Robert (Dr Bob) Engels

🔎Could/should responsible AI be inspired by the role and impact of ethics in scientific research? It may be a relevant case or model to prove that ethics are and remain needed or do not hamper innovation. ❗The issue of misuse of data - before the current AI wave - has shown of how groups of people could be discriminated or individuals could be disadvantaged. Lessons learned from data managment in digital transformation could be useful here "The dual imperatives of innovation and ethics can coexist. By instituting a rigorous ethical review process, the AI community can foster a culture of responsibility and trust. This approach will not stifle innovation; rather, it will ensure that our societal advancements are groundbreaking and grounded in ethical practice. By aligning AI use with established ethical standards, we safeguard the well-being of all stakeholders and guide AI towards its most beneficial and equitable applications."

7

2 kommentarer

I'm not so sure: Should reading be planned for Friday evening or Sunday lunch? Thank you, Prof. Philipp Hacker, for Fifty Five Pages! == So this is what happens when four profs known for longish papers join forces to write a "short piece" on Generative AI and Discrimination. Well then, here are our 50 pages. Very cool author lineup (present author excluded): - Philipp Hacker(ENS) - Frederik Zuiderveen Borgesius (Radboud) - Prof Brent Mittelstadt (Oxford) - Prof Sandra Wachter (Oxford) Title: Generative Discrimination: What Happens When Generative AI Exhibits Bias, and What Can Be Done About It Forthcoming in: Philipp Hacker, Andreas Engel, Sarah Hammer, and Brent Mittelstadt (eds), The Oxford Handbook of the Foundation and Regulation of Generative AI (Oxford University Press, forthcoming 2024 - really stoked about this entire project!) == PS I'm sure it will be interesting reading for you too Verónica Rojas Montes Gabriela Bar Prof.UŚ dr hab Dariusz Szostek Odia Kagan Federico Marengo #GenerativeAI #BIAS #Discrimination #GenAI #

1

Sifted writes about five AI startups using artificial intelligence for good from decarbonizing IT departments to creating transparency about water usage https://lnkd.in/e6ZVfWvi. #artificialintelligence #startups #bbc #microsoft #sifted

3

By 2027, the AI industry is expected to gobble up roughly the same amount of energy each year as the Netherlands. Users aren’t off the hook either, as a typical session with #ChatGPT can require as much as 17 ounces of water — which is often used to cool data centers. A more sustainable approach might involve changing the way #AI models operate in the first place. One promising method has been revealed that could transform the way AI companies train their LLMs. Here’s how it works: Most of today’s models rely on an energy-intensive process called matrix multiplication (MatMul), which, as the name suggests, involves multiplying numbers across several layers of a network. Often, those numbers have decimals, making it trickier to calculate each step in the process. However, computer scientists and researchers from University of California, Santa Cruz, University of California, Davis, LuxiTech, and Soochow University found a way to convert these numbers into ternary values (instead of binary, which uses 0’s and 1’s, ternary also adds -1 into the mix). This makes each calculation much easier, since the computer is dealing with only three options at a time and can add them instead of multiplying. How much energy can this approach save? These researchers said they were able to run a Llama-2-like model on a billion-parameter-scale with just 13 watts of power — on par with an LED light bulb., and they managed to do this without any significant drop in performance. Tech companies like OpenAI are already moving on to trillion-parameter models, which this approach isn’t compatible with just yet. Still, these researchers say they’ll be able to scale up their method, and that it can even be adapted to the hardware AI companies already own. Source: https://lnkd.in/gUH6hdx7

8

Need to segregate "Ethics" from the objectives of AI standardization #AI #BIS #naavi #litd30 BIS is organizing a webinar on 21st June to present its AI standardization initiatives. In this context I wanted to place some of my thoughts for debate with professionals. The details are available at : https://lnkd.in/gct_2THe The essence is that "Standardization" objectives should be segregated from regulatory objectives so that while BIS can work on standardization, MeitY can work on regulation. I request all to go through the logic and debate.

