Don't get confused with "open source" for "open weights" in GenAI

Last week, I saw a post on LinkedIn by Yann LeCun on the open-source efforts from Meta and Mark Zuckerberg being champion of open-source AI. But the definition of Open-source in AI, especially on LLM can be misleading in many ways because Meta does not release Llama as open-source but only the weights of the model.

To understand this in detail, let’s dive a bit deeper into the open-source and closed-source AI.

The distinction between open-source AI and closed-source AI has become a critical point of discussion in the rapidly advancing field of artificial intelligence. This differentiation holds significant implications for transparency, collaboration, and innovation within the AI community. Moreover, confusion often arises regarding the terms "open-source" and "open weights," particularly in the context of large language models (LLMs). 

Defining open-source AI and closed-source AI

Open-source AI pertains to AI models, algorithms, and software whose source code is made publicly accessible. This openness allows individuals to view, modify, and distribute the code. The foundational principles of open-source AI encompass transparency, collaboration, and community-driven innovation. Examples of notable open-source AI projects include TensorFlow by Google and PyTorch by Facebook, which offer extensive libraries and tools for AI development.

Closed-source AI involves proprietary models and software where the source code is not shared with the public. Companies that develop closed-source AI maintain exclusive control over their codebase, limiting access to the internal workings of their algorithms. This approach is often driven by commercial interests, intellectual property protection, and competitive advantage. Prominent examples of closed-source AI include certain models developed by OpenAI, such as GPT-4o, Gemini, Claude etc which have restricted access to their underlying code.

Avoiding the Confusion Between "Open Source" and "Open Weights"

Open weights refer to the practice of making the trained parameters of a model available to the public. While open weights enable others to utilise a pre-trained model for various applications, they do not provide insight into the underlying architecture, data, or training process used to develop the model. Essentially, users can leverage the model’s capabilities without comprehending its inner workings or making modifications.

In contrast, Open-source involves not only sharing the weights but also the entire codebase, including the training scripts, data preprocessing methods, and any other components necessary to fully understand and reproduce the model. Open-source projects encourage a deeper level of engagement, allowing researchers and developers to build upon existing work, identify potential biases, and enhance the model's performance through community-driven efforts.

The Significance of Clear Definitions: Open-source AI

1. Transparency and Trust: Open-source AI promotes transparency, allowing the community to scrutinise and comprehend the decision-making processes of AI models. This transparency is vital for building trust, especially in applications with significant societal impact, such as healthcare, finance, and criminal justice.

2. Innovation and Collaboration: Open-source AI fosters a collaborative environment where researchers and developers can collectively advance the field. By sharing knowledge and resources, the community can address complex challenges more effectively and drive innovation at a faster pace.

3. Ethical Considerations: Access to the full source code allows for thorough examination of potential biases and ethical implications embedded in AI models. This scrutiny is essential for developing fair and unbiased AI systems that align with societal values.

4. Educational Value: Open-source AI serves as an invaluable educational resource. Students, researchers, and practitioners can learn from existing models, understand best practices, and experiment with modifications to enhance their skills and knowledge.

5. Commercial Impact: For businesses, the choice between open-source and closed-source AI has significant implications. Open-source AI can reduce development costs and accelerate time-to-market, while closed-source AI might provide a competitive edge through proprietary technology.

The Significance of Clear Definitions: Open-weights (or just “weights”)

1. Limited Transparency: Open weights allow the use of pre-trained models but do not offer insight into how the model was created or trained, limiting full transparency.

2. Utilization without Modification: Users can leverage the capabilities of the model with open weights but cannot modify or fully understand the underlying mechanisms without additional information.

3. Partial Collaboration: While open weights can foster some level of collaboration by enabling others to build applications using the pre-trained model, they do not support the deep collaboration seen in fully open-source projects.

4. Educational Constraints: Open weights can be a learning resource, but without access to the entire codebase and training data, they provide a limited educational experience compared to fully open-source AI.

5. Ethical Review Limitations: Open weights do not allow for a thorough examination of potential biases or ethical issues in the model, as the full training context is not available.

6. Commercial Use: Businesses can use open weights to integrate advanced capabilities into their products quickly, but they may face challenges in adapting the model without full access to the source code and training process.

Conclusion

The distinction between open-source AI and closed-source AI, along with the clarification of open weights, is a fundamental aspect of contemporary AI. As the AI community continues to grow and evolve, it is imperative to foster a clear understanding of these concepts to promote transparency, innovation, and ethical development. Most importantly, it is essential to avoid confusing "open source" with "open weights" (or just weights) in AI, as this misunderstanding can hinder collaboration and progress. By embracing open-source principles, the AI community can harness the collective intelligence of researchers and developers worldwide, advancing the field in a manner that benefits society as a whole.

