Symbl.ai

Symbl.ai’s Weekly AI Roundup is here to get you up to speed on all things GenAI🚀!  Key Highlights: Meta launched and open sourced Llama 3.1. Mistral AI announced Mistral large 2. Stability AI published a research paper on the text to audio model. Coalition for Secure AI (CoSAI) formed by major tech companies to address AI security. OpenAI’s introduced plans for a new “prototype” search engine, SearchGPT. The LLM Alignment event by Data Science Dojo is coming up and ICML 2024 is happening right now! Product & Research: 👉 Llama 3.1: Meta announced Llama 3.1, the open source AI model. The instruction-tuned model is available in three versions (3) - 8B, 70B and 405B and supports a context window of 128K. Llama 3.1 70B outperforms GPT 3.5 T on MMLU chat & HumanEval, and Llama 3.1 405B outperforms GPT 4o on GSM8K and ARC Challenge. 👉 Mistral Large 2:  Mistral AI announced Mistral Large 2, an extension of the Mistral Large family with more advanced capabilities. The model has a 128k context window size, offers multilingual support,  and supports multiple coding languages. 👉 Stable Audio Open Research Paper: Stability AI published a research paper that describes the architecture and training process of its text-to-audio model. Notably, there are three components. 1) an autoencoder for waveform compression. 2)  a T5-based text embedding for text conditioning. 3) a transformer-based diffusion model (DiT) operating in the autoencoder's latent space. The model can generate high-quality stereo audio at 44.1kHz from text prompts, and can synthesize realistic sounds and field recordings. News: 👉 Coalition for Secure AI (CoSAI): Google, OpenAI, Microsoft, Amazon, NVIDIA, Intel Corporation, and others are forming the coalition to address AI security. Aim is to address a “fragmented landscape of AI security” by providing access to open-source methodologies, frameworks, and tools.  👉  SearchGPT: OpenAI’s new “prototype” search engine that will help users find what they are looking for by using generative AI to gather links and provide answers in a conversational tone. OpenAI has engaged partners for feedback. Conferences & Event: 👉 [ICML] Int'l Conference on Machine Learning: The conference is taking place from July 21st to 27th at Messe Wien Exhibition Congress Center, Vienna, Austria. ICML is the premier gathering of professionals dedicated to the advancement of ML and is globally renowned for presenting and publishing cutting-edge research. For ICML 2024 live updates and highlights, see the link in comments. 👉 LLM Alignment - Techniques for Building Human-Aligned AI (by Data Science Dojo): Attend session on 11th September to understand how LLMs are evolving to align with human values. The session will cover popular fine tuning and optimization techniques and discuss ideas for real-world applications in customer service and healthcare. Did you come across other exciting developments in AI this week? Let us know in the comments ⬇️

Symbl.ai’s Weekly AI Roundup is here to get you up to speed on all things GenAI🚀!  Key highlights: Mistral AI released Codestral Mamba - specialized in code generation. Hugging Face launched a family of small language models - SmolLM and the training corpus on which they are trained. GPT 4o-mini released by OpenAI. Microsoft gives up the OpenAI board seat. The EU’s AI act published in bloc’s Official Journal. Hybrid Chatbots session by Data Science Dojo is around the corner.  Research & Product launch: 👉 Codestral Mamba: Mistral AI launched Codestral Mamba, a Mamba2 language model specialized in code generation. The 7B parameter model outperforms other similar size models on CruxE and HumanEval Javascript benchmarks. 👉 SmolLM: Hugging Face launched SmolLM, state-of-the-art small language models available in three sizes: 135M, 360M, and 1.7B parameters. Hugging Face has also released SmolLM-Corpus, the training corpus on which models are trained. Their evaluations demonstrate that SmolLM models outperform other models in their size categories across a diverse set of benchmarks, testing common sense reasoning and world knowledge. 👉 GPT 4o-mini: This model supports text and vision inputs, with future plans for multimedia support, a context window of 128K tokens, and up to 16K output tokens per request. GPT-4o mini outperforms other small models at reasoning tasks involving both text and vision, scoring 82.0% on MMLU. It also excels in mathematical reasoning and coding tasks and scored 87% on MGSM, outperforming other small models.    News: 👉 Microsoft gives up observer seat on OpenAI board (article by Ina Fried, Axios): On Tuesday evening, Microsoft announced that it is relinquishing its observer seat on OpenAI's board, citing "significant progress" over the past 8 months and expressing confidence in the company's direction. By relinquishing its observer seat, Microsoft eliminates a potential issue for antitrust regulators who might have concerns about the company's significant involvement with OpenAI. 👉 EU’s AI Act gets published in bloc’s Official Journal, starting clock on legal deadlines (article by Natasha Lomas, TechCrunch): The EU AI Act has been published in the Official Journal. The law will come into force on August 1, 2024, with full applicability by mid-2026. Banned AI use cases include social credit scoring, compiling facial recognition databases, and real-time remote biometrics by law enforcement in public places, except during search for missing or abducted persons. Events & Conferences: 👉 Hybrid Chatbots - Merging LLMs and Classification Models (by Data Science Dojo): Attend session on 21st August to explore a hybrid approach to conversational bots, combining classification models and LLMs. Learn strategies for faster deployment, cost optimization, and real-world applications in customer service and healthcare. Did you come across other exciting developments in AI this week? Let us know in the comments ⬇️