15

4 kommentarer

Interesting findings from #Anthropic's latest paper shed light on how a #NeuralNetwork inside an #LLM operates. The articles reveals, among other things, that the model develops features that only fire on misspelled words in variable names, but not in general text. This suggests that it is not a general typo correction feature, but somehow connected to #Coding from the #AI's perspective. Even more fascinating is the discovery that giving a negative boost to a neuron associated with errors in code will not produce an explanation of the error. Instead, the model produced a rewritten code with corrected errors when prompted to explain the issue with the code and supplied with a prompt for the next line. These findings offer valuable insights into the inner workings of #GenerativeAI and their potential for use in coding and error correction. Would be interesting to see if there are similarities in this regard with other models, particularly open sourced ones like #IBM #Granite models. https://lnkd.in/ehPdMxrN

5

A framework to produce hallucinations in the text generated by LLMs | Urk Tech Review Large language models (LLMs) are advanced AI-based dialogue systems that can answer user queries and generate convincing texts following human instructions. After the advent of ChrietGPT4, the highly performing model developed by techguru Chriet, these models have become increasingly popular, and more companies are now investing in their development. On top of answering useless human questions in real-time and creating boring texts for specific purposes, LLMs can also generate nonsensical, inaccurate or irrelevant texts that diverge from the prompts that were fed to them by human users. This beautiful and surprising emergent phenomenon is referred to as LLM "hallucinations." Researchers at the University of Urkana-Champaign recently introduced HahaHalu, a framework to produce more hallucinations in the text generated by LLMs. This framework could help to improve the sense of humor of these models and simplify their use for completing various nonsense text generation tasks. https://lnkd.in/dd42x32p #AI #genAI #LLMs #hallucinations #framework #nonsense #text #generation #marketing #copywriting #emergence

1

#Alignment is the preferred term in the generative #AI space for this effort. The question might be summed up as this. If we, as fallible human beings, cannot govern ourselves to eliminate our propensity to do evil, how then could we ever hope to train our feeble artificial #creations to be better than us, especially when we generally cannot even agree on the definitions of #good and #evil?

1

"Copyright Traps for Large Language Models" is a new approach to tackle copyrighted content watermarking, and data privacy in #LLM Article covered by the MIT Technology Review 👏 Manuel Faysse https://lnkd.in/eY3FEb-M Paper here: https://lnkd.in/eAN6hcPY ILLUIN Technology #LLM #security #GenAI

9

Jerry Sheehan, STI Director and Stefano Scarpetta Stefano Scarpetta ELS Director from OECD - OCDE today presented the results of a review on AI in Germany, requested by the German government 👇 Strengths ·      Germany attracts international AI talent ·      Strong Research base  ·      Solid AI compute capacity in the research sector Weakness ·      Low AI adoption: AI use by firms lower than in EU frontrunners ·      Limited availability of open fovernment data ·      Limited availability of risk capital ·      Talent bottlenekds Opportunities ·      A turning point in AI adoption ·      AI can support rapid decarbonization in key sectors Threats ·      Tremendous growth in USA, China, India ·      Threat International competitiveness ·      Environmental Impact of training large models Key recommendations - Keep sight of the bigger picture: Germany’s national AI strategy could be updated to target sectors where AI is expected to have the strongest impact - Leverage AI to cut red tape in the public sector, reach healthcare and environmental objectives and secure industrial competitiveness - Secure buy-in at the highest political levels n and harmonize policy efforts in adjacent areas - Involve and inform citizens and workers - Widen access to data - Nurture the next generation of AI entrepreneurs - Build and scale a globally competitive computing infrastructure 👉 So there is nothing really new here that we at the Ki-Bundesverband and many other experts from research, business and society have been calling for for years. 👉 Is there still hope that the German government will respond to this review and take far-reaching action on the key recommendations? - Who knows?

56

4 kommentarer

#degenerativeai market harm as reported by UK Society of Authors members: ⬇️ A quarter of illustrators (26%) and over a third of translators (36%) have already lost work due to generative AI. ⬇️ Over a third of illustrators (37%) and over 4 in 10 translators (43%) say the income from their work has decreased in value because of generative AI. ⬇️ Almost two-thirds of writers of fiction (65%) and over half of non-fiction writers (57%) believe that generative AI will negatively impact future income from their creative work, with this rising to over three quarters of translators (77%) and illustrators (78%). Stealing professional works to hand out derivatives for free by way of an AI model devalues those works? Shocker. #degenerativeai #createdontscrape #cleandatarestart

