FOD#40: Trust and Responsibility in the Age of AI

Next Week in Turing Post:

• Wednesday, Token 1.19: Explainable AI.

• Friday, Guest post about Modern AI and its Unhyped Capabilities.

I’m a professional journalist who has worked in tech for a few decades. Since the bloom of social media, it’s been a tough times for journalism as so many voices appeared and the cacophony was deafening.

AI-generated content adds even more infotrash. But surprisingly enough, I think that AI is here to bring us back to the quality of journalism, both as a risk factor and as an enabler. Two other articles from last week made me think about this. The first one from Semafor introduced their new offering: Semafor's Signals. Using Microsoft and OpenAI tools, Signals provides diverse insights on global news, adapting to digital shifts and AI challenges. Reed Albergotti, the technology editor of Semafor, wrote:

I totally agree with this point; in the age of AI, there is nothing more important than to have voices/media whom you trust. And here comes the professional journalist. The responsible journalist. Who is this person? That’s a tricky question since ‘responsible’ in the context of AI becomes a joke. In the era of AI, the question of what constitutes responsible journalism gains new dimensions. Last week, for example, Goody-2 was launched, a chatbot designed to avoid misinformation by providing vague responses and being “responsible”.

AI can be dangerous and used as – for example – for audio-jacking, but in terms of journalism, it offers a bunch of amazing tools that significantly enhance reporting, editing, and content distribution. For instance, automated fact-checking platforms like Full Fact in the UK utilize AI to quickly verify claims made in public discourse, enhancing the accuracy and reliability of news reporting. Data journalism has also been revolutionized by AI, with tools like Datawrapper allowing journalists to create interactive charts and visualizations without extensive coding knowledge. Moreover, The New York Times' experiment with personalized article recommendations showcases how AI can curate content tailored to individual readers' interests, potentially increasing engagement and subscription rates.

Last week, The Platformer was also contemplating the future of the web and journalism.

And here I agree again: no one is coming to save journalism, but with AI – as risk and enabler – journalism can finally return to its essence. Reflecting on the journey of journalism through the digital and AI revolutions, it becomes clear that while challenges abound, the essence of journalism as a pillar of democracy remains intact. Embracing AI thoughtfully allows journalism to return to its core mission: to inform, educate, and hold power to account – to have responsibility – thereby ensuring that it continues to thrive as a trusted guide in an increasingly complex world.

Twitter Library

News from The Usual Suspects ©

Vesuvius and Pompeii

• Using AI, three students read more than 2,000 Greek letters buried in Pompeii in 79 AD and won $700,000.

Roblox

• The game company introduced AI-powered real-time chat translations in 16 languages.

Sam Altman

• Sam Altman seeks $5-7 trillion for global AI chip production expansion. (That’s a lot…). Gary Marcus offers 7 reasons why the world should say no (that’s not that many…)

OpenAI meanwhile

• OpenAI hits $2 billion annual revenue being among fastest-growing tech firms.

• OpenAI is working on two AI agents to automate diverse tasks.

Microsoft

• A summary of recent research from Microsoft and around the world that can help us create a new and better future of work with AI.

NVIDIA

• Nvidia wants to start producing personalized chips for AI companies

Google

• Google rebranded Bard to Gemini. Read Ethan Mollick about

  mixed strengths and weaknesses.

   

A few

• Nvidia, OpenAI, Microsoft, and nearly 200 other companies joined the US AI Safety Institute Consortium (AISIC) to support the safe development and deployment of generative AI.

The freshest research papers, categorized for your convenience

Large Language Models and Their Enhancements

• More Agents Is All You Need: Demonstrates how increasing the number of agents in LLMs enhances performance through a sampling-and-voting method. Read the paper

• Tag-LLM: Adapts general-purpose LLMs to specialized domains using custom input tags for domain- and task-specific behavior. Read the paper

• BiLLM: Introduces a 1-bit post-training quantization approach for LLMs, maintaining high performance under ultra-low bit-widths. Read the paper

• Direct Language Model Alignment from Online AI Feedback: Enhances model alignment through online feedback, improving exploration and performance. Read the paper

• The Hedgehog & the Porcupine: Presents Hedgehog, a learnable linear attention mechanism that mimics softmax attention in Transformers. Read the paper

• An Interactive Agent Foundation Model: Proposes a novel AI framework for domains like Robotics and Healthcare, integrating visual autoencoders, language modeling