event organized as part of the French Presidency of the Council of the EU.</p><p>Stay tuned for more information!</p><p>Igor Carron,</p><p><a href="https://lighton.ai/">LightOn</a>, CEO and Co-founder</p><p>To register for our newsletter, <a href="https://lighton-2.hubspotpagebuilder.com/newsletter">you can subscribe here</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=296a3985f3ef" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2021, the year AI ate HPC … and more]]></title>
            <link>https://medium.com/@LightOnIO/2021-the-year-ai-ate-hpc-and-more-fbce08279528?source=rss-1a8df9096570------2</link>
            <guid isPermaLink="false">https://medium.com/p/fbce08279528</guid>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[photonics]]></category>
            <category><![CDATA[foundation-models]]></category>
            <category><![CDATA[gpt-3]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[LightOn]]></dc:creator>
            <pubDate>Thu, 30 Dec 2021 20:14:21 GMT</pubDate>
            <atom:updated>2021-12-30T20:14:21.779Z</atom:updated>
            <content:encoded><![CDATA[<h3>2021, the year AI ate HPC … and more</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RQn2mWWLRtK9yF_Xxovffw.png" /></figure><p>Back in 2011, Marc Andreesen announced that <a href="https://d12xxm04.na1.hubspotlinksstarter.com/Btc/2M+113/d12xxm04/VW9DbS4VckmXW5HYBXn2b7fZSW22vCbp4CJVj8MhPPS73lSc3V1-WJV7CgT5gW554m-p5-WH0_W276jv42hkNsHW4F9JXq5FKJC-W5g3Q5X38BYxsW3cYDr15Sx93VW9fsb7n3CGvpVW1-V4KL5dH0NJMkHvrymK3_XW4nPXFG1snCSKW4rS2rt7c97DbW2mLCzb7ZY5ghW7q6gcX5TbTGCW9hszGs7Bp7chW2P8qcW7gPXlqW6wLD8R22x9F8W6qYd3S2mJmmRW2XRP9n3cKhz1MP3XDxmGqnyW8NnPP88W3lhhW6BL1PW329Ln1W3XNW3p7DJzCrN1ByCJCnVNtPV-4G3M257y0GF4HvPkXGL1t38_B1"><strong>Software was eating the world</strong></a> while everyone was trying to make sense of the realities of the cloud versus brick and mortar businesses. Eight years later, Tarry Singh articulated how <a href="https://d12xxm04.na1.hubspotlinksstarter.com/Btc/2M+113/d12xxm04/VW9DbS4VckmXW5HYBXn2b7fZSW22vCbp4CJVj8MhPPQZ5knJ3V3Zsc37CgYyGW7_0T8Q2XsLKvV_PHYR3PHzxgW7DTR2c5-KrS-W7SBXxL2n5J-yW5CjjjT3X7y2TVcfj0m4f81McW3Q_6N_744Yf1W8J7lB86zzV58W6MGz8G5gtV8gW18-7G12fYLd-VZ6QXZ28DXm_W5174_S7jKtcYN3GRZTq-XxZ5W5LG30v5T6W5gW6Q8CP48QYZ4XW7CdP7P68vn8fW88_D8K35s59mN4yBXtfy-jrdW2tW1Js4_vhWrV19GRG6kTYxXW157XM48VqyBJW5hZqJ52g12x4VDwZyQ20NCN3W3_TwmN3q9lCLW3Y2KdG17qndTW7rQX4F4qxJZYW718Ljr6QkM3DW6RmF9P8XT9fxW4131hL7tcjyKW2f_bh76rldBmW8xNVp-944r3hW3lzwfh3h-bvs3c2W1"><strong>AI was eating software</strong></a>; a year before GPT-3 and Codex would give solid ground to this prediction. Fast forward two years later, we just witnessed how AI ate HPC and we believe those are the first steps towards how AI is eating Learning, Creative and Office work.</p><p>Let me explain.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qznP28wj1Bi8W1Tyau0fSA.png" /></figure><p>At <a href="https://lighton.ai/">LightOn</a>, we have been working on getting AI to be transformative for everyone. For that to happen, we used the Jean Zay French national supercomputer for two different yet somehow related reasons this past year. First, our <a href="https://d12xxm04.na1.hubspotlinksstarter.com/Btc/2M+113/d12xxm04/VW9DbS4VckmXW5HYBXn2b7fZSW22vCbp4CJVj8MhPPRS3lSbNV1-WJV7CgP-0W1Bt24G6bpyH8W3xxMJC1MDJXdN5X9p47fLy4tW6Hrny22W87QWW6KjY-16VYDhtW8zQ_HR875CczW7TMvVP6YZYGwW5xw-fq3q1ZsMVbqFMS53NpyHW2tzr1n1pWKRJN1LYSbvbT443W2Htky43HMmhQW42HQv_3B3dSYW3BpGqL1vvzTXW3q6lxs550rYMN3YN9fYM_gYjW2fDMPZ94MJLmMKYQM0qGClwW6m_Rpr4_GVF0N2PLW8dpppsqW1sCZzd31nC19W4CC07S4jGHtR3lrF1"><strong>LightOn’s Optical Processing Unit hardware</strong></a> was integrated into this top105 supercomputer. Even though LightOn’s hardware is analog and uses a technology currently unknown to supercomputing, there are several good reasons the future of computing will use this technology. Relatedly, in a co-design fashion, we also used the Jean Zay facility to implement and run code for the building of Large Language/Foundation Models that we believe are key to Transformative AI. In March, we trained the largest French language model ever called Auriga and made it available to everyone through our PAGnol demo.