Noeon Research

“We do not need to scale data to get new capabilities. Instead we have intellectual capabilities already encoded in how we decompose the process, how we run the inference.” Last month, Noeon Research’s Chief Scientist Andrey Filchenkov presented a talk at the EUinJapan DEL Japan AI Bridge forum. Borrowing the “neat – scruffy” paradigm of AI systems, Andrey explained how large language models had to become scruffy in order to be practical, and why this puts limits on their capabilities. He also explained how Noeon Research is working on a new, neat unified framework to pursue artificial general intelligence. #Noeon #AGI #technology #agenticai #agenticsystems #AI Watch Andrey’s talk in full below:

Noeon Research is seeking a highly skilled LLM / NLP Engineer to join our team. We are developing a novel natively agentic AI system with general capabilities. More information about this position: https://lnkd.in/dTXJ2ntP 🔴 Responsibilities: ▫ Develop and implement solutions leveraging open-source large language models (LLMs) to address specific business needs. ▫ Fine-tune, customize, and optimize open-source LLMs for various applications, ensuring high accuracy and efficiency. ▫ Evaluate and compare different open-source LLM architectures to select the most suitable models for given tasks. ▫ Integrate LLMs into existing systems, ensuring seamless functionality and performance. ▫ Stay updated with advancements in the field of natural language processing (NLP) and open-source LLMs, incorporating new methods and tools as appropriate. ▫ Collaborate with cross-functional teams to define product requirements and deliver solutions aligned with organizational goals. 🔴 You will be a good fit if you have: ▫ Proven experience in working with large language models (e.g., GPT, LLaMA, T5, BERT). ▫ At least 2 years of experience in Machine Learning or Natural Language Processing (NLP). ▫ Hands-on experience in fine-tuning and deploying language models for specific tasks. ▫ Proficiency in Python and ML frameworks (e.g., TensorFlow, PyTorch). ▫ Experience with NLP libraries (e.g., Hugging Face Transformers, spaCy, NLTK). ▫ Proactive mindset to stay updated with the latest advancements in AI and LLMs. ▫ Fluent in conversational and written business English (C1+). 🔴 What we are offering: ▫ Full-Time Remote Job from anywhere in the world. However, EMEA / APAC preferred for a better time zone overlap with the Tokyo HQ. ▫ Competitive Salary. Depending on the candidate's skill level, we are willing to offer 14-17M JPY and equity compensation (options). ▫ Flexible Schedule. Outside scheduled team meetings, teammates are free to work on their tasks independently. ▫ Medical Allowance. Compensation of overseas health insurance and additional medical costs. ▫ Language Courses. Support for English and Japanese language training. If you believe you meet these requirements and are passionate about driving innovation in AI architecture, we encourage you to apply for the LLM / NLP Engineer position at hiring@noeon.ai #techjobs #hiring #NLPEngineer #LLMEngineer

Tomorrow (30 October) we have the pleasure of presenting our research at the EUinJapan DEL Japan AI Bridge Event (link to: https://lu.ma/uk2s3jl2), organized by Andrea Sansica and Ilya Kulyatin. Our Lead ML Engineer, Andrey Filchenkov, will talk about "General Intelligence from First Principles" and how Noeon is working on a novel natively agentic AI system with general capabilities. We will also have a table for attendees to ask about our research. If you're coming, please feel free to stop by and say "hi"! #Noeon #startups #AGI #technology #agenticai #agenticsystems

Some popular social media websites, like LinkedIn, have an #AI assistant that can help you finish an email draft, or rewrite a post using their AI model. But at what point does your original content no longer become your own, and instead become "artificial"? Where is the boundary? How do others determine which parts of your writing are real or fake? Detecting this boundary is the goal of a new study "AI-generated text boundary detection with RoFT", involving a member of our team, German Magai, and his co-authors. This work received the outstanding paper award at Conference on Language Modeling. The study examined several ways to adapt state-of-the-art artificial text detection classifiers to the boundary detection setting. Here are some of the key findings and outcomes: 1. Differences in length distribution across domains seem to play an important role, obstructing out-of-domain generalization. 2. The “Recipes” domain of dataset ROFT often contains numbered lists, and the first sentence tends to be comparatively long. As a result, the "Recipes" topic is hard for all classifiers to generalize to. 3. Cross-model generalization of many classifiers fails not only for the best AI text generators but also for the worst generators. At the same time, bad generators are very easy for human raters, because of the high level of semantic inconsistencies and grammar issues. 4. Human-written texts in ROFT tend to have 2-3 peaks of perplexity that may correspond to a plot twist or change of narrative focus. The perplexity drop at the beginning of the fake text is conflated with the same drop due to discourse changed by perplexity-based detectors. It’s becoming the new norm to use tools to help complete or refine human texts using auto-generation. In what ways has it been most helpful for you? Let us know in the comments. https://lnkd.in/eZumvEkv #COLM2024 #ATD #GenAI #RoFT #ML #technology

