Applied GenAI Bites goes Multimodal
Text is only the Beginning
Hey everyone! 👋
Thank you for being part of the Applied Generative AI community!
What a great event we had at AWS! Over 100 people showed up for the event, thank you!
Agenda for Today:
Small Language Models
Create your avatar GitHub Repo
Voice Assistants
Swiss AI Startups that have raised money this year
Swiss Companies in YC
Interesting Resources
Recap of the last Event at AWS offices
Small Language Models, Avatars, Voice Assistants, and AI Videos.
Vittorio from PremAI discussed Small Language Models and their advantages and disadvantages compared to larger LLMs.
Advantages:
Fast: SLMs are quick to process because they require less computational power and memory compared to larger models.
Cheap: Since they demand fewer resources, SLMs are cost-efficient in both the training, deployment, and inference phases.
Customizable: Smaller models can be fine-tuned more easily to specific tasks or domains, allowing for greater flexibility in application.
Great for Privacy: SLMs can be run locally on devices, reducing the need for data to be sent to external servers, thus enhancing privacy.
Disadvantages:
Limited Contextual Understanding: SLMs struggle to grasp complex or nuanced contexts, especially over long conversations or text.
Reduced Accuracy and Performance: They tend to have lower precision in tasks compared to larger models, particularly in complex scenarios.
Limited Creativity and Variability: SLMs often generate more repetitive or less diverse outputs compared to their larger counterparts.
Data Dependency: Their performance heavily relies on the quality and amount of training data, leading to potential bias or gaps.
Pajtim (in the picture on the left) from AWS made all of us feel like superheroes for one night by delving into the nitty-gritty of creating personal AI avatars. Here is the GitHub repository. It's very helpful if you want to learn how to build applications using AWS.
Cost-effective setup of Comfy-UI: Link
Personalized Avatar Stack: Link
We even had a hilarious moment. During an interaction between me and a Cleopatra avatar (whom I had instructed to be “charming and intelligent”), she had an interesting interpretation of her prompt.
By the way, I'd love to bring this technology into schools, offering an alternative learning approach for kids and students who don't like to concentrate on books and are more visual learners. This could benefit people with dyslexia or simply all of us with shorter attention spans following the widespread adoption of social media. If you're in contact with any school principals or teachers, please let me know!
Voice Assistants
I am a strong believer in the imminent and ubiquitous adoption of voice assistants. Inbound and Outbound customer support and sales will change significantly in the next years. At Duenders, we are working on various projects that utilize them. Here are some examples:
AI Receptionist for hotels, restaurants, gyms, and spas. It can answer questions about the business, make bookings, and collect feedback.
Try it out by calling: +41 43 505 22 01. Or listen to an example here:
Interview Analyzer: an AI performs customer or employee feedback interviews and automatically extracts insights from the transcripts, which are then visualized in a dashboard or report.
Personal Assistant: tired of spam calls? We've had requests for an AI that can answer calls, collect information about the caller, reason about the call, and potentially flag it as spam. All the information is then sent by email with a summary so that the user can decide whether to return the call or not.
Try it out here: +41 43 505 22 10. Or listen to an example here:
Swiss AI startup that raised money Year-to-Date
Source: The Week in Swiss Startups newsletter. In bold the ones that were not included in the last issue of Applied AI Bites.
DemoSquare: CHF 1.2M - Pre-seed
Optiverse: CHF 25k - Convertible loan (Founderful Campus)
hoshii: Undisclosed
Lakera: $20M - Series A
Ex nunc: CHF 50k - Loan
DemoSquare: CHF 1.2M - Pre-seed
Neural Concept: $27M - Series B
Jua.ai: $16M - Seed
EthonAI: CHF 15M - Series A
DeepJudge: $10.7M - Seed
Retinai: $6.18M - Series A
Synthara AG: $11.4M - Seed extension ($5.5M equity + $5.9M grants)
Rivia: CHF 3M - Seed
UpCircle: CHF 150k - Grant
identic ai: CHF 100k - Grant
csky.ai: CHF 25k - Convertible Loan
Studyflash: CHF 25k - Convertible Loan
Logmind: Undisclosed - Seed
Nuuro: Undisclosed - Pre-seed
Swiss Companies in YC
Y Combinator is probably the most famous startup accelerator in the world. Thousands of companies apply for each batch, and the probability of getting in is less than 1%. A big thank you to my friend Gianmaria Sbetta for putting this information together. If you're interested in startups in Switzerland, he is THE person to get in touch with.
2024
- Stacksync (YC W24): Real-time and two-way sync between CRMs and Databases
- Stack Auth (YC S24): Open-source auth & user management for developers
- Sonia (YC W24): Mental Healthcare
2023
- nunu.ai: Building the first multimodal agents to play and test games
- sizeless (YC S23): Make ML models reproducible and safe
- Tremor: The UI toolkit for Dashboards
- Rex.fit 🦖 (YC W23): Your personal AI fitness and nutrition coach
- Anarchy Labs: Gitlab for LLM Teams
- Nango (YC W23): A single API for all your integrations
- atla: LLMs to evaluate other LLMs
- SID.ai (YC S23): SID connects Large Language Models (LLMs) to the data they need
- Univerbal: Language learning with a conversational AI Tutor
Interesting Resources
Text-to-SQL is the capability of LLMs to write queries for relational databases. This allows non-technical people to build visualizations and dashboards using natural language. Even if you know how to write SQL, you can appreciate the faster and more natural experience. PremAI has put together a report titled "State of Text-to-SQL". You can read it at this Link.
Anthropic Claude Definitive Meta-Prompt: This prompt can be used in conjunction with your specific task to create a better, more structured, and higher-quality output prompt. It's worth looking at even just to learn what proper prompt engineering looks like. Link.