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*nCigYB5bOoGsLgvsuuHNXw.gif" /></figure><p>In July, we launched the <a href="https://d12xxm04.na1.hubspotlinksstarter.com/Btc/2M+113/d12xxm04/VW9DbS4VckmXW5HYBXn2b7fZSW22vCbp4CJVj8MhPPRy3lSbtV1-WJV7CgL7wW5LGWrl4Hd1WCW69Sh-4146fFBVdwwz711Yy2qW1f9WNr4DZmTMW4S18ns7p4p6bW4k0cqS76vM1gW5_szny46NbXfW4KMfzW3HlQ_9W6xmMBv4GGqrNW3blVVb7DyYnPN7__QFVjK-pVW8KyPK82FkvMMW58rk3c7dZFN0W8L2PjR66wdrRW7362c18dxD4MW2vk_rr8JspNyW7T5qhN5VSQVkW8cktW44kgW3zN2SFLDqJ05HCW8Sp-Dr4Vm7qG3l3F1"><strong>Muse API</strong></a>, making our language models available for business use. Initially released in private beta, <a href="https://d12xxm04.na1.hubspotlinksstarter.com/Btc/2M+113/d12xxm04/VW9DbS4VckmXW5HYBXn2b7fZSW22vCbp4CJVj8MhPPRS3lSbNV1-WJV7CgNmpW8qvbkH1z6zVwV7Jdyj1LtMksVcDlg-9dYH5WW2fCvyj26m77JW11rXq38ghnJTN7v1cgDK-4xjVM4YkV4pwC7RW1gJwJj6q0hm3N9hb_SNTtFVNW1dd4s41rCjB0W1gBsNd7h-Ry3W1B9Jck7gVmZ6W8Jnx0Z8W7Hy2W37xxkz2TdX7xN7xJ7ZhfTwCFW5tb-qx70Jp1KW2HB1Xy7h6NDFW4ZFl0P4tJ-Q0W1SnZF52ryPfcW5zLzkm27kl3kW4DKbl_8yfCRbW59760p7p1_v92p-1"><strong>Muse</strong></a> has quickly gained its first customers, and a public commercial version with five languages is to be released in early 2022. Some of these early customers are using this new AI to redefine SEO or the experience for website creation.</p><blockquote><em>“True happiness comes from the joy of deeds well done, the zest of creating things new” </em><strong>Antoine de Saint-Exupéry</strong></blockquote><p>Eventually, a major impact of these Large Language Models trained on HPC infrastructures will be the ability <strong><em>for everyone to personally learn faster and for office workers worldwide to get the job done in a fashion never seen before</em></strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/890/1*3GpOlzlvwcNh0unT-44dxw.png" /></figure><p>If you are a start-up company or an individual starting a business around this promise, don’t hesitate to join the <a href="https://d12xxm04.na1.hubspotlinksstarter.com/Btc/2M+113/d12xxm04/VW9DbS4VckmXW5HYBXn2b7fZSW22vCbp4CJVj8MhPPRS3lSbNV1-WJV7CgDL2W6Bc77b3WsYwZW5hGj_45lGxMDW67rBwL1Xs8wkW71QqSp2FDS2TW2jLDb3835nkjW60r8Dc4zF3_vW7zsHrF46q2NPVCpGk64WtcqcW2P99wl8Bb3hcW8dcgZ21SRY76W2rVb8X7tckh4V8vrjB3YkYQzW59kbD92_CB2mW8c91SM4nV6S5W3bR41p2dZx26W3qk9Fp7Zy9KTW5k89Gn960z4gW83FZBr8VMNBdW4l4WQs9072QYW1qkbC08ZZML6W8psnN05hdDPFW1S9clV2kxq4M3d2x1"><strong>Muse Partnership program</strong></a>, and let’s start a discussion around how <a href="https://d12xxm04.na1.hubspotlinksstarter.com/Btc/2M+113/d12xxm04/VW9DbS4VckmXW5HYBXn2b7fZSW22vCbp4CJVj8MhPPRS3lSbNV1-WJV7CgJb0W3B5Kkl2V31BfW7lG4NX1PDck-W4WWC_r67y4s5W5Yx4rp2298GpW6DCHXP6kDf2fW72cK3q55T1W7W509Btr6j8KBjW7BYS5t705fpyW6Sb-Pv492Gr5W3wfRTc94kmjgN82RN7Z6H5JjW6mz2Zn4PN7XJW1zF9bj13189HN8_gfWfB7DvNW1cRyf74LldPvMVYq1B8_pTYW4hQCw85F0jDgW3f64ln3xk5_LN3jLrw7HdG7LW50Rp4597Rh-ZW4_T0hR6p9H1MW3lpvTm6n1fSt3d5b1"><strong>Muse</strong></a> can help you.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/840/1*YlC2rRGuICXEFBU2Tr8JQA.jpeg" /></figure><p>These models will also have the same effect in creative work and in the discovery process. We are currently <a href="https://d12xxm04.na1.hubspotlinksstarter.com/Btc/2M+113/d12xxm04/VW9DbS4VckmXW5HYBXn2b7fZSW22vCbp4CJVj8MhPPT13lScZV1-WJV7CgY7CW4c6x7C9bm7QRW3816Vj1C1NJJW906bX74rPPFMW2JSTXp5HVPjJV7MVzw1KdFSBW9dQR9p26hfZ-W44QPJc783kfxW7FQzNn5hNJn_W3BQtlb643sM3N7rQRsMMgNN-W573zHC2RxXgVW4FB-gS4cf26mW1-pbkQ6Ltt77W15h6rm75y2z7W7S7533347lRhW4vQf-h8tZ_D4W17VBKJ3klQxgW8jt8Kc1P5fYCN19cfH6jtBDwW1l0MkS6j8WjnW8N_H4T7M31HVW3qv0lJ6SR7vWW7RsjTW5VFm4VVpD9kW6zF1g6W7_fpz-2dQrV4W3jrb-h4pr5krW4w9jHW1xScNzMbYY9xyjW3CW13_Dm11KQ85sN1kz6-Wp83n73lz71"><strong>partnering with TII in the UAE to build an exascale HPC facility</strong></a> for the purpose of <a href="https://d12xxm04.na1.hubspotlinksstarter.com/Btc/2M+113/d12xxm04/VW9DbS4VckmXW5HYBXn2b7fZSW22vCbp4CJVj8MhPPRy3lSbtV1-WJV7CgTXjN7fZMPLG-v1CW3pp-jq7K6lFjMWqWtp8_BgSW1CDp5v62H9-rW8FwkP04B7Q9ZW78gFXJ41nBhGN3rdHXh2F7xwW2lC07w6hkKj8Vlr1gK55nb91V7Wfm51nQVwDW8G9_mG1PTSHyW6vtC698qKT5RW5ltscL3tfbBxN7Ky-cM_r1RNW3D7HDC3LgNl4W5rW7867-NBThW95plPk6qZhz8W2ZmB4Y4K8VhBW8WW-CR6zcm_QW5_yJwQ6p9wbK3nlY1"><strong>training and using these models to serve scientific and engineering developments</strong></a>.