AI2 Incubator’s Post

At the beginning of this year, I wrote in Insight #13: """ 2024 Prediction: VoiceGPTs We wrap up Insight #13 with our prediction for 2024. Similar to many other 2024 predictions, we anticipate multimodal models to take center stage. We are particularly excited about models that combine text and speech modalities, enabling seamless end-to-end conversations that are voice-based. This is in contrast to the current pipelined approach of sandwiching an LLM with a pair of speech-to-text and text-to-speech models that results in highly stilted, walkie-talkie-like experience. Multimodal text and speech models, which we refer to as VoiceGPTs, will elevate the popular ChatGPT experience beyond the confines of the keyboard. Imagine having a natural conversation about any topic with a VoiceGPT on your Alexa, Siri, or Home device. This is a highly non-trivial technical challenge. We will only see a preview of such technology in 2024. """ Fast forward 9 months, I was proven to be a bit too cautious. First, OpenAI announced in May GPT-4o, with scifi-like demos evocative of the movie Her (and drawing the ire of Scarlett Johansson). GPT-4o's voice mode was rolled out to users earlier this month. OpenAI continues to show the way, leaving the rest of the industry (Anthropic, Google, Meta, etc.) scrambling to catch up. The company that is closest to catching up here is however a French AI research lab "with a $330 million budget that will make everything open source". It is called Kyutai. Last week they shared Moshi (model, weights for Moshi and its Mimi codec, streaming inference code in Pytorch, Rust and MLX, and a fantastic technical report). Amazing! Moshi's technical report is our pick for past week's spotlight paper at Harmonious. https://lnkd.in/gFEVEdTq #ai2incubator #harmonious

Harmonious' spotlight paper this week: OLMoE: Open Mixture-of-Experts Language Models Authors: Allen Institute for AI; Contextual AI; University of Washington; Princeton University    This paper presents OLMoE, an innovative language model leveraging a sparse Mixture-of-Experts architecture, which achieves remarkable efficiency and performance with its 7 billion parameters. I found the emphasis on key design choices and their detailed analysis of MoE training particularly insightful. The open-source nature of their work fosters transparency and collaboration in the AI community. However, the high computational resources required for pretraining may limit accessibility for many academic institutions. Lastly, I am curious about whether the outcomes observed in smaller models will hold true in significantly larger models. #harmoniou #ai2incubator

Harmonious' weekly paper roundup for the week of August 26, 2024. The reviewed papers collectively delve into various advancements in AI models, particularly focusing on multimodal, vision-language, and inference strategies. Several papers explore the enhancement of Large Language Models (LLMs) through innovative techniques such as improved inference patterns for long contexts, the utilization of mixed encoders, and energy-efficient on-device processing (WiM, Eagle, Dolphin). Another recurring theme is multimodality, with in-depth studies on optimizing LLMs for cross-modal alignment and real-time interactions in complex environments (Law of Vision Representation, GameNGen, CogVLM2). Further contributions include advancements in text-to-image diffusion models, audio language modeling, and AI-generated content in music, reflecting the expanding scope of AI applications (SwiftBrush v2, WavTokenizer, Foundation Models for Music). The practical impact of these models is underscored by initiatives to enhance the functionality and accessibility of benchmarks and operational pipelines, ensuring robust performance in real-world scenarios (SWE-bench-java, LlamaDuo, MME-RealWorld). https://lnkd.in/gw7za8-X #ai2incubator #harmonious

About 3 years ago (Fall 2021) I wrote about LLMs and mused about the concept of task-centric AI as an emerging addition to Andrew Ng's initiative on data-centric AI: "What's with the brouhaha around LLMs? Learning efficiency! Below is the famous GPT-3 graph that got everyone's attention ... Instead of building 10 models with 1,000 labels per, we could build 1,000 models with 10 labels per ... In the task-centric world, LLMs could open up the opportunity to help less technical folks build and use AI models without relying on an expensive data science team. No-code AI, powered by the XXL transformers near you? Scale.ai and Snorkel.ai are the poster-child unicorns of the data-centric AI world. Who will emerge as the representatives for the LLM task-centric world?" While the term task-centric AI did not catch on—GenAI is the chosen one (I am neither Andrew Ng nor marketing expert), it essentially describes the AI world we have today. Is there something as transformative lurking on the horizon? I have not seen anything like it following recent AI research (and occasionally sharing thoughts on harmonious.ai). There will be one for sure, and let's be on the lookout. https://lnkd.in/gtQcm-cJ PS: The Yoodli team is on a tear. Go Esha Joshi & Varun Puri! #ai2incubator #startup #llm #genai #task_centric_AI

Seattle Tech Week is Jul 29 to Aug 2 and we're hosting the party of the year! Are you an AI founder? AI investor? AI researcher? AI professor? RSVP now: https://lu.ma/ai2bbq Come hang with 700 AI researchers, professors, entrepreneurs, investors, engineers, and more! Celebrate the best of AI in the PNW with live music, startup science fair, cold beer and BBQ sliders w/ veggie options too. Musical guest: Steve Hall (https://lnkd.in/gQDCdkQA) NOTE: Registration required. Tickets will be checked at the door. Vendors/recruiters—please be respectful. This isn't the place to hustle. 😊

This is the 10th edition of Harmonious’ weekly paper roundup series. This past week I did not find any paper that merits the spotlight designation. Thus I am experimenting with giving a brief overview of papers that are organized around the following topics. - LLM applications: agents, chatbots, RAG, document understanding, coding, and others. - LLM prompting: techniques such as CoT to help us get the most out of LLMs. - Multimodal LLMs. This is an area where many folks expect a lot of advances in the next wave. - Synthetic data and other novel ways to generate training data. Despite the rise of LLMs and zero/few-shot learning, the data bottleneck is still present. It’s helpful to find creative ways to get data, not only for fine tuning but also for prior generation ML approaches. - Benchmarks and evaluations: we need to understand the strengths and limitations of LLMs. - LLM fine tuning/many shot learning. This is an important option wherever prompting is not sufficient. - Context: topics such as context length and limits, effective use of context, context compression, etc. - LLM efficiency, primarily for fine tuning and inference. This is obviously important for real world deployment. - LLM internals: how they work. - LLM frontier: what’s the next big leap beyond transformers? State-space models? Self-evolution? - LLM announcements: e.g. Llama, Phi, etc. I created this taxonomy based on reading a few hundred papers for the first 9 editions of the weekly paper roundup series. I also roughly order the topics in the order of relevance to practitioners (obvious caveat: this is highly subjective). I may adjust this taxonomy if necessary. Not every topic will have papers for a given week. Read more about the papers from the week of April 22 here: https://lnkd.in/gVWK2kgh Sign up at https://harmonious.ai/ to get a weekly update delivered to your inbox. #harmonious.ai #ai2incubator