Applied AI Bites ?? #3

Applied AI Bites ?? #3

Hi GenAI lovers! Mondays may have a rep for being rough, but the content of today will hopefully be more interesting than your start of the week's never-ending inbox. ??Enjoy!

Agenda for Today:

  • Are EU regulations working?
  • 185 GenAI Use Cases list
  • Swiss AI Startups that have raised money this year
  • RAG Advanced Techniques Part 1 (slightly technical)


Are EU regulations working?

OpenAI decided not to release the advanced voice mode (which allows voice-to-voice conversations to be more expressive) in Europe.

Meta unveils its Llama 3.2 models, but the multimodal versions are not available in Europe due to ongoing data privacy concerns. ?????????????? ??????????'?? ???????? ?????? ???? ???????????? ??????????????????.

???????? ???? ??????????????????, as Meta's open-source models are often the fundamental building blocks for developing on-premises, highly secure solutions with strong data privacy protections. ???????? ?????? ???? ???????? ??????!

Here are some more thoughts on the topic: Link

Is the EU sabotaging itself with its own regulations?

EU regulations are getting the opposite output of what they are meant for.



GenAI Use Cases Waterfall ??

Are you still struggling to come up with GenAI Use Cases? Here are 185 from Google and its partners. Pretty much all industries are covered. Link


Swiss AI startups that raised money Year-to-Date ??

Source: The Week in Swiss Startups newsletter. In bold are the ones that were not included in the last issue of Applied AI Bites.


RAG Advanced Techniques: Part 1 ??

When talking about RAG there are only two types of people: those who think it’s easy and those who have tried to put a chatbot or RAG system in production.

What is RAG? Retrieval Augmented Generation is the biggest use case of GenAI so far. It is the fundamental block of building a “ChatGPT over your proprietary data“ systems.

What is the basic implementation? RAG combines retrieval and generation. It uses embeddings (vector representations of text) to find relevant information from a database, then feeds this into a language model to generate informed responses. This process enhances AI outputs with specific, up-to-date knowledge.

Main Challenges: Chunking, Retrieval of correct text chunks, Avoid hallucinations


Advanced Techniques Part 1

Query Rewriting: Make LLMs enrich queries with relevant background information to improve search precision. Example:

Original query: What are the health benefits of meditation?

Enhanced query: Considering recent neuroscience research and long-term clinical studies, what are the proven physiological and psychological health benefits of regular meditation practice?


Step-back Prompting: Make LLMs broaden the query to capture related concepts and ensure comprehensive coverage. Example:

Original query: What are the health benefits of meditation?

Expanded query: What are the overall impacts of mindfulness practices, including meditation, on physical and mental well-being?


Sub-query Decomposition: Make LLMs divide complex questions into interconnected aspects for thorough exploration. Example:

Original query: What are the health benefits of meditation? Faceted queries:

Decomposed query:

  • How does regular meditation affect stress levels and cortisol production?
  • What changes in brain structure and function have been observed in long-term meditators?
  • Can meditation practices influence sleep quality and insomnia symptoms?
  • What are the potential benefits of meditation for managing chronic pain conditions?


HyDE (Hypothetical Document Embedding)

  • Generate a plausible response to the query using the AI model.
  • Convert this hypothetical answer into a vector representation.
  • Find the most similar actual documents in the database using this vector.

The generated hypothetical reply often yields higher similarity scores during retrieval compared to the original question, due to its richer contextual information.


Reranking

Reranking in RAG systems improves the relevance and quality of retrieved documents by reassessing and reordering them. It addresses the limitations of initial retrieval methods, which often use simple similarity metrics, by applying more sophisticated relevance assessments. This ensures the most pertinent information is prioritized for the generation phase, enhancing overall system performance.

* Image Credits: Langchain


There are many more techniques to explore. Did you enjoy this deep dive? If yes, please let me know, and I will include Part 2 in the next issue! And don’t forget to share this article with a friend :)


Talk to you soon!

Ruggiero

Roberto Dal Corso

I help SME business owners achieve predictable revenue growth, leveraging AI-driven strategies and a proven business growth accelerator | keynote speaker | board advisor | time-crunched cyclist.

1 个月

Lineup looks super interesting!

回复
Thomas Kuster

Tech Attorney & Partner at LEXR | AI, SaaS, Deeptech

1 个月

Thanks for sharing! I slightly disagree on the regulation bit though - Big tech probably knows better than anyone else how to handle the extensive EU regulations, and the decision not to offer some of their innovations is in my opinion more of a powerplay/lobbying move than an actual issue with the regulation as such - that's why they usually don't say which regulation exactly prevents them to release something in Europe. Not saying the EU didn't overdo it here, just that big tech isn't the real victim, the additional compliance burden and the legal uncertainty is more harmful to smaller companies and start ups imo.

要查看或添加评论,请登录

社区洞察

其他会员也浏览了