Stay Ahead in AI with Weekly AI Roundup; read and listen on AITechCircle:
Welcome to the weekly AI Newsletter, your go-to source for practical and actionable ideas. I’m here to give you tips that you can immediately apply to your job and business.
Before we start, share this week’s updates with a friend or a colleague:
Today at a Glance:
- Open LLMs in the context of the EU AI Act
- Generative AI Usecase: Personalized Lessons in Education with Large Language Models
- AI Weekly news and updates, Favorite Tip Of The Week & Things to Know
- Open Tech Talk Podcast, the latest episode on Building an AI-Driven Business with Gray Mabry
|
Large Language Models: How open are they?
There is an ongoing debate on the merits of Open-Source Large Language Models versus proprietary models, and why the development in open source is crucial. This debate is rooted in the original development that took place in the industry, championed by the unsung heroes who support the open-source mantra and help drive innovation.
As the AI era approaches, the same questions about open-source LLMs and how they are vital for the industry’s growth come to mind. I found an excellent paper, ‘Rethinking open source generative AI: open-washing and the EU AI Act,’ and spent this weekend going through it.
Generative AI systems often claim to be open, but how genuinely open are they?
This raises the question: what qualifies as open source in generative AI?
The authors of this paper did wonderful work in creating an evidence-based framework that identifies 14 dimensions of openness, including training datasets, scientific and technical documentation, licensing, and access methods.
This framework surveys over 45 generative AI systems covering text and text-to-image applications.
The notion of releasing LLMs as ‘open’ models is questionable itself. Achieving genuine openness in generative AI faces significant challenges. The complexity of current system architectures and training procedures means openness cannot be simply categorized. Additionally, some systems are only nominally open; free access does not necessarily mean they are truly transparent. Furthermore, many models share only their weights, keeping other critical aspects hidden, undermining professional standards and deviating from the open-source movement’s core principles.
Open-Washing
An excellent way described in the paper is with the term “Open-washing.”
“Companies operating in the generative AI space currently appear to be converging on a strategy known as open-washing [60]: collect brownie points for openness without disclosing critical information of training and tuning procedures, thereby largely escaping the scientific scrutiny and legal exposure that would come with full openness.”
All the Large Language Models were evaluated based on the three key categories:
- Availability: Evaluated based on the source code, pretraining data, base weights, fine-tuning data, fine-tuning weights, and licensing under a recognized open-source license.
- Documentation: Assessed by code documentation, architecture details, preprint papers, published peer-reviewed papers, model cards, and datasheets explaining data collection and curation methods.
- Access: Determined by the availability of a downloadable package and an open API.
They assessed 14 characteristics, assigning each a score of open (1 point), partially open (0.5 points), or closed (0 points). For instance, an API requiring user registration was considered partially open.
We are all going through the phase of selecting the LLMs for different use cases to address specific organizational requirements.
It’s worth reviewing it during your evaluation of LLM for your organization’s use cases.
You can find the updated version of the above at Opening up ChatGPT: tracking openness of instruction-tuned LLMs
|
Weekly News & Updates…
Last week’s AI breakthroughs marked another leap forward in the tech revolution.
- JASCO from Meta is a new music-generation model with improved controllability using conditioning inputs like chords or beats. link
- DataComp-LM (DCLM) from Meta is a framework for building and training large language models (LLMs) with diverse datasets. Its 6.9B base model is competitive with Mistral/LLAMA3/Gemma/Qwen2 on most benchmarks. It offers a standardized corpus of over 300T unfiltered tokens from CommonCrawl, effective pretraining recipes based on the open_lm framework, and an extensive suite of over 50 evaluations
- Graph Foundation Model (GraphFM) is trained on 152 datasets with over 7.4 million nodes and 189 million edges spanning diverse domains.
- GPT-4o mini from OpenAI is priced at 15 cents per million input tokens and 60 cents per million output tokens, more than 60% cheaper than GPT-3.5 Turbo.
The Cloud: the backbone of the AI revolution
- From inference to RAG: Choosing CPUs for efficient generative AI application deployments Link
- NVIDIA Unveils Reference Architecture for AI Cloud Providers. Link
- How to build user authentication into your gen AI app-accessing database. Link
|
Gen AI Use Case of the Week:
Generative AI use cases in the Education Sector:
Personalized Lessons in the Education Sector with Large Language Models
Implementing Large Language Models (LLMs) in the education sector to create personalized lessons addresses critical challenges such as diverse learning needs, resource constraints, engagement issues, and scalability.