</p><p>Stay tuned, the true AI revolution is really coming!</p><p><a href="http://www.linkedin.com/in/IgorCarron">Igor Carron</a></p><p><a href="https://lighton.ai/">LightOn</a>, CEO and co-founder</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fbce08279528" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LightOn Photonic coprocessor integrated into European AI Supercomputer]]></title>
            <link>https://medium.com/@LightOnIO/lighton-photonic-coprocessor-integrated-into-european-ai-supercomputer-793e5eca98bb?source=rss-1a8df9096570------2</link>
            <guid isPermaLink="false">https://medium.com/p/793e5eca98bb</guid>
            <category><![CDATA[supercomputer]]></category>
            <category><![CDATA[photonic-computing]]></category>
            <category><![CDATA[hpc]]></category>
            <category><![CDATA[hardware]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[LightOn]]></dc:creator>
            <pubDate>Tue, 21 Dec 2021 19:17:38 GMT</pubDate>
            <atom:updated>2021-12-21T19:17:38.403Z</atom:updated>
            <content:encoded><![CDATA[<p>Four years ago to the day, <a href="https://lighton.ai/">LightOn</a>’s <a href="https://medium.com/@IgorCarron/first-light-at-lighton-8ddd0f20388f">first Optical Processing Unit (OPU)</a> had its first light in a Data Center showing that our technology was data center ready.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qznP28wj1Bi8W1Tyau0fSA.png" /></figure><p>It is with immense pride and pleasure to announce that LightOn’s OPU has been installed in one of the world’s Top500 supercomputer as part of a pilot program with GENCI and IDRIS/CNRS.</p><p>The team at LightOn is immensely proud to write the future of computing in this world-first integration of a computing photonic device in an HPC infrastructure.</p><p>The press release can be found <a href="https://lighton.ai/wp-content/uploads/2021/12/Press-Release-LightOn-Photonic-coprocessor-integrated-into-European-AI-Supercomputer-Dec-21-2021.pdf">here</a>.</p><p>Thank you GENCI and IDRIS/CNRS for making this happen!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=793e5eca98bb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LightOn: Bringing containers to light]]></title>
            <link>https://medium.com/@LightOnIO/lighton-bringing-containers-to-light-61b6d655f8d2?source=rss-1a8df9096570------2</link>
            <guid isPermaLink="false">https://medium.com/p/61b6d655f8d2</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[docker]]></category>
            <category><![CDATA[photonics]]></category>
            <category><![CDATA[data-center]]></category>
            <category><![CDATA[linux]]></category>
            <dc:creator><![CDATA[LightOn]]></dc:creator>
            <pubDate>Fri, 08 Oct 2021 07:47:07 GMT</pubDate>
