. . AI: Dichotomy: US Leading in Model Cost and Performance While the EU Leads Regulation and Safety (3.14.24)

Friends, the EU Parliament has passed the first major regulations on AI. This development is expected to have far-reaching implications on the AI landscape, setting the stage for a new era of regulation and oversight. Meanwhile, on the other side of the world, the US is leading the charge with the most cost-efficient and powerful models. Anecdotally, when speaking to AI founders, we're hearing interest in avoiding deep entanglement with parties in the EU.

In other news, we see a new appointment at Google, and jousting between Anthropic and OpenAI for pole positions on the leaderboard.

Thanks for reading!

-Marshall Kirkpatrick, Editor

EU Parliament Approves First Major AI Regulatory Rules

First impacted: World Leaders, Regulation and Compliance Teams
Time to impact: Medium

The EU Parliament has given the green light to the world's first major regulatory rules for AI, known as the EU AI Act, with 523 votes in favor, 46 against, and 49 abstentions. The Act, which categorizes AI technology into risk levels, is expected to be implemented in May 2024, following final checks and approval from the European Council, despite concerns from some EU countries about potential impacts on competition at a global scale. [World’s first major act to regulate AI passed by European lawmakers] Explore more of our coverage of: EU AI Regulation, Tech Competition, Risk Management. Share this story by email

Anthropic Launches Claude 3 Haiku, A Bang for Buck Leader

First impacted: Enterprise AI, AI developers
Time to impact: Short

Anthropic has launched Claude 3 Haiku, a model they describe as the fastest and most affordable in its class, capable of processing 21K tokens per second for prompts under 32K tokens and with a cost of $0.25 and $1.25 per million tokens for input and output, respectively. The company highlights that this model, designed for enterprise workloads, offers quick results, robust security, and the ability to process large volumes of documents at half the input cost of other models in its performance tier like GPT3.5 and Gemini Pro 1.0. [Claude 3 Haiku: our fastest model yet] Explore more of our coverage of: Anthropic, AI Models, Enterprise Workloads. Share this story by email

Anthropic's Claude-3-Opus and Sonnet Challenge GPT4-Turbo for Top Position

First impacted: AI developers, Multilingual content creators
Time to impact: Short

The AI models Claude-3-Opus and Sonnet are gaining traction on the lmsys.org chatbot arena leaderboard, with Claude-3-Opus now sharing first place with GPT-4-Turbo and Sonnet outperforming GPT-4-0314. Lymsys runs a blind taste-test system where tens of thousands of people have put a prompt into the arena and then answered whether they liked model A or model B's response best. It's pretty fun. https://arena.lmsys.org/ The post by Lymsys.org also mentions that Claude-3-Opus has shown impressive capabilities in multi-lingual domains, leading to plans for separate leaderboards for different domains of interest. [via @lmsysorg] Explore more of our coverage of: AI Models, Multi-Lingual Domains, Claude-3-Opus. Share this story by email

Xavi Amatriain Joins Google as VP of Core ML/AI

First impacted: AI developers, Google Employees
Time to impact: Medium

Xavier (Xavi) Amatriain has taken on the role of VP of Product for Core ML/AI at Google, where he will oversee AI products. He comes into the role with experience leading AI and ML teams at Netflix, Quora and Linkedin. [via @xamat] Explore more of our coverage of: Google AI, VP Appointment, Core ML/AI. Share this story by email

Nous Research Launches Hermes 2 Pro Aimed at Financial Data

First impacted: Financial Analysts, Software Developers
Time to impact: Short

Nous Research has launched the Hermes 2 Pro, a language model that they say is capable of executing function calls and retrieving financial data such as stock prices and company fundamentals. The model, which works through a command line interface and JSON mode inference, allows users to add custom functions and Pydantic models to create JSON schema, and uses a prompt format, ChatML, to engage in multi-turn chat conversations. [GitHub - NousResearch/Hermes-Function-Calling] Explore more of our coverage of: Language Models, Financial Data, Function Execution. Share this story by email