AI Pulse – Edition 1
Table of Contents
1. Government & Policy Initiatives
2. Legal Matters
3. Technology Updates
4. AI Economics
5. Research
1. Government & Policy Initiatives
U.S. AI Safety Institute Collaborates with Anthropic and OpenAI on AI Safety Research
The U.S. Artificial Intelligence Safety Institute, under the Department of Commerce’s National Institute of Standards and Technology (NIST), has announced agreements with Anthropic and OpenAI for formal collaboration on AI safety research, testing, and evaluation. The agreements will allow the Institute to access new models from both companies before and after their public release, providing an opportunity to evaluate model capabilities and associated safety risks as well as to collaborate on methods to mitigate these risks. Link
Controversial California’s AI Safety Bill Clears Legislature
California lawmakers have passed the contentious AI safety bill, SB 1047, which now awaits the signature of California’s Governor Gavin Newsom. The bill mandates safety testing, including third-party audits of safety practices, for advanced AI models such as those that cost over $100 million to develop or require a significant amount of computing power. It also requires AI developers to outline methods for disabling AI models if they malfunction, effectively implementing a kill switch. The bill empowers the state attorney general to sue non-compliant developers, particularly in the event of an ongoing threat, such as AI taking over government systems. OpenAI’s Chief Strategy Officer, Jason Kwon, previously expressed concerns over the new bill, stating that it could slow progress and cause companies to leave the state. He stressed that AI regulations should be left to the federal government, arguing that a federally-driven set of AI policies, rather than a patchwork of state laws, will foster innovation and position the U.S. to lead the development of global standards. Link
Biden-Harris Administration Targets Ineffective AI Chatbots
The Biden-Harris Administration has launched the “Time Is Money” initiative to address corporate practices that waste consumer time, including ineffective AI chatbots. Under the initiative, the Administration is looking to actively address the challenges and limitations of consumer chatbots, such as receiving inaccurate information, and has signaled plans by the Consumer Financial Protection Bureau (CFPB) to release new rules or guidance setting out when the use of automated chatbots or automated artificial intelligence voice recordings is permissible. Link
2. Legal Matters
SAG-AFTRA and Narrativ Establish Ethical AI Agreement
SAG-AFTRA has partnered with Narrativ to offer its members the opportunity to ethically license their digital voice replicas for use in digital audio advertising. The agreement ensures that performers have control over the use of their digital voices, including informed consent, fair compensation, and the ability to set personal ad preferences. Link
3. Technology Updates
OpenAI Launches Fine-Tuning for GPT-4o and Enhances File Search Controls in Assistants API
OpenAI has introduced new features for its developers, including the launch of fine-tuning for GPT-4o, expanding on the previously available GPT-4o-mini fine-tuning. This allows developers to customize the structure and tone of model responses or ensure it follows complex domain-specific instructions more consistently. Through September 23, developers can take advantage of 1M free training tokens per day for GPT-4o fine-tuning and 2M free tokens for GPT-4o-mini fine-tuning. Additionally, OpenAI has improved the File Search controls in its Assistants API to enhance response relevance. Developers can now inspect and re-rank search results, adjust the ranker, and set a relevance score threshold for file chunks used in generating responses. Link
Microsoft Enhances AI Offering with New Phi Model
Microsoft has announced several updates to its AI offerings, including enhancements to the Phi family of models. The Phi-3.5-Mixture of Expert (MoE) model integrates 16 smaller experts into a single system with a total of 42 billion parameters, while however only activating 6.6 billion parameters at any given time. This design offers the computational efficiency and latency of a smaller model while maintaining the high-quality output of a larger one. Additionally, Microsoft has introduced a new 3.8 billion parameter mini model, Phi-3.5-mini, which now supports over 20 languages, expanding its global usability. Link
Anthropic Reveals System Prompts for Claude AI Models and Makes Artifacts Available for All Users