            <atom:updated>2021-10-08T07:47:07.798Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*b8rWFzhZWtDmZAY6IoLmxw.png" /></figure><p>Since Dec 2017 (closed beta) and April 2020 (official launch), LightOn has been offering online access to optical computing through the <a href="https://cloud.lighton.ai/">LightOn Cloud service</a> that features its Optical Processing Unit technology. This was the first time optical processing was made available on a cloud platform, worldwide.</p><p>At LightOn, we keep the innovation going on all aspects of hardware and software, to better serve our users. We recently migrated our cloud infrastructure to containers with great features for LightOn Cloud users. We had fun doing it, and wanted to share some of the details that stood out during this project.</p><h3>Containers 101</h3><p>You have probably heard of containers here and there: it’s now a widely used technology that changed how cloud services work. Mainly popularized by<a href="https://docker.com/"> Docker</a> and<a href="https://kubernetes.io/"> Kubernetes</a>, it’s a great ensemble of techniques that allows services to run isolated from the computer host, in a portable way. The name <em>containers</em> really stands for the portability it offers: once you’ve built a container image, you can run it without any other dependency than a common runtime should your service be a simple Python script, or a complex application with many runtime dependencies.</p><p>Containers revolutionized the way developers define applications and services, because an image can be built from a self-contained file, commonly named Dockerfile, allowing for the description of the runtime dependencies in code.</p><h3>How containers can help LightOn Cloud</h3><p>Until now, when booking a time slot on <a href="https://cloud.lighton.ai/">LightOn Cloud</a>, the server that holds the desired OPU is restarted from a predefined image, in order to have a clean state for each user session. This worked great for a year and a half. However it showed severe limitations for both users and support staff. Rebooting a datacenter-grade computer takes several minutes, so users have to wait for up to 10 minutes from the moment the booking is made, until it is ready. In order to maintain or enhance the environment provided to users, our staff needs to make changes directly into boot images, which is error-prone and difficult.</p><p>Seeing these major pain points we decided to move our infrastructure into a container environment. Instead of completely starting the server, the environment for LightOn Cloud will be in a single container, which would be much faster to start, and easier to maintain.</p><h3>Hardware access in a container</h3><p>But <a href="https://cloud.lighton.ai/">LightOn Cloud</a> has <a href="https://lighton.ai/photonic-computing-for-ai/">great hardware</a>, right? This is what spices up our project…</p><p>Each cloud server is connected to both a GPU and an <a href="https://lighton.ai/photonic-computing-for-ai/">OPU</a>, which is why data scientists and researchers use LightOn Cloud. If containers are isolated from the host, how do you access these hardware accelerators? Docker allows us to add a tunnel to this isolation, by mapping devices and other filesystem objects from the host to the container. This can be done with the following directive in a docker-compose.yaml file:</p><pre>devices:<br>  - /dev/nvidia0<br>  - /dev/nvidiactl<br>volumes:<br>	- /usr/lib/x86_64-linux-gnu/libnvidia-ml.so.${NVIDIA_DRIVER_VERSION}:/usr/lib/x86_64-linux-gnu/libnvidia-ml.so.${NVIDIA_DRIVER_VERSION}:ro<br>  - /usr/lib/x86_64-linux-gnu/libcuda.so.1:/usr/lib/x86_64-linux-gnu/libcuda.so.1:ro<br>	- /usr/bin/nvidia-smi:/usr/bin/nvidia-smi:ro</pre><p>In this snippet we tell it that Nvidia devices are available in the container, as well as the associated shared libraries. Since it’s a native mapping, you can expect bare-metal performance with these settings.</p><p>You also need to allow it at the c-groups device level. C-groups is at the core of the Linux kernel resources management, giving access to devices from kernel to user-space. Devices in Linux are identified with major and minor numbers, and those can be found in /proc/devices. We first query for the GPU device number:</p><pre>opu@compute-01:/home/opu$ grep nvidia /proc/devices<br>195 nvidia-frontend</pre><p>The c-groups directive is then added to the docker-compose.yaml:</p><pre>device_cgroup_rules:<br>  # 195 pour nvidia<br>	- &#39;a 195:* rmw&#39;</pre><p>The <a href="https://lighton.ai/photonic-computing-for-ai/">OPU device</a> is no more different than the GPU, since it’s seen as a Linux device as well. Problem solved!