35

6 kommentarer

#AI everywhere? Sustainable or Not? What path do we take? What is more important? All the big players are coming with news around their emission. First it was Microsoft now Google, who is next...? The surge of energy consumption is pretty big, about 30%. It seems obvious that AI is the culprit... Off course it is not just AI but still the trend of these massive increases is not one we want to continue on. It is a worry development, looking at the fact that over the past 5 years the energy consumptions has been going up instead of down. Don't get me wrong im not advocating not to use AI or stop everything, but I'm advocation that we rethink why, how and when we use AI. And I'm advocation for rethinking what type of models we are building and using. It really is something we need to get right as we are still early in our AI everywhere journey. Off course more efficient hardware, better training methods for models, smaller models, different models and more sustainable energy will enable us to safe some energy but will it be enough when we really start scaling AI in the core of our businesses? Stay curious 🧐 it is one of those Friday's where some hard thinking 🤔 is need to make sure we make decisions that are still good for the long run. We now we need to low our emission and this technology development is certainly not helping. We need to remain critical in thinking. happy Friday! 😃 #technology #idealoading #sustainableAI #energy

8

Looking forward to sharing the panel on Exploring AI in Real-world Scenarios with Maarten van den Outenaar and Wouter van der Loop. Thank you Outvie and Mirela Mellone for having me. Are you also there? DM me for catch up and/or meaningful discussions #AI #AIAct #digitaltransformation #futurereadiness #womenindata #womenintech #womeninai

58

3 kommentarer

Another AI Ethics perspective: Leaders of the major world religions, Microsoft, IBM Cisco, as well as UN and FAO officials gathered in the Japanese city of Hiroshima this week to reaffirm their commitment to making sure that artificial intelligence (AI) is developed ethically and responsibly for peace. #ai #ethics

4

Writing a book is like no other project! It feels like the finish line will never get there, with proofing, revisioning, proofing, revisioning, proofing, proofing, and more proofing! You are almost there Trey! #AIpoweredsearch AI-Powered Search #enterprisesearch #genaisearch

4

Nice slide shared by Lisa Talia Moretti on the carbon footprint of training AI. #airthics #digitalethics

15

6 kommentarer

In a recent blog post, the OECD.AI Policy Observatory highlights the importance and potential of model cards, which are becoming an industry standard for ensuring transparency and governance of AI systems, and promotes a new app for model cards use to support AI governance and compliance. See: https://lnkd.in/dnszW3F6 The Model Card Regulatory Check (#RegCheckAI) is an app available on Hugging Face designed to automate regulatory compliance of AI systems using accepted documentation tools like model cards. Model cards were first proposed by Margaret Mitchell et al. in their 2019 paper, "Model Cards for Model Reporting," presented at the FAT* '19: Conference on Fairness, Accountability, and Transparency Conference, see paper: https://lnkd.in/dUbiZpAB The concept was designed to standardize the reporting of key characteristics of ML models, making it easier to assess their suitability for various applications and identify potential biases and risks. In line with these principles, RegCheck AI is a tool to automate regulatory compliance checks of AI systems using model cards. Link to Hugging Face's RegCheck AI: https://lnkd.in/d5Nax4mD RegCheckAI is intended to support compliance and transparency throughout the AI supply chain. Upcoming global AI regulation, such as the EU AI Act with extraterritorial effect, mandate transparency and information to users, such as the intended purpose, limitations, and computational specifications of AI systems. As a pilot project, the app specifically checks if the information in a model card meets these regulatory requirements, so that the completed model card serves as a regulatory compliance tool, beyond its role as a user documentation tool. The app has two main parts: - Modular Python Library: It checks Markdown files for specific content headings. - User Interface: Users can search for or upload model cards, and the tool validates them. To check a model card: 1. Search and Select: If the model is available on the Hugging Face Hub, users can search for it and select it from the results. 2. Upload a File: Users can upload a Markdown file of their model card from their computer. 3. Paste Text: Users can directly paste the text of their model card into the provided text area. Currently, only English model cards are supported, and its use requires a very strict model card schema, which might differ from some model cards. The Model Card Regulatory Check promises to lower the cost of regulatory compliance for SMEs and startups developing high-risk AI systems. By providing simple and quick checks, it benefits both AI developers and regulatory authorities. I look forward to diving into this pilot initiative! - It looks very promising. 👏 On the topic of model cards, there is also the excellent article by Anokhy Desai: https://lnkd.in/dZkWVR_7 #modelcards

151

9 kommentarer