In an effort to enhance transparency, Anthropic has published the system prompts for its latest models Claude 3 Opus, Claude 3.5 Sonnet, and Claude 3 Haiku, marking a first in the industry. The system prompts reveal the models’ operational boundaries and desired traits such as intellectual curiosity and impartial handling of controversial topics. Alongside this release, Anthropic has also expanded the availability of its Artifacts feature to all users of its platform. Artifacts allow for the creation and management of various creative outputs like code snippets and interactive dashboards directly within the app. Link
Google Introduces Multiple Gemini Updates, Including Custom Gems, Imagen 3 and New Experimental Models
Google has rolled out several updates to its Gemini platform. These include the introduction of custom Gems that allow for the creation of personalized AI experts, improvements to its Imagen 3 text-to-image model as well as the release of three experimental models including a new smaller variant, Gemini 1.5 Flash-8B, as well as enhanced Gemini 1.5 Pro and Gemini 1.5 Flash models. Additionally, it has now also made available Structured Outputs. Link
X.ai Releases Grok-2 and Grok-2 Mini
X.ai has released its new models Grok-2 and Grok-2 Mini in beta. Both models come with advanced chat, coding and reasoning capabilities over the preceding Grok-1.5 model. Grok-2 was initially tested on the LMSYS leaderboard under the pseudonym “sus-column-r” where it outperformed Claude 3.5 Sonnet and GPT-4-Turbo at the time of the launch in terms of its overall Elo score. Both models are currently accessible to X premium users and are due to become available via the platform’s enterprise API. Link
4. AI Economics
OpenAI Considers Corporate Restructuring Amid Funding Talks
OpenAI is reportedly in talks to raise billions in a new funding round that could value the company at over $100 billion. The fundraising round, led by venture capital firm Thrive Capital, could see participation from Apple and Nvidia, alongside existing OpenAI partner Microsoft, which already owns a 49% share of OpenAI’s profits after contributing ~ $13 billion to the startup in 2019. To attract more investors, OpenAI is also said to be considering a change in its corporate structure that would involve simplifying the current complex non-profit structure. The current structure, which issues equity via its for-profit subsidiary governed by a non-profit board, has been under scrutiny, with some investors suggesting a shift to a simpler for-profit structure would be more beneficial. Link
Big Tech’s AI Investment Surge Continues
Big Tech companies, including Microsoft, Alphabet, Amazon, and Meta, have collectively increased their AI investment to over $100 billion in the first half of 2024. Despite scepticism from Wall Street regarding the returns on these investments, the firms remain committed to further expanding their AI infrastructure over the next 18 months. Link
5. Research
MIT Releases AI Risk Repository to Guide Safe AI Adoption
Researchers at Massachusetts Institute of Technology (MIT) have released a comprehensive AI Risk Repository to serve as a common frame of reference for understanding and mitigating potential risks associated with AI deployment. The repository categorizes over 700 AI risks into a causal taxonomy that classifies AI risks by its causal factors and a domain taxonomy that categorizes AI risks into seven domains and 23 subdomains. Link
FinalSpark Introduces Biocomputing with Human Neurons
Swiss company FinalSpark is pioneering the field of biocomputing with its “Neuroplatform”, a computer platform powered by human-brain organoids. The platform aims to support AI with 100,000 times less energy than currently required to train state-of-the-art generative AI. The organoids, which are clusters of lab-grown cells, are connected to electrodes that stimulate the neurons within the living sphere and link the organoids to conventional computer networks. The neurons are trained to form new pathways and connections, similar to how a human brain learns. Link
OECD Explores AI Integration in Public Governance
The Organisation for Economic Co-operation and Development (OECD) has released a policy paper titled Governing with Artificial Intelligence: Are Governments Ready? which seeks to provide a roadmap for governments in navigation AI. The paper emphasizes key benefits in AI adoption including productivity enhancement, responsiveness and inclusivity, and accountability and oversight while warning of risks such as bias and fairness, transparency and explainability, and data privacy and security. To mitigate these risks, the OECD recommends developing strategic frameworks, building capacity, setting regulatory standards, and fostering international collaboration. Link