</p><h3>Bringing things together</h3><p>Once we have our devices accessible from the container, we need to provide the software stack. As you usually do in Docker, we start from an official image,<a href="https://hub.docker.com/r/continuumio/anaconda3/"> the Anaconda image</a>. The Dockerfile instructs on installing<a href="https://github.com/lightonai/lightonml"> lightonml</a>, and all the main libraries that data scientists are used to working with.</p><p>When the container starts, jupyterlab and SSH service are both started in the user’s entrypoint, and a<a href="https://traefik.io/"> Traefik service</a> is put along in the docker-compose, that will do https proxying from the host to the container.</p><p>All we need to do then is to answer booking requests by orchestrating the containers’ start and stop from our cloud scheduler, customizing them with user details in environment variables that are passed to the Docker engine.</p><h3>Conclusion</h3><p><a href="https://cloud.lighton.ai/">LightOn Cloud</a> is now fully containerized and now gives users the ability to start their booking in less than 5 seconds, instead of “traditional” 8 to 10 minutes.<a href="https://cloud.lighton.ai/"> You can try it yourself now, for free, through our trial and the LightOn Cloud for Research program!</a></p><p><strong>We’d like to thank the awesome guys at </strong><a href="https://enix.io/en/"><strong>Enix</strong></a><strong> that provided their invaluable expertise on containers for this project.</strong></p><p><em>Author:</em> <a href="https://www.linkedin.com/in/chbrosso/"><em>Charles Brossollet</em></a><em>, Lead developer at LightOn</em></p><p><em>All references to LightOn trademarks are the property of LightOn SAS. All other trademarks mentioned herein are the property of their respective owners.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=61b6d655f8d2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LightOn Newsletter, September 2021]]></title>
            <link>https://medium.com/@LightOnIO/lighton-newsletter-september-2021-38cf8534b8a6?source=rss-1a8df9096570------2</link>
            <guid isPermaLink="false">https://medium.com/p/38cf8534b8a6</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[new-foundation-model]]></category>
            <dc:creator><![CDATA[LightOn]]></dc:creator>
            <pubDate>Fri, 10 Sep 2021 18:09:08 GMT</pubDate>
            <atom:updated>2021-09-10T18:25:36.975Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/872/1*L0OJl4OtI2za6X9RVp_ohA.png" /></figure><h3><strong><em>An Urdu joke is better told in Urdu than if it were translated into English or French</em></strong></h3><p>Dear reader,</p><p>You probably have noticed our <a href="https://lighton.ai/"><strong>new website</strong></a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AEdUyDpiIeDs3UdfyHsd1A.png" /></figure><p>Let me provide you with some context as to why we are prominently featuring <a href="https://muse.lighton.ai/"><strong>Muse</strong></a>, our API for foundation models.</p><p>In 2012, Deep Neural Networks rocked the Computer Vision world. Eight years later, the release of GPT-3 is similarly changing the course of Natural Language Processing and Artificial Intelligence as witnessed by the recent rare <a href="https://arxiv.org/abs/2108.07258"><strong>ArXiv report of more than 100 Stanford researchers</strong></a> on the opportunities of these foundation models.</p><p>Up until a year ago, <a href="https://lighton.ai/"><strong>LightOn</strong></a>’s main focus had been to get hardware technology to open new avenues of computing. Our photonic technology started in the lab and now works in Data Centers while performing massively parallel computations. We are continuing its scaling so that LightOn OPUs can eventually train and even perform inference for these massive models.