Kind regards,

Shen Pandi

TheGen.AI News

Tektonic AI Unveils $10M Seed Funding to Enhance Automation with GenAI

A few years back, robotic process automation (RPA) was a major trend in enterprise software, aiming to automate repetitive business tasks. However, it seems it didn't fully meet expectations. Now, the emergence of generative AI might be the breakthrough needed for these systems.

Seattle's startup, Tektonic, which blends generative AI with traditional symbolic methods, has just emerged from stealth mode and announced a $10M seed funding round from Point72 Ventures and Madrona Ventures.

Tektonic's platform allows users to employ natural language to facilitate workflow automation, particularly in complex areas like quotes and renewals that involve numerous manual steps and vary significantly between businesses.

Co-founded by Nic Surpatanu, who has had leadership roles at companies like Tanium, UiPath, and Microsoft, Tektonic aims to enhance automation by integrating generative AI, which provides adaptability and a better grasp of user intent—capabilities that RPA lacked. This adaptability is crucial, as RPA systems can break with any significant interface change, requiring ongoing maintenance.

Surpatanu believes that while generative AI can't yet support fully autonomous systems, it can significantly extend the range of automatable scenarios, potentially increasing process automation from 50% to 80%.

Technically, Tektonic uses foundation models and open models to extract entities and manage lower-level actions, aiming to reduce manual work and allow sales reps to focus more on customer interactions by partnering with AI agents that understand and manage their processes.

According to Madrona Ventures' Steve Singh, Ted Kummert, and Palak Goel, generative AI's ability to navigate across data silos and coordinate tasks can revolutionize process automation.

Although still in the early stages, Tektonic is working with design partners to refine its system. Surpatanu envisions the company evolving into a SaaS model within three to five years, connecting directly to business APIs. Currently, implementing Tektonic requires setting up a container system within a business's virtual private cloud.

TCS Unveils WisdomNext: A Pioneering GenAI Aggregation Platform for Industry Transformation

Tata Consultancy Services (TCS) (BSE: 532540, NSE: TCS), a global leader in IT services, consulting, and business solutions, has introduced TCS AI WisdomNext™. This new platform consolidates various Generative Artificial Intelligence (GenAI) services into a unified interface, facilitating the large-scale adoption of advanced technologies by organizations at reduced costs and within regulatory limits. The platform is crafted to eliminate obstacles that prevent customers from developing and deploying business solutions, and it supports real-time experimentation with different vendor, internal, and open-source Large Language Models (LLMs).

AI and GenAI are increasingly integral across business value chains, yet selecting and experimenting with the appropriate foundational models remains challenging for solution designers globally. These foundational models are continually changing, each offering unique capabilities in terms of usage, cost, and effectiveness. According to TCS’ AI for Business Study, which explores AI's impact on enterprises, business leaders recognize AI’s benefits but often struggle with its implementation. TCS AI WisdomNext aims to simplify this process, helping businesses select suitable models and streamline the creation of new business solutions using GenAI tools. It also facilitates the reuse of existing components to speed up design processes.

Siva Ganesan, Head of the AI.Cloud Unit at TCS, commented, "TCS AI WisdomNext enables our clients to fully leverage GenAI, enhancing their data utilization, fostering business innovation and efficiency, and providing a competitive advantage. Customers are impressed with the platform’s ability to navigate the complex and rapidly evolving AI market and quickly develop cutting-edge solutions. We're addressing real business challenges and assisting our clients in redefining the use of GenAI, with many finding the rapid adoption and tangible business outcomes particularly exciting."

During its initial testing phase, TCS utilized this robust tool with several major clients to deliver significant value and develop prototypes. These include speeding up sales processes for a US outdoor advertising company with real-time inventory updates and integrated mapping for quote generation; boosting productivity and efficiency in application migration and modernization for a top American insurance company; and improving customer interactions with a smart mortgage-assistant for a prominent UK bank.

Scott Kessler, Executive Vice President and Chief Information Officer at Northeast Shared Services, noted, "GenAI offers an opportunity to significantly enhance our knowledge capital on multiple fronts. With the TCS AI WisdomNext™ platform, we can enrich our enterprise knowledge, seamlessly integrating data and insights to improve efficiency, drive innovation, and focus more on our customers within our grocery business, adding value at every stage."

