Andrea Mondello, CPBPM’s Post

How does Anthropic’s Responsible Scaling Policy (RSP) align with the EU AI Act? The obligations for a General Purpose AI model (Claude 3) are set out under Article 53 of the AI Act. Here’s how Anthropic’s RSP lines up: 1. Technical Documentation • AI Act: Keep technical documentation available for examination by the EU AI Office. • RSP: Publishes model cards detailing capabilities, limitations, evaluations, and use cases (page 6). 2. Documentation for Providers • AI Act: Maintain documentation for providers integrating AI models, balancing transparency and IP protection. • RSP: Partners must adhere to Anthropic’s safety protocols, ensuring responsible scaling and safety measures (page 10). 3. Summary of Training Data • AI Act: Publish a summary of AI model training data using the AI Office’s template. • RSP: Shares evaluation results and includes training data summaries in model cards where possible (pages 6 and 11). But: 4. The AI Act mandates that General Purpose AI models respect EU copyright law. More on the EU text and data mining exception for General Purpose AI models to come. #AICompliance #TechPolicy #ResponsibleAI #AIAct #AnthropicAI

Anthropic just published a significant update to its Responsible Scaling Policy, introducing new safeguards and governance measures for advanced AI systems and highlighting future benchmarks requiring additional safety protocols. 💡 The policy introduces ‘Capability’ and ‘Required’ Thresholds to trigger enhanced safety measures when AI models reach certain risk levels. 💡 The two new thresholds focus on AI capabilities related to bioweapons and autonomous AI research. 💡 Anthropic emphasized the need for the risk approach to be ‘exportable,’ hoping that it will become an industry standard and help shape regulation. 💡 Anthropic will regularly evaluate its AI models, while a ‘Responsible Scaling Officer’ role will oversee policy implementation and compliance. 💡 The company also pledged increased transparency, including public disclosure of capability reports and external expert input. Anthropic's update represents one of the most comprehensive frameworks yet for responsibly developing advanced AI and continues to cement the company’s standing as the ‘safety-focused’ lab. The move may also foreshadow some major acceleration imminent with new releases in the near future. #ai #technology #anthropic https://lnkd.in/gdi_pY68

In what ways does Artificial Intelligence (AI) really threaten our society 🌍 ? Will exisiting legislation 📜 sufficiently protect us? And what further can be done? Together with Bart van der Sloot from Tilburg University 👩💻 my colleague Leon de Beer takes a look beyond the hype and explores the systemic risks of AI, possible interventions, and the way forward. Great read!

"Reasoning through arguments against taking AI safety seriously" - a marvellously clear blogpost by Yoshua Bengio. It's no wonder that he is the computer scientist with the highest impact in the world as measured by h-index. Bengio addresses paragraphs to each of the following groups of people: Those who think AGI and ASI are impossible or are centuries in the future; Those who think AGI is possible but only in many decades; Those who think that we may reach AGI but not ASI; Those who think that AGI and ASI will be kind to us; Those who think that corporations will only design well-behaving AIs and existing laws are sufficient; Those who think that we should accelerate AI capabilities research and not delay benefits of AGI; Those concerned that talking about catastrophic risks will hurt efforts to mitigate short-term human-rights issues with AI; Those concerned with the US-China cold war; Those who think that international treaties will not work; Those who think the genie is out of the bottle and we should just let go and avoid regulation; Those who think that open-source AGI code and weights are the solution; Those who think worrying about AGI is falling for Pascal’s wager

The foundation AI model value chain is complex, with many players, differing motivations, and varying influence when it comes to ensuring safety for people and society. At a recent Partnership on AI workshop, hosted by GitHub, experts came together to create a clear map of the open foundation model value chain. We hope this collective analysis of a complicated ecosystem is a resource for researchers to interrogate the source of different risks more effectively and for practitioners and policy-makers to develop targeted interventions for responsible, real-world impact. Read more here: https://lnkd.in/gvKhSfEU Thank you to PAI’s Head of AI Safety Madhulika Srikumar, GitHub, and the workshop attendees who contributed to this vital work. #AIGovernance #Policy #RiskMitigation #AIPolicy #AIEthics #ResponsibleAI #AIRegulation #TechPolicy

At the IDAIS-Venice 2024, global leaders and experts in AI endorsed a consensus statement recognizing AI safety as a global public good. Amid rapid advancements in AI, the statement urges international cooperation to establish emergency preparedness, safety assurance frameworks, and independent AI safety research. Key recommendations include developing global governance regimes, instituting AI Safety Institutes, and creating mechanisms for verifying AI safety claims. This global effort aims to avert catastrophic risks posed by advanced AI systems, emphasizing transparency, accountability, and collaboration across borders. #ResponsibleAI https://lnkd.in/ghkhBrFM