</p><p>Another side to LightOn’s story comes out of our hardware-algorithm in-house co-design efforts: Better hardware requires being deeply knowledgeable about software and algorithms. For the past year, our <a href="https://d12xxm04.na1.hubspotlinksstarter.com/Btc/2M+113/d12xxm04/VX5DG92_N3_bW769MB-7hnR0cW2K0rfL4xj2ytN1MKFZG3lSb9V1-WJV7CgSSwW86pxTd4R0R81W243M8-7RS8VsW7nlWXk1vZsWjW7CjwHs1y0BXhW5n-tWn2K_SNLW3RdZHt831hSjN1SYFlk9bLNJW97xDZw1jxt7dW4dZcPr4KHxy1W1WhKF43g9R5NW3xfmgT1TMZG6W6lL_Qy3TwyWqW4hPlsT5dts-kW7fW_6v93LRLwW8PK-S_97fwFhW6TH-2n5pkZC8W5VW6Qw8hHn5cW1LFwWz4Hzxbz3p4d1"><strong>AI team</strong></a> has focused on getting to be the best at building these extremely complex foundation models. We released <a href="https://pagnol.lighton.ai/"><strong>PAGnol</strong></a>, a french GPT-3/foundation model, and have continued our efforts by releasing a Spanish and an Italian one. <a href="https://lighton.ai/"><strong>LightOn</strong></a> is also part of the team involved in the Big Science project that aims at building a 200 billion parameter model. As the <a href="https://lair.lighton.ai/akronomicon/"><strong>Akronomicon</strong></a> leaderboard shows, we enjoy being in the company of giants 😉.</p><p>Our next step is to make these models available and we are doing so through <a href="https://muse.lighton.ai/"><strong>Muse</strong></a>, the API for the foundation models available at <a href="https://lighton.ai/"><strong>LightOn</strong></a>. Our vision is to provide these foundation models to the 80%+ of the world population that does not interact daily in English. Nuances in an Urdu joke given in Urdu will probably not translate well in English and that is why, we believe, every model ought to be language-specific. While our future offering will feature an English model, it will eventually cover natively 40 major languages. Initial commercial availability with few languages will start in early 2022. <a href="https://muse.lighton.ai/"><strong>Muse</strong></a>’s name was chosen because the API is expected to be a conduit for inspiration to its users in the same way Zeus’ daughters have been for literature, science, and the arts for the past 2600 years.</p><p>The <a href="https://muse.lighton.ai/"><strong>Muse</strong></a> API is currently running in private beta with several customers. If you or your business are interested in accessing it, please <a href="https://lighton.ai/contact-us/"><strong>contact us</strong></a>.</p><p>Forward we go!</p><p>Igor Carron, CEO, LightOn</p><p><em>Muse Press release:</em></p><h3><strong>Create, Process, Understand and Learn with LightOn Muse 🎭</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cj2AeAmB4QS9iGtw.png" /></figure><p>At LightOn we are proud to unveil Muse, an <a href="https://d12xxm04.na1.hubspotlinksstarter.com/Btc/2M+113/d12xxm04/VX5DG92_N3_bW769MB-7hnR0cW2K0rfL4xj2ytN1MKFZG3lSb9V1-WJV7CgJzhW3LH8DN6mmvNxW3YwZSy2c8rW-W5-b2lg17R8QxW8M226M1FHmy9W9gPqZt8HmjSJW2vzbyF7y3hDXN2KSG_yHGF8NW3XHclr5zxkNpW74dZrd4L7_pCW81xFS44PxvjbW4c4gH1874f4-W6vbxQq3F6C3dW7nTdm05jGxkYVDr6j86KC3HGW5k5BLf168m9yW8VmMYz6-mgrRW87fs9G63bp2BW1_pLf85g3h1T3gSJ1"><strong>extreme-scale language API</strong></a> to enable the integration of <strong>human-like intelligence in any product or application</strong>, in any language.</p><p>Muse is <strong>already available in private beta</strong> and we are working closely with our beta <strong>users</strong> and <strong>customers</strong> to fine-tune our offering in order to <strong>serve their business needs</strong> in the best way possible.</p><p>Our existing customers span across multiple industries including <strong>SaaS platforms</strong>, <strong>SEO</strong>, or <strong>EdTech</strong>, taking advantage of the broad capabilities of the LightOn Muse API.