R “Ray” Wang, CEO and Principal Analyst at Constellation Research, added, “Clients are looking for a GenAI platform that combines pre-built industry blueprints, state-of-the-art technologies, and orchestration capabilities to support successful future adoption and innovation in GenAI. Service providers that invest thoughtfully in GenAI, providing transformational support with a focus on security and ethical practices, will be the trusted partners that businesses rely on to fulfill the promises of GenAI. These AI-first partners are poised to deliver a versatile, multi-model, multi-modal, and multi-cloud platform."

Galileo Launches Specialized Evaluation Models to Enhance Enterprise GenAI

Galileo, a developer of generative artificial intelligence (GenAI), has launched a new series of evaluation foundation models (EFMs) called Galileo Luna EFMs. These models are designed to streamline the process of integrating trustworthy AI into business operations, according to a press release from the company on Thursday (June 6).

Vikram Chatterji, co-founder and CEO of Galileo, emphasized the necessity for enterprises to swiftly and accurately evaluate AI responses for issues such as hallucinations, toxicity, and security risks. He noted that traditional methods like human evaluation or using large language models (LLMs) are often too costly and slow. The Luna EFMs, therefore, aim to address these challenges by being specifically tailored for evaluation tasks, making them more efficient in terms of cost, speed, and accuracy.

The release also mentioned that these models have been integrated across all Galileo platforms to intercept harmful inputs and enhance both system security and operational efficiency. Notably, major companies, including Fortune 50 consumer packaged goods brands and Fortune 10 banks, are already using these EFMs to manage millions of GenAI application queries each month.

Alex Klug, head of product, data science, and AI at HP, echoed the sentiment about the limitations of existing evaluation methods and praised the Luna models for tackling the primary concerns of cost, latency, and accuracy in enterprise AI evaluations.

In addition to this launch, Galileo had previously announced in February the release of a new solution focused on retrieval augmented generation (RAG) and agent analytics, aimed at fostering the development of trustworthy AI solutions.

A report by PYMNTS Intelligence in collaboration with AI-ID, titled "Banking on AI: Financial Services Sector Harnesses Generative AI for Security and Service," highlights that while GenAI is transforming sectors like finance and banking by improving customer interactions and risk models, it also brings challenges such as data security and decision-making risks. Consequently, regulators and stakeholders are increasingly focusing on issues like model explainability to mitigate the risks associated with GenAI.

TheOpen.AI News

Stability AI Unveils 'Stable Audio Open' for Creative Audio Production

Recently, Stability AI, a leader in generative AI technologies, has unveiled an open-source AI text-to-audio model named “Stable Audio Open,” targeting creative individuals. This new tool allows users to convert text prompts into up to 47 seconds of high-quality audio. It's particularly adept at generating instrument riffs, ambient sounds, drum beats, and Foley recordings.

Stable Audio Open is released under the Stability AI Non-Commercial Research Community License, which restricts its use to non-commercial purposes only. Additionally, users must agree to Stability AI's privacy policy to use the model.

The platform offers customization by allowing users to fine-tune the model with their own audio data. For example, a drummer could refine the model using their personal drumming recordings to create unique beats.

Technically, the model boasts 1.21 billion parameters and is composed of an autoencoder, a T5-based text embedding, and a transformer-based diffusion model (DiT). It utilizes a total of 486,492 audio recordings for its datasets, sourced from Freesound and the Free Music Archive (FMA), all under Creative Commons licenses.

To avoid copyright issues, Stability AI has taken diligent steps, including collaborating with Audible Magic to check for copyrighted music in the Freesound samples and conducting metadata searches in the FMA subset.

It’s important to note that Stable Audio Open differs significantly from Stability AI’s commercial offering, Stable Audio, which provides subscribers with the ability to generate three-minute tracks with more sophisticated music structures and additional advanced features.

In conclusion, Stable Audio Open presents a valuable resource for amateur audio producers and hobbyists eager to explore and create unique audio outputs.

Explore AI Chatbot Performance with CheckMate, the New Open-Source Platform

A multidisciplinary team from the University of Cambridge, comprising computer scientists, engineers, mathematicians, and cognitive scientists, has developed an open-source platform named CheckMate for evaluating large language models (LLMs). This platform enables users to interact with LLMs and assess their effectiveness in real-time.

In a study published in the Proceedings of the National Academy of Sciences, the researchers tested CheckMate using three different LLMs—InstructGPT, ChatGPT, and GPT-4—to assist in solving undergraduate mathematics problems. They observed that while there was generally a positive link between the accuracy of the LLMs and their perceived helpfulness, there were cases where the models provided incorrect yet seemingly useful responses. In some instances, errors made by LLMs were mistakenly perceived as correct, particularly in models designed for conversational purposes.