Nathan Hamiel gets real about the current state of #genAi-generated garbage. Even if it doesn’t deepfake the populace into picking up their pitchforks, the feedback loop of junk still does damage. The real 💰is this: “I’ve said all along that regulating underlying technology is a losing proposition. What should be regulated are use cases.” I’d go a step further. I don’t think either path is particularly useful anymore in the U.S., but regulation is absolutely not a preventative control. There are decades of unenforceable or unenforced restrictions on non-Ai tech, as prosecutors are definitely not equipped to discern and we tech folks can rarely agree long enough to classify the really offensive behavior. At least by strengthening rules on the use cases we might end up with something that is enforceable, albeit often too late to have any meaningful benefit to risk reduction. #ResponsibleAi #SecureAi #VerifiableAi #PrivateAi

Many government are trying to understand the risk & benefits of these new #GenerativeAI technologies. One thing is for sure - #GenAI & the new models that follow - are here to stay ! buff.ly/48X58vO #innovation #disruption #ArtificialIntelligence #AI

Reasoning through arguments against taking AI safety seriously - Yoshua Bengio "I worry that with the current trajectory of public and political engagement with AI risk, we could collectively sleepwalk – even race – into a fog behind which could lie a catastrophe that many knew was possible, but whose prevention wasn’t prioritized enough." In this fascinating post Yoshua Bengio discusses why AI safety should be taken seriously. He considers the pros and cons associated with many of the arguments in this space including specifically addressing the following: - For those who think AGI and ASI are impossible or are centuries in the future - For those who think AGI is possible but only in many decades - For those who think that we may reach AGI but not ASI - For those who think that AGI and ASI will be kind to us - For those who think that corporations will only design well-behaving AIs and existing laws are sufficient - For those who think that we should accelerate AI capabilities research and not delay benefits of AGI  - For those concerned that talking about catastrophic risks will hurt efforts to mitigate short-term human-rights issues with AI - For those concerned with the US-China cold war - For those who think that international treaties will not work - For those who think the genie is out of the bottle and we should just let go and avoid regulation - For those who think that open source AGI code and weights are the solution - For those who think worrying about AGI is falling for Pascal’s wager - For those who discard x-risk for lack of reliable quantifiable predictions #AI #artificialIntelligence #AIsafety #ResponsibleAI #AIethics #EthicalAI #AIgovernance #EthicsinAI https://lnkd.in/evqPc5MY

AI Giant Anthropic Sounds 18-Month Alarm on AI Regulation In a stunning warning that's sending shockwaves through Silicon Valley, AI powerhouse Anthropic has issued what amounts to an urgent SOS to governments worldwide: Regulate AI within 18 months, or risk potential catastrophe. The warning comes at a critical moment, just days before the U.S. presidential election, highlighting the intersection of technological advancement and political urgency. What makes this particularly noteworthy is that it's coming from within the industry itself – specifically from a company known for its cautious approach to AI development. The numbers are startling. In just one year, AI's capabilities have skyrocketed. On complex coding tasks, AI performance jumped from a modest 1.96% success rate to an astonishing 49%. Even more concerning, these systems are now matching PhD-level experts in advanced scientific knowledge, particularly in sensitive areas like chemical and biological research. "The window for proactive risk prevention is closing fast," Anthropic warns in what might be the most significant corporate risk assessment of 2024. Their internal "Frontier Red Team" has discovered that current AI models can already assist with sophisticated cyber operations – and the next generation will be even more powerful. But here's the twist: Anthropic isn't calling for a shutdown. Instead, they're proposing a nuanced approach they call "targeted regulation." Think of it as a safety harness for AI development, not a cage. Their proposal centers on three key pillars: 1. Transparency: Companies must publish clear safety policies and risk evaluations 2. Security Incentives: Governments should reward robust safety practices 3. Simplicity: Regulations should be clear and focused, avoiding bureaucratic bloat What makes this particularly fascinating is the timing. With AI capabilities advancing exponentially – scientific understanding improved by 18% in just three months – we're watching a real-time race between technological progress and regulatory frameworks. The message is clear: The AI industry is at a crossroads. Either we implement thoughtful regulation now, or risk facing what Anthropic calls "poorly-designed, knee-jerk regulation" later – potentially getting "the worst of both worlds." This is more than just another tech company statement; it's a watershed moment in the AI industry's relationship with government oversight. The question now isn't whether to regulate AI, but how quickly we can do it right. 🧵 Follow for more breaking tech analysis and AI developments...