</p><p>Some of the Muse <strong>highlights</strong> include:</p><ul><li>business-ready intelligence primitives accessible in just a few lines of code, such as best in class <strong>summarization</strong>, <strong>steerable generation</strong>, <strong>text classification</strong>, and <strong>search</strong>,</li><li>powerful and unique foundation models trained on <strong>civilization-scale high-quality curated data</strong>,</li><li>for the first time, the ability to serve the latest AI model, <strong>natively trained in many languages</strong>, starting with French, Spanish, and Italian — with more coming soon!</li></ul><p>You can access Muse in private beta today to empower your products with business-ready intelligence primitives, accessible in just a few lines of code.</p><p><a href="https://muse.lighton.ai/">Learn more about the Muse API</a></p><h3><strong>Next LightOn AI Meetup on Thursday, September 16</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/799/0*OU2fb0VIkaaE3l94.png" /></figure><p>For our 15th AI Meetup, we will host <strong>Federico Bianchi</strong>, a Postdoctoral Researcher at Università Bocconi. Federico will talk about <strong>Contrastive Language-Image Pre-training for the Italian Language</strong> (CLIP). Join us on Thursday, September 16, 2021, from 4:30 PM to 5:30 PM CEST.</p><p><a href="https://www.meetup.com/LightOn-meetup/events/280482787/">Subscribe to our free AI Meetup here</a></p><h3>The next episode of the BigScience workshop</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/318/0*nKmH28_7kD7dN28y.png" /></figure><p>The next big live event <strong>BigScience Episode 2</strong> is happening on September 20th and you can register <a href="https://bigscience.huggingface.co/en/"><strong>here</strong></a>. The event will be co-located with INLG 2021, the <a href="https://inlg2021.github.io/pages/registration.html#registration"><strong>14th International Conference on Natural Language Generation</strong></a><strong>. </strong>Our very own Julien Launay will talk about <strong>Scaling Modelling Decisions to the &gt;100B scale</strong>.</p><p><a href="https://bigscience.huggingface.co/en/">Register for the BigScience workshop</a></p><h3><strong>Igor Carron @ TechBBQ</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/357/0*UAeLv_CrDE85hVFA.png" /></figure><p>Our CEO and Co-founder, Igor Carron will be at <a href="https://techbbq.dk/"><strong>#TechBBQ21</strong></a> on 16–17 September 2021 in Øksnehallen, Copenhagen to talk about <strong>The next generation of sustainable machine learning.</strong></p><h3>Laurent Daudet @ ECOC</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/357/0*pVb0wTERImFHBfXE.png" /></figure><p>On Monday 13 September, our CTO and co-founder Laurent Daudet will talk at the <a href="https://www.ecoc2021.org/programme/workshops"><strong>Neuromorphic Computing workshop at ECOC 2021</strong></a>. Don’t miss the chance to watch his presentation about <strong>Scaling up AI with Photonics Computing.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=38cf8534b8a6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Create, Process, Understand, and Learn with LightOn Muse]]></title>
            <link>https://medium.com/@LightOnIO/create-process-understand-and-learn-with-lighton-muse-7e7e35c7f63a?source=rss-1a8df9096570------2</link>
            <guid isPermaLink="false">https://medium.com/p/7e7e35c7f63a</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[new-foundation-model]]></category>
            <category><![CDATA[gpt-3]]></category>
            <category><![CDATA[api]]></category>
            <dc:creator><![CDATA[LightOn]]></dc:creator>
            <pubDate>Fri, 03 Sep 2021 07:44:03 GMT</pubDate>