The findings suggest that LLMs that can express uncertainty, react adaptively to user corrections, and succinctly explain their reasoning tend to be more effective as assistants. However, the study underscores the importance of users carefully verifying LLM outputs due to their potential inaccuracies.

Albert Jiang, a co-first author and member of Cambridge's Department of Computer Science and Technology, emphasized the necessity of both quantitative and qualitative evaluations of LLMs to better understand their interaction with humans. He highlighted the limitations of traditional evaluation methods that rely on static input-output pairs, which do not account for the dynamic interactions typical in real-world applications.

Katie Collins, another co-first author from the Department of Engineering, mentioned the mixed perceptions among mathematicians regarding LLM capabilities, ranging from overestimating their ability to generate complex mathematical proofs to underestimating their basic arithmetic skills. The CheckMate platform aims to provide a more nuanced understanding of what LLMs can and cannot do.

To validate their tool, the team engaged 25 mathematicians, from students to senior professors, who used CheckMate to work through mathematical problems with the help of an LLM. These participants rated each response for its accuracy and helpfulness, unaware of which model they were interacting with. This experiment not only tested the LLMs' capabilities but also helped identify areas for potential improvement in AI literacy and LLM development.

TheClosed.AI News

OpenAI and Google DeepMind Employees Issue Open Letter Highlighting AI Risks

On Tuesday, a group of employees and former employees from leading AI firms issued an open letter highlighting the industry's inadequate safety measures and advocating for stronger whistleblower protections. This letter, demanding a "right to warn" about AI risks, marks a significant public critique from insiders of an often secretive sector. Signatories include eleven individuals from OpenAI and two from Google DeepMind, with one also having prior experience at Anthropic.

The letter criticizes AI companies for holding significant undisclosed information about their technologies' capabilities, safeguards, and associated risks. It points out the minimal obligations these companies have to disclose such information to governments and none to civil society, expressing skepticism about their willingness to share it voluntarily.

In response, OpenAI upheld its commitment to safety, mentioning its internal reporting mechanisms and its practice of not releasing technologies without robust safety measures. Google had not responded to comments at the time of the report.

The document underscores long-standing concerns about AI's potential dangers, which have escalated with the technology's rapid development, often outpacing regulatory frameworks. Despite AI companies' public assurances of safe development, the letter reflects ongoing concerns within the community about insufficient oversight and the possible exacerbation of social harms.

The letter, initially covered by the New York Times, demands better protection for AI company employees who raise safety issues. It calls for adherence to four transparency and accountability principles, including assurances against forcing employees into non-disparagement agreements that prevent them from discussing AI-related risks, and establishing a way for employees to anonymously communicate concerns to company board members.

OpenAI Responds to Criticism with New Paper Highlighting Efforts to Mitigate AI Risks

Recently, following allegations by a former employee that OpenAI had been careless with AI development, the company published a research paper on Thursday detailing their efforts to make the mechanisms of AI models like GPT-4 more transparent. OpenAI’s intent with this publication is to bolster the explainability of AI models, reassuring the public of their commitment to managing AI risks. The paper focuses on making GPT-4's advanced features more user-friendly for researchers, developers, and AI enthusiasts.

The core aim of the newly introduced tool is to simplify the complex architecture of GPT-4, making it easier for users to understand how the model generates its outputs. AI models such as GPT-4, known for producing human-like text responses, can often seem opaque and mysterious in terms of their decision-making processes. The tool developed by OpenAI is designed to clarify these processes by isolating and defining the key concepts within GPT-4’s structure.

The research also delves into sparse autoencoders, a type of neural network that learns efficient data representations by activating only a few neurons at a time. This sparsity helps the model focus on significant data features, improving its efficiency in tasks like image recognition and data reconstruction. The paper outlines various techniques to enforce sparsity and shows how these can enhance model performance.

OpenAI believes that this tool will make GPT-4 more accessible, allowing developers and researchers to better analyze the model's outputs and refine AI applications for greater accuracy and reliability.

The paper also highlights ongoing internal challenges at OpenAI, mentioning that the research was led by the now-defunct "superalignment" team, which was dedicated to addressing the long-term risks associated with AI. Notably, the paper includes contributions from former team leaders Ilya Sutskever and Jan Leike, who have both departed from the company. Sutskever, a co-founder and former chief scientist of OpenAI, was involved in the contentious decision to temporarily dismiss CEO Sam Altman last November.

Unbabel's New AI Model Surpasses OpenAI's GPT-4 in Language Translation

Unbabel, a technology company known for blending machine and human translation services, has developed a new AI model, TowerLLM, that reportedly surpasses OpenAI's GPT-4o and other leading AI systems in translating between English and six key European and Asian languages. This achievement underscores the potential of large language models (LLMs) in business applications, particularly in translation services where GPT-4o had previously been the top contender.