Business Challenges
- Diverse Learning Needs: Students have varying learning styles, paces, and levels of understanding.
- Resource Constraint: Limited availability of personalized tutoring for each student.
- Engagement Issues: Keeping students engaged with the learning material is challenging.
- Scalability: Providing personalized education at scale is challenging.
AI Solution Description
Large Language Models (LLMs) can create personalized lesson plans tailored to each student’s needs. Here’s how it works:
Collect data on student performance, learning styles, and preferences through assessments, quizzes, and interaction logs.
Use LLMs to generate customized lesson plans, explanations, and practice exercises based on the analyzed data.
Continuously update the personalized content based on ongoing assessments and feedback to ensure the lessons remain relevant and practical.
Expected Impact/Business Outcome
- Revenue: Increased enrollment due to offering personalized learning experiences, leading to higher tuition fees.
- User Experience: Enhanced student engagement and satisfaction through tailored learning materials, leading to better retention rates.
- Operations: Reduced strain on educators by automating the creation of personalized lesson plans, allowing them to focus on more complex tasks.
- Process: Streamlined educational processes with automated, adaptive content delivery.
- Cost: Lower costs associated with hiring additional tutoring staff due to automation.
Required Data Sources
- Student performance records
- Learning style assessments
- Interaction logs from educational platforms
- Feedback and survey data from students and educators
Strategic Fit and Impact
This use case has a high strategic fit and impact due to its potential to transform the education sector by making personalized learning accessible and scalable. It aligns well with the goals of enhancing educational outcomes and operational efficiency.
Rating: High Impact
|
Favorite Tip Of The Week:
Here’s my favorite resource of the week.
- Eureka Labs from Andrej Karpathy, AI+Education company, and the first course will be: , LLM101n. This undergraduate-level class guides the students through training their own AI, similar to a smaller version of the AI Teaching Assistant.
Potential of AI
- Kyutai Labs just released Moshi, a real-time native multimodal foundation model. Try out here
|
Things to Know…
OpenAI’s five-step plan for achieving Artificial General Intelligence (AGI) offers a clear framework for monitoring AI development. This classification system shared with employees, outlines five levels of AI capability.
Conversational AI (Level 1)
Problem-Solving Virtuosos (Level 2)
Autonomous Agents (Level 3)
Innovators and Creators (Level 4)
Organizational Equivalents (Level 5)
|
The Opportunity…
Podcast:
- This week’s Open Tech Talks episode 140 is “Developing AI Products From Tech Stack to User Feedback with Jason Agouris” Jason’s extensive experience in systems integration across retail, fintech, wholesale, and supply chain logistics makes him our go-to guru for data integration and strategy.
Apple | Spotify | Youtube
Courses to attend:
- Pretraining LLMs from DeepLearning AI. This course explores the creation of large language models (LLMs) like Llama, Grok, and Solar using a technique called pretraining, which is the first step of training an LLM
- Generative AI Fundamentals Specialization from IBM, covers the fundamental concepts, capabilities, models, tools, applications, and platforms of generative AI foundation models.
Events:
Tech and Tools…
- STORM: Synthesis of Topic Outlines through Retrieval and Multi-perspective Questions Asking
- LLaMA-Factory – A WebUI for Efficient Fine-Tuning of 100+ LLMs
|
Data Sets…
- Multi-Domain Wizard-of-Oz dataset (MultiWOZ), a fully-labeled collection of human-human written conversations spanning multiple domains and topics. At a size of 10k dialogues
- Discrete Reasoning Over the Content of Paragraphs (DROP). It is a QA dataset that tests a comprehensive understanding of paragraphs. In this crowdsourced, adversarially-created, 96k question-answering benchmark, a system must resolve multiple references in a question, map them onto a paragraph, and perform discrete operations over them
Other Technology News
Want to stay on the cutting edge?
Here’s what else is happening in Information Technology you should know about:
- Figma explains how its AI tool ripped off Apple’s design, as reported by TheVerge. Generative AI tools started making designs similar to Apple’s Weather app.
|
Earlier week’s Post:
That’s it!
As always, thanks for reading.
Hit reply and let me know what you found most helpful this week – I’d love to hear from you!
Until next week,
Kashif Manzoor
|
The opinions expressed here are solely my conjecture based on experience, practice, and observation. They do not represent the thoughts, intentions, plans, or strategies of my current or previous employers or their clients/customers. The objective of this newsletter is to share and learn with the community.
|
|