Are your AI bots causing damage to your CX? A survey showed that 66% of 35K chatbot interactions received a 1/5 CSAT rating. The rise of AI voice agents in frontline interactions has highlighted risks that impact CX and compliance: 👉 Non-compliance: collecting excess data than needed or improper handling of sensitive data 👉 Lack of humanness: AI's inability to mimic human interactions 👉 AI hallucinations: AI providing incorrect advice to customers With Call Score API, the risks can be mitigated by having a human in the loop to define a call evaluation framework. For e.g., you can define a scorecard with following criteria: ✅ Humanness: ‘Did AI maintained the conversational flow? and ’ Did AI demonstrate the context understanding to correctly answer customer queries’ ✅ Data collection: ‘Did AI only collected necessary customer details needed to address the issue’ & ‘Did the AI handle sensitive information as per relevant regulations’ Learn more about Call Score API - https://lnkd.in/g8jBn_Gm Try the API today. See the link in comments on how to get started.

Can LLMs detect deception? Does multimodal information from conversations help them with this challenging task? This is one of the questions that our summer intern, Benjamin R. has been exploring in his work at Symbl.ai. We are looking at the general problem of multimodal feature transfer in LLMs, and how to utilize information from data across different modalities to improve the output of these models. Benjamin's work builds on and strengthens our collaboration with Prof. Larry Heck and the AI Virtual Assistant (AVA) Lab at Georgia Institute of Technology. The AVA Lab is dedicated to researching next-generation virtual assistants, challenging assumptions about human-computer interaction design, and multimodal interactions. Their work explores a wide range of conversations, from task-oriented to open-domain chit-chat, integrating explicit and implicit commonsense knowledge with higher-level inference and reasoning. Read more about Benjamin's work here: https://lnkd.in/gGDKMEfR This work is currently under review, and we will provide a link to the full paper as well as extended results shortly. Reach out to us directly if you are interested in this work! 

Human led call evaluations are time and cost intensive! Research shows only 1 to 3% of business conversations are covered in manual QA, with subjective and error-prone assessments.    Our applied generative AI product: Call Score API ensures 100% coverage and reduces assessment time from hours to minutes by generating scores with qualitative feedback at scale. For example, in customer support calls, QA analysts traditionally listen for empathy. With our API, you can set criteria with questions and their priorities: 👉 “Did the agent listen actively and show understanding of the customer's issue?” (High) 👉 “Did the agent provide clear and helpful solutions?” (High) 👉 “Did the agent ensure the customer felt heard and valued?” (High) Call Score API will generate scores using that criteria, enabling targeted coaching for customer facing teams and ultimately improving customer experience. You can use this API to expand into custom criterias and scores like problem-solving, communication skills, adherence to company policies and more.         Automate call scoring by creating a free platform account on Symbl.ai. Learn more - https://lnkd.in/g8jBn_Gm