TowerLLM was evaluated against several major models including GPT-4o, GPT-4, GPT-3.5 from OpenAI, as well as models from Google and DeepL. It showed superior performance, especially in translating English to Korean, where it led by approximately 1.5%, although it was slightly outperformed by GPT-4 in English-German translations.

Further testing in professional domains like finance, medicine, law, and technical writing revealed that TowerLLM consistently outperformed the best models from OpenAI by 1% to 2%. Although these results are yet to be independently verified, they suggest that GPT-4, despite its previous dominance, might now be facing competition from newer, differently trained AI models.

Unbabel, headquartered in San Francisco and Lisbon, trained TowerLLM to be multilingual using a large dataset of multilingual text. This approach not only enhanced its translation capabilities but also its performance on multilingual reasoning tasks compared to similar-sized open-source models from other companies like Meta and the French AI startup Mistral.

The model was further refined with a dataset of high-quality translations between language pairs, assisted by another AI model developed by Unbabel, named COMETKiwi, which evaluates translation quality. João Graça, Unbabel’s CTO, emphasized that unlike other LLMs that are primarily trained with English text and only incidentally learn translation, TowerLLM was intentionally trained with a rich multilingual dataset, with fine-tuning on high-quality translations being crucial for its enhanced performance.

This success of TowerLLM aligns with a broader trend where smaller AI models trained on high-quality datasets are matching or surpassing larger models. An example is Microsoft's Phi 3 model, which, despite having fewer parameters, outperformed larger models by using a "textbook-quality" dataset. Graça highlighted that the key differentiator among models now is not the algorithmic design, which remains largely the same across the industry, but the quality of data and how it is presented to the model during training.

TowerLLM is available in two versions, one with 7 billion parameters and another with 13 billion. An earlier version launched in January nearly matched GPT-4’s capabilities and worked with 10 language pairs, but the latest model surpasses GPT-4 and supports 18 language pairs.

Towards AGI: 1st Flagship in-person Event in London, UK.

We are thrilled to announce "Towards AGI: 1st Flagship Event," an in-depth exploration of advancements towards artificial general intelligence. This inaugural in-person event, hosted by the Explore Group, will take place in London and presents a unique opportunity for AI enthusiasts and professionals to connect and discuss the implementation of Generative AI in their organizations. The focus of this event will be on GenAI in Banking and Insurance.

Scheduled for June 13, 2024, from 5 PM to 8 PM, the event will be held at Portsoken House, 155-157 Minories, London EC3N 1LJ. Don’t miss out on this chance to engage in insightful discussions and expand your network in the field of AI.

We are delighted to invite you to an evening that promises to be both stimulating and informative. The main event will feature a Fireside Chat titled "GenAI in Banking," lasting 40 minutes. During this session, experts will explore banking and asset management use cases, discussing how organizations are balancing governance concerns while exploring Generative AI. The chat will also include a Q&A session, allowing the audience to interact directly with the speakers and ask insightful questions.

Following the Fireside Chat, we will have Live Demonstrations lasting 30 minutes, showcasing real-world applications and use cases of Generative AI. This segment will include a demonstration of an insurance underwriter workflow with an embedded GenAI assistant and a walkthrough of an AI Chatbot designed for central metadata management in banks.

Join us for an evening of insightful discussions, live demonstrations, and networking opportunities as we delve into the world of Generative AI in banking and insurance. We look forward to seeing you there!

Click here to fill the event form

Keep reading

In our quest to explore the dynamic and rapidly evolving field of Artificial Intelligence, this newsletter is your go-to source for the latest developments, breakthroughs, and discussions on Generative AI. Each edition brings you the most compelling news and insights from the forefront of Generative AI (GenAI), featuring cutting-edge research, transformative technologies, and the pioneering work of industry leaders.

Highlights from GenAI, OpenAI and ClosedAI: Dive into the latest projects and innovations from the leading organisations behind some of the most advanced AI models in open-source, closed-sourced AI.

Stay Informed and Engaged: Whether you're a researcher, developer, entrepreneur, or enthusiast, "Towards AGI" aims to keep you informed and inspired. From technical deep-dives to ethical debates, our newsletter addresses the multifaceted aspects of AI development and its implications on society and industry.

Join us on this exciting journey as we navigate the complex landscape of artificial intelligence, moving steadily towards the realisation of AGI. Stay tuned for exclusive interviews, expert opinions, and much more!