Issue 62 23rd July 2024

Highlights

GPT-4o mini: advancing cost-efficient intelligence

Introducing our most cost-efficient small model

Scarlett Johansson says OpenAI’s Sam Altman would make a good Marvel villain after voice dispute

Actor, who claimed ChatGPT update used an imitation of her voice, says she declined to provide her own as ‘it went against my core values’ Scarlett Johansson has spoken out against OpenAI and deepfake technology, saying it was “so disturbing” and she was “so angry” after the company seemingly mimicked her voice for its ChatGPT system Sky.

Are We Ready for Multi-Image Reasoning? Launching VHs: The Visual Haystacks Benchmark!

Humans excel at processing vast arrays of visual information, a skill that is crucial for achieving artificial general intelligence (AGI). Over the decades, AI researchers have developed Visual Question Answering (VQA) systems to interpret scenes within single images and answer related questions. While recent advancements in foundation models have significantly closed the gap between human and machine visual processing, conventional VQA has been restricted to reason about only single images at a time rather than whole collections of visual data.

Large language models don’t behave like people, even though we may expect them to

A new study shows someone’s beliefs about an LLM play a significant role in the model’s performance and are important for how it is deployed.

TTT models might be the next frontier in generative AI

TTT models, a new architecture, could effectively replace transformers if they scale up as their creators suggest they will.

Video

Copilot L33t Sp34k | Moving from Researcher to Practioner in AI Security

Harriet Farlow, CEO of Mileva Security Labs, discusses how AI security has moved from an academic topic to being implemented across organizations and the best practices for doing this.

Articles

The world is not quite ready for ‘digital workers’

CEO Sarah Franklin got such intense pushback on her company’s plans that she suspended them after three days One thing seems for sure: people are not ready for “digital workers” just yet. That’s the lesson learned by Sarah Franklin, the CEO of Lattice, a human resources and performance management platform that offers performance coaching, talent reviews, onboarding automation, compensation management and a host of other HR tools to more than 5,000 organizations around the world.

Microsoft at ICML 2024: Innovations in machine learning

The competitive dynamics of AI agents and a method for learning and applying temporal action abstractions represent just some of Microsoft’s contributions to ICML 2024.

Advanced Data Modelling

Data model layers, environments, tests and data quality explained.

LLM Agents Demystified

We will first introduce ReAct [2], a general paradigm for building agents with a sequential of interleaving thought, action, and observation steps. In addition to the tools provided by users, by default, we add a new tool named finish to allow the agent to stop and return the final answer.

Breaking Instruction Hierarchy in OpenAI's gpt-4o-mini

Johann Rehberger digs further into GPT-4o's "instruction hierarchy" protection and finds that it has little impact at all on common prompt injection approaches.

Deep Dive on the Hopper TMA Unit for FP8 GEMMs

The Hopper (H100) GPU architecture, billed as the “first truly asynchronous GPU”, includes a new, fully asynchronous hardware copy engine for bulk data movement between global and shared memory called Tensor Memory Accelerator (TMA). While CUTLASS has built-in support for TMA via its asynchronous pipeline paradigm, Triton exposes TMA support via an experimental API.

Mistral and Nvidia Bring AI to Local Machines

Mistral AI and Nvidia have introduced Mistral NeMo, an artificial intelligence (AI) model designed to bring advanced AI capabilities to standard desktop computers, widening access for large and small businesses. Mistral NeMo’s key feature is its ability to process up to 128,000 words simultaneously, handling complex tasks like document analysis and code generation without relying.

Creating and verifying stable AI-controlled systems in a rigorous and flexible way

Neural network controllers provide complex robots with stability guarantees, paving the way for the safer deployment of autonomous vehicles and industrial machines.