#NoeonResearch develops a novel, natively agentic #AI architecture with a number of essential design properties in mind, the most important of which include: ⁃ 𝐈𝐧𝐧𝐞𝐫 𝐚𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭. The system represents the goals in a discrete and explicit form and pursues them consistently and without deviations. ⁃ 𝐈𝐧𝐭𝐞𝐫𝐩𝐫𝐞𝐭𝐚𝐛𝐢𝐥𝐢𝐭𝐲. Discrete knowledge representation enables the user to inspect the system’s knowledge base and reasoning trajectory at any level of abstraction. ⁃ 𝐑𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠 & 𝐝𝐚𝐭𝐚 𝐬𝐞𝐩𝐚𝐫𝐚𝐭𝐢𝐨𝐧. Abstract reasoning abilities are independent of domain-specific knowledge, ensuring that our system is stable and efficacious in any domain regardless of the task complexity. By having these properties, our architecture will be capable of solving intellectual tasks – such as information retrieval, reasoning and generation – in a robust and reliable way. This will enable its use across most value-adding activities such as software development, enterprise automation and industrial design, as well as safety-critical areas including high-precision manufacturing, renewable energy and complex infrastructure. Learn more about our approach at https://noeon.ai/. #AI #AIResearch #AICommunity #ArtificialIntelligence

Closing off the summer of conferencing, our very own Blaine Rogers and Simeon Ganiushin are in Las Vegas for Ai4 - Artificial Intelligence Conferences! Hit them up if you want to grab a coffee and learn more about #NoeonResearch. #AI4 #NoeonAI #AI #Innovation 

Our team - Darya Zvereva, Yuriy Chernyi and Andrey Filchenkov is in sunny Jeju, South Korea, to attend the IJCAI International Joint Conferences on Artificial Intelligence Organization and mingle with fellow AI researchers. #AI #research #noeonAI #IJCAI

Noeon Research is in Vienna for [ICML] Int'l Conference on Machine Learning! Our team (German Magai, Andrey Filchenkov, Simeon Ganiushin and Blaine Rogers) is looking forward to engaging with fellow researchers and practitioners to cross-pollinate the latest ideas in #AI. Join us at the conference for a chat and don't forget to check out German's poster on applying LLMs to algebraic topology -https://lnkd.in/gg9RQe8r. #Noeonai #AI #LLM #ML #ICML2024

"To err is human – but AI isn’t one. So, can we build systems that are always correct?" Despite their continuing evolvement, #LLMs still stumble on straightforward questions, often failing dramatically. This issue is rooted in their design - merely scaling up training data or computing power isn't enough to overcome these shortcomings. At #NoeonResearch, we are developing an alternative architecture that consistently provides accurate answers for questions that can be resolved through intellectual effort. Our innovative system can autonomously detect gaps in knowledge and resolve data conflicts, ensuring reliable and precise outcomes. Learn more about our alternative approach at noeon.ai. #noeonai #AI #DeepTech #Innovation #ArtificialIntelligence #MachineLearning #TechInnovation #FutureOfAI #AIResearch

Congratulations to our Research Engineer German Magai on having his paper accepted for the [ICML] Int'l Conference on Machine Learning. In the paper, German and his co-authors, Kirill Brilliantov, Fedor Pavutnitskiy and Dmitrii A. Pasechniuk proposed a new way to generate cycles in simplicial groups using machine learning methods. Reformulating the problem of generating cycles in simplicial groups as the task of generating elements from the intersection of subsets of (a form of) Dyck-n words, they use GPT-2 for sample generation and achieve a substantial and scalable improvement over classical algorithms. This work is a first step towards the automatic generation of nontrivial elements of homotopy groups of spheres and the application of large language models to problems of homotopy topology.  Full paper: https://lnkd.in/ep7YEDYF German and Noeon Research will attend #ICML in Vienna on July 21–27. Catch us at the conference for a chat on interpretable #AI and alternative architectures for reasoning! #ICML2024 #Noeon