TLDR AIOctober 30, 2023

TLDR AI 2023-10-30

5 sections15 entrieslinks blockcomplete parse

Parse notes

Subscription/account footer truncated before parsefooter_truncated

01HEADLINES & LAUNCHES

GOOGLE INVESTS $2B INTO ANTHROPIC

Google has committed to a substantial investment in Anthropic, a competitor to OpenAI, earmarking up to $2 billion for the AI startup.

cnbc.com2 min read

FIRST DIRECTOR OF THE FRONTIER MODELS FORUM NAMED AND BIG TECH ADDS $10M TO SAFETY INITIATIVE FUNDING

Chris Meserole is the first director of the Frontier Models Forum. Many large tech companies have added funding to the AI safety fund.

blog.google4 min read

AMAZON ADS LAUNCHES AI-POWERED IMAGERY FOR ADVERTISERS

Amazon Ads has introduced an AI-powered image generation feature in beta that creates lifestyle and brand-themed images to improve ad performances. The tool, which requires no technical expertise, enables brands to create more engaging ads, reportedly increasing click-through rates by 40%. It will be rolled out progressively to advertisers based on customer feedback.

aboutamazon.com2 min read

02RESEARCH & INNOVATION

EMULATE TRAINING A LARGE MODEL JUST BY TRAINING A SMALL ONE

Emulator Fine Tuning (EFT) comes from the direct preference optimization group and explores what happens if you fine tune a small model and then project that fine tuning onto a large model. This counterintuitive strategy works shockingly well.

arxiv.org28 min read

IDENTIFYING OUT-OF-DISTRIBUTION DATA

ML models can sometimes wrongly recognize unfamiliar data, thinking they've seen it before. A new method, NNGuide, helps models make better decisions, especially when identifying unknown data.

arxiv.org14 min read

TOWARDS MONOSEMANTICITY: DECOMPOSING LANGUAGE MODELS WITH DICTIONARY LEARNING

Researchers use sparse autoencoders to tackle neural networks' polysemanticity, extracting clearer features for improved model understanding. These features influence outputs and exhibit complex interactions.

transformer-circuits.pub60 min read

03ENGINEERING & RESOURCES

AUDIOFLARE

An all-in-one AI audio playground using Cloudflare AI Workers to transcribe, analyze, summarize, and translate any audio file.

github.comGitHub repository

A SYSTEM FOR EVALUATING LARGE LANGUAGE MODELS

This study presents "JudgeLM", a new method to efficiently evaluate Large Language Models in versatile situations. The authors developed a vast dataset and benchmarking system.

github.comGitHub repository

MULTI-MODAL PROMPT INJECTION IMAGE ATTACKS AGAINST GPT-4V

The new GPT-4V allows for interactive conversations with image inputs but also introduces new vulnerabilities through visual prompt injection attacks, enabling exfiltration of data and execution of hidden instructions encoded in images.

simonwillison.net3 min read

04MISCELLANEOUS

DEEP LEARNING IN RUST (BOOK)

Rust is a popular systems programming language with strong memory safety. Burn is a framework built on these principles for efficient ML. Importantly, it offers compatibility with the Onyx runtime.

burn.dev

THANKS TO AI, THE FUTURE OF PROGRAMMING MAY INVOLVE YELLING IN ALL CAPS

OpenAI's prompts reveal many quirks about prompt engineering. For example, the prompts use 'please' a lot and important instructions are often all capitalized. Capitalizing instructions likely works as the models were trained on huge numbers of examples where responses clearly paid more attention to capitalized sentences. This could result in a future where people shout at their computers to get them to work better.

arstechnica.com3 min read

IS AI ALIGNMENT ON TRACK? IS IT PROGRESSING... TOO FAST?

AI alignment is a complex issue that lacks concrete benchmarks or numbers for measurement. Predictions of AI-induced doom rely on narratives rather than rigorous models. The absence of hard numbers makes it challenging to assess the state of AI alignment.

guzey.com8 min read