AI Edition - It's AI Time - 142 [04-09-2024]

AI Edition - It's AI Time - 142

Why AI can't spell 'strawberry'

Large language models sometimes fail at tasks like counting letters due to their tokenization methods. This highlights limitations in LLM architecture that affect their understanding of text. Nevertheless, advancements continue, such as OpenAI's Strawberry for improved reasoning and Google DeepMind's AlphaGeometry 2 for formal math.

First MLPerf benchmarks for Nvidia Blackwell, AMD, Google, Untether AI

Nvidia's new Blackwell chip demonstrated top per GPU performance in MLPerf's LLM Q&A benchmark, showcasing significant advancements with its 4-bit floating-point precision. However, competitors like Untether AI and AMD also showed promising results, particularly in energy efficiency. Untether AI's speedAI240 chip, for instance, excelled in the edge-closed category, highlighting diverse strengths across new AI inference hardware.

Can AI Scaling Continue Through 2030?

AI training is growing at an unprecedented 4x per year, outpacing mobile adoption and genome sequencing rates from past technological expansions. Research indicates scaling AI training could feasibly continue until 2030, constrained primarily by power availability and chip manufacturing capacity. Training runs up to 2e29 FLOP may become possible, marking substantial progress akin to the leap from GPT-2 to GPT-4, provided hundreds of billions are invested. Data scarcity and latency also pose challenges, but these may be surmountable through multimodal and synthetic data generation strategies and advanced network topologies.

AI-Implanted False Memories

A study from MIT Media Lab found that generative chatbots powered by large language models significantly increased the formation of false memories during simulated crime witness interviews, inducing over three times more immediate false memories than a control group.