Today:

• OpenAI Shows ‘Human-Level’ Problem Solving

• OpenAI Enhances Synchron's BCI Chat

• Meta Enhances LLM Reasoning Efficiency

• SoftBank Acquires Graphcore AI Chipmaker

• Amazon AI Assistant Rufus Launches

• Copilot AI Reads Handwriting Now

• OpenAI Faces Safety Scrutiny

OpenAI Scale Ranks Progress Toward ‘Human-Level’ Problem Solving

OpenAI has developed a five-level system to measure its progress towards creating advanced AI capable of surpassing human intelligence. During an all-hands meeting, OpenAI announced it is currently at Level 1 but nearing Level 2, called "Reasoners," which can perform basic problem-solving akin to a person with a doctorate. 

This system, intended to enhance understanding of AI safety and future developments, will be shared with investors and the public. OpenAI's CEO, Sam Altman, anticipates reaching artificial general intelligence within this decade.

BLOOMBERG

Synchron unveils OpenAI-powered BCI chat feature

Synchron has integrated OpenAI's generative AI into its brain-computer interface (BCI) system, enhancing its functionality with a new chat feature. The BCI, designed to decode neural signals to control digital devices, uses an endovascular approach to implant a device on the brain's motor cortex. 

This system allows individuals with severe paralysis to control devices hands-free. The integration of AI enables automated prompts for communication based on contextual inputs like user emotion. Synchron emphasizes user privacy by not sharing brain data with OpenAI. The company aims to make BCI technology more accessible and secure through its innovative delivery method and has strong financial backing.

MASS DEVICE

Meta researchers distill System 2 thinking into LLMs, improving performance on complex reasoning

Meta researchers have developed a technique called "System 2 distillation" to improve large language models (LLMs) on complex reasoning tasks without the need for slow, intermediate steps. This method allows LLMs to internalize the reasoning processes typically required for deliberate problem-solving, making them faster and more efficient. 

The researchers found that their approach can match or even exceed the accuracy of traditional System 2 methods while reducing computational costs. However, some tasks, like complex math reasoning, still require deliberate reasoning. This technique holds promise for optimizing mature LLM pipelines and freeing up time for more challenging tasks.

VENTUREBEAT

Japan's SoftBank acquires British AI chipmaker Graphcore

Japan's SoftBank Group has acquired British AI chipmaker Graphcore for an undisclosed amount, ending speculation about Graphcore's future. Despite once being seen as a rival to Nvidia, Graphcore struggled financially, cutting its workforce and shutting down operations in multiple countries. 

The acquisition provides Graphcore with resources to compete globally, although challenges remain, including the need for more investment from British pension funds. CEO Nigel Toon highlighted the significant investment required in the rapidly growing AI chip market. The deal also opens possibilities for collaboration with Arm Holdings, another SoftBank-owned chip designer.

REUTERS

Japan's SoftBank acquires British AI chipmaker Graphcore

Japan's SoftBank Group has acquired British AI chipmaker Graphcore for an undisclosed amount, ending speculation about Graphcore's future. Despite once being seen as a rival to Nvidia, Graphcore struggled financially, cutting its workforce and shutting down operations in multiple countries. 

The acquisition provides Graphcore with resources to compete globally, although challenges remain, including the need for more investment from British pension funds. CEO Nigel Toon highlighted the significant investment required in the rapidly growing AI chip market. The deal also opens possibilities for collaboration with Arm Holdings, another SoftBank-owned chip designer.

REUTERS

Microsoft’s Copilot AI now understands your terrible handwriting

Microsoft's Copilot AI in OneNote now includes a feature to read and analyze handwritten notes. This capability, currently in beta, allows users to write notes with a stylus, then use AI to summarize, ask questions, or generate to-do lists. Copilot can also convert handwriting into text for easier editing and sharing. 

Available to Copilot for Microsoft 365 subscribers and Copilot Pro users, this feature enhances productivity by accurately deciphering and formatting handwritten content. While the AI performs well with legible handwriting, its effectiveness with more challenging handwriting remains to be seen.

THE VERGE

OpenAI is plagued by safety concerns

OpenAI faces increasing scrutiny over its safety protocols, with recent reports highlighting rushed safety tests and internal discontent. An anonymous employee revealed that OpenAI celebrated product launches before ensuring their safety. Following the departure of key figures like cofounder Ilya Sutskever and safety researcher Jan Leike, concerns about prioritizing product development over safety have intensified. 

Despite public relations efforts and collaborations, such as with Los Alamos National Laboratory, to bolster its safety image, internal and external critics demand more transparency and stricter safety measures to prevent potential risks associated with advanced AI technologies.

THE VERGE

🧠RESEARCH

MAVIS: Mathematical Visual Instruction Tuning

The paper introduces a new approach to improving multi-modal large language models (MLLMs) for solving math problems involving visuals. It focuses on enhancing visual encoding of math diagrams, aligning diagrams with language, and boosting mathematical reasoning. MAVIS employs specialized datasets and training stages to achieve these improvements, addressing gaps in current MLLM capabilities.

Multimodal Self-Instruct: Synthetic Abstract Image and Visual Reasoning Instruction Using Language Model

The paper introduces a method for improving multimodal models' understanding of abstract images like charts and maps. Using large language models, they create synthetic visual reasoning instructions for various scenarios. This benchmark reveals current models' limitations in abstract image comprehension. Fine-tuning with these instructions enhances performance in visual reasoning tasks.

LLaVA-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models

The paper presents LLaVA-NeXT-Interleave, an advanced model that enhances large multimodal models (LMMs) to handle multi-image, video, and 3D tasks. It introduces the M4-Instruct dataset with 1.18 million samples across 14 tasks and 41 datasets. The model shows improved performance in multi-image and video benchmarks while maintaining single-image task efficiency and enabling cross-modality task transfers.

Video-STaR: Self-Training Enables Video Instruction Tuning with Any Supervision

The paper introduces Video-STaR, a novel self-training method for enhancing Large Vision Language Models (LVLMs) using any labeled video dataset. Video-STaR cycles between generating and fine-tuning instructions, improving video understanding and adapting to new tasks. It shows significant performance boosts, including a 10% improvement in general video QA and substantial gains in specific tasks like Kinetics700-QA and FineDiving.

ANOLE: An Open, Autoregressive, Native Large Multimodal Models for Interleaved Image-Text Generation

The paper introduces Anole, an advanced large multimodal model (LMM) designed for seamless interleaved image-text generation. Addressing limitations of previous models, Anole offers native integration without adapters, supports both image and text generation, and avoids separate diffusion models. Built from Meta AI's Chameleon, Anole uses a data and parameter-efficient fine-tuning strategy, achieving high-quality, coherent multimodal outputs. The model, training framework, and data are open-sourced.

🛠️TOP TOOLS

TypeflowAI - Easily share or embed AI tools into your website to boost SEO, increase your traffic and get more leads

RenderNet - Create AI characters with complete control and consistency

Reclaim AI - The smart(est) calendar app that optimizes everyone’s schedules for better productivity, collaboration, and work-life balance.

Dream Flare - Interactive AI videos and stories

Pieces Copilot+ - on-device AI coding assistant that boosts developer productivity by helping you solve complex development tasks through a contextual understanding of your entire workflow.