            <atom:updated>2021-09-03T07:44:03.086Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A simple API to bring human-like understanding &amp; skills to any application, in any language</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tkWWOLdJ9c-fV9tqBLNBHw.png" /></figure><p><strong>PARIS, France, August 31st, 2021</strong> — Paris-based AI startup <a href="https://lighton.ai/">LightOn</a> today unveils <a href="https://muse.lighton.ai/">Muse</a>, a service that integrates human-like intelligence in any product or application, in any language.</p><p>A year ago, OpenAI sparked a paradigm shift in AI by releasing GPT-3, a generally capable large language model. Up until now, such technology has only been commercially available in English.</p><p>Using its unique know-how in combining <a href="https://lair.lighton.ai/akronomicon/">extreme-scale AI</a> and <a href="https://lighton.ai/photonic-computing-for-ai/">innovative hardware</a>, LightOn is furthering these advances by offering access to large models in various languages. Key features of the LightOn Muse API include:</p><ul><li>business-ready intelligence primitives accessible in just a few lines of code, such as best in class summarization, steerable generation, text classification, and search,</li><li>powerful and unique foundation models trained on civilization scale high-quality curated data,</li><li>for the first time, the ability to serve the latest AI model natively trained in many languages, starting with French, Spanish and Italian — with more coming soon.</li></ul><p>SaaS platform developers can now access the incredible power of state-of-the-art AI models without any knowledge of Natural Language Processing or Machine Learning. Current customers span across multiple industries including SaaS platforms, SEO, or EdTech, taking advantage of the broad capabilities of the LightOn Muse API.</p><p>Igor Carron, LightOn’s CEO, declared about the release of the Muse API</p><blockquote>“<em>Thanks to the LightOn Muse API, our clients can now produce new services unthinkable even a year ago, all this with a local touch. As a result, the API can substantially improve SaaS client’s experiences and enhance the stickiness of their services. This new API product is the result of LightOn’s recognized deep expertise in efficient hardware, making this new extreme-scale AI both scalable and sustainable. With this new tool, we lower the barrier to entry of Transformative AI across businesses worldwide</em>”</blockquote><p>Sylvain Peyronnet, CEO of Babbar.tech said about LightOn Muse technology:</p><blockquote><em>“Babbar uses the LightOn Muse API to help copywriters write better text, faster. By incorporating our unique predictive SEO technology into the steerable generation options provided by the Create endpoint, at Babbar we deliver millions of uniquely optimized words to our customers every month.”</em></blockquote><p>The LightOn Muse API is now available in private beta for business customers. More information at <a href="https://lighton.ai/">https://LightOn.ai</a></p><p><strong>About LightOn</strong></p><p>LightOn unlocks Extreme-Scale Machine Intelligence by re-imagining the whole computational hardware and algorithmic stack. LightOn’s portfolio spans several products ranging from photonic hardware to software APIs serving some of the largest scale AI models in the world. Founded in 2016, LightOn is funded by Otium Capital, Anorak Ventures, and Quantonation. <a href="https://lighton.ai/">LightOn.ai</a></p><p>This press release is also available in pdf in <a href="https://lighton.ai/wp-content/uploads/2021/08/LightOn-MUSE-Press-Release-August-31-2021.pdf">English</a>, <a href="https://lighton.ai/wp-content/uploads/2021/08/LightOn-MUSE-Communique-press-31-Aout-2021.pdf">French</a> and <a href="https://lighton.ai/wp-content/uploads/2021/08/LightOn-MUSE-Communicato-Stampa-Agosto-31-2021.pdf">Italian</a>.</p><p>All references to LightOn trademarks are the property of LightOn SAS. All other trademarks mentioned herein are the property of their respective owners.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7e7e35c7f63a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>