Personalize call scoring with Call Score API’s new features - ‘Custom Criteria’ & ‘Scorecard’ 🎉 Call Score API is a Gen AI powered, low-code API that evaluates call quality and participant performance at scale - not only of humans but also AI agents. You can use the API to automate scoring of calls across real time and async channels for audio, video and transcript data and build coaching plans, improving CX. Read more here on how to get started with your calls or transcripts: https://lnkd.in/gr3fMxri ✅ Custom Criteria: Define your own criteria with a checklist of questions and priority levels for more granular evaluations. ✅ Scorecard: Combine various criteria (either custom, managed criteria, or a mix of both) to comprehensively evaluate conversations. What does that mean? Unparalleled customizability and programmability to AI-assisted teams across multitude of domains and verticals: ✅ Sales: Ensure adherence to specific sales methodologies like BANT and MEDDIC for quality assurance and drive targeted coaching to improve conversion rate. ✅ Recruitment: Measure biases, empathy, and job description alignment to Improve candidate experience. ✅ Customer Service: Prioritize traits such as problem solving and empathy to improve customer satisfaction (CSAT) and enhance customer loyalty.    Build your own call scores by creating a free platform account on Symbl.ai and reach out to our Sales team for a personalized demo. Links in comments.

Symbl.ai’s Weekly AI Roundup is here to get you up to speed on all things GenAI🚀!  Key highlights: Meta researchers introduce LLMs with fewer than a billion parameters for mobile deployment. Google DeepMind researchers introduce PEER, a layer design that utilizes product key technique for sparse retrieval from a million experts. Anole - an open, autoregressive and native multimodal models for interleaved image-text generation. Elon Musk announces Grok 2 & 3. A security flaw in the OpenAI ChatGPT Mac app has been exposed. AI Engineering Workshop by AI User Group is coming up!  Product & Research: 👉 MobileLLM: Meta research focuses on designing LLMs with fewer than a billion parameters for on-device applications, leveraging deep and thin architectures, coupled with embedding sharing and grouped-query attention mechanisms. The resulting models - MobileLLM -  exhibit substantial advancements in zero-shot commonsense reasoning, question answering, and reading comprehension tasks compared to previous SoTA methods. 👉 Mixture of A Million Experts: Google DeepMind researchers introduce PEER (parameter efficient expert retrieval) architecture, a fine-grained Mixture-of-Experts (MoE) design that decomposes wide dense feedforward layers into numerous tiny experts. Notably, it uses a learned index structure to efficiently route to over a million experts. Experiments on language modeling tasks demonstrate that PEER layers outperform dense FFWs and coarse-grained MoEs in terms of performance-compute trade-off. 👉 Anole: an open-source, autoregressive, natively trained large multimodal model for interleaved image-text generation. It builds on Chameleon and uses a fine-tuning process with a curated dataset of ~6K images. It demonstrates the ability to follow nuanced instructions to produce high quality images and interleaved text-image content. News: 👉 Grok 2 & 3 (article by Emily Price, PCMag): Elon Musk confirms the launch of Grok-2 in August. He claimed that it would “exceed current AI on all metrics." He also tweeted that Grok 3 will be released at the end of the year "after training on 100k H100s". 👉 OpenAI’s ChatGPT Mac app was storing conversations in plain text (article by Jay Peters, The Verge): The concerning security issue meant that if a bad actor or malicious app had access to your machine, they could easily read your conversations with ChatGPT and the data contained within them - demonstrated by Pedro José Pereira Vieito. After recognizing the issue, the company announced a new version which encrypts these conversations. Events & Conferences: 👉 AI Engineering Workshop (hosted by AI User Group): Register for the event on July 18, 2024 from 9:00 am - 6:00 pm. Learn to build open-source AI with Hugging Face and to build AI agents capable of reflection, tool use, memory, and executing tasks. Did you come across other exciting developments in AI this week? Let us know in the comments ⬇️

We are thrilled to announce that our CEO, Surbhi Rathore, has been nominated for the VentureBeat Women in AI Awards 2024 in the AI Entrepreneurship category 🎉! It is a proud moment for us as a women-led business to be nominated for this award alongside other incredible leaders like Allie K. Miller, Erin Rapacki, Nicole Eagan, and others. See the link in comments for complete list. The award honors women who have started companies demonstrating significant promise in the AI space, considering factors such as business traction, technological innovation, and impact. Surbhi's nomination is a testament to her leadership under which Symbl.ai has propelled real-time voice AI for front line teams like sales and customer support. Developers and businesses have improved customer experience, human productivity, and more at a reduced cost with our specialized understanding and generative models. As we eagerly await the awards ceremony on July 10 in the VB Transform event in San Francisco, join us in congratulating Surbhi Rathore for this amazing feat. If you are attending the 3 day conference, let us know in the comments and we would love to meet you there! Read the announcement here: https://lnkd.in/dHH79Fhx.   #WomenInAI #VBTransform2024 #Leadership #AIEntrepreneurship

Symbl.ai’s Weekly AI Roundup is here to get you up to speed on all things GenAI🚀!  Key highlights: Google launched Gemma 2 in 9B and 27B parameter sizes and Context Caching feature for Gemini API to optimize cost by token reuse. Hugging Face introduced BigCodeBench that evaluates LLMs on programming tasks. The Center for Investigative Reporting sued OpenAI and Microsoft for allegedly using their content without permission. VB Transform 2024 conference and AI for product manager event are around the corner. Product & Research: 👉 Google launched Gemma 2: Gemma 2 is available in both 9 billion and 27 billion parameter sizes, offering higher performance and greater inference efficiency compared to its predecessor. The 9B Gemma 2 leads its class, outperforming Llama 3 8B and other models in its category. 👉 BigCodeBench: Hugging Face launches BigCodeBench, which evaluates LLMs on practical and challenging programming tasks. Notably, GPT-4o and DeepSeek-Coder-V2-Instruct have the highest scores and lead the score board. 👉 Context caching: Allows caching input tokens and then referencing these for subsequent requests. That makes it useful for use cases such as Chatbots with extensive system instructions or repetitive analysis of lengthy video files. Using cached tokens instead of repeatedly passing the same corpus reduces costs at certain volumes. News: 👉 OpenAI, Microsoft sued by Center for Investigative Reporting (an article by Hayden Field, CNBC): The Center for Investigative Reporting (CIR), the oldest nonprofit newsroom in the U.S., sued OpenAI and Microsoft for copyright infringement. The lawsuit alleges OpenAI used CIR's content without permission or compensation. Similar suits have been filed by The New York Times, Chicago Tribune, and New York Daily News. Conferences & Events: 👉 VB Transform 2024 by VentureBeat: Register for the three day conference from July 9 -11 in San Francisco, CA. Learn how Graph-based RAG is transforming Enterprise AI, how AI requires a new kind of OS, blending open-source LLMs with proprietary models, and more. 👉 AI for product managers: Attend the webinar by Data Science Dojo on July 24th to explore how AI is transforming the domain of Product Management by enhancing efficiencies and reshaping strategies. Did you come across other exciting developments in AI this week? Let us know in the comments ⬇️

We are excited to announce that our CEO, Surbhi Rathore, will be joining the panel discussion "Breaking Barriers: Strategies for Women to Thrive in AI Careers - From Classroom to Boardroom" at the BMM 2024 conference today at the San Jose Convention Center, California. The BMM conference is a vibrant gathering of the global Maharashtrian community, celebrating culture, language, and identity while embracing technological advancements. Hear from leaders, including Sonali Sambhus (Board Advisor at Worth AI ), Shilpa Kolhatkar (Director of Worldwide AI Initiatives at NVIDIA), and Sara Thomas Deshpande (General Partner at Maven Ventures), as they share their experiences and the challenges they overcame to make a significant impact. If you are coming to the conference, don’t miss this opportunity to learn and be inspired by these leaders. We look forward to meeting you there! #WomenInAI #WomenInLeadership #Inspiration #Empowerment #WomenInBusiness  #bconnect #bmm2024 #kaaybay #baymarathi #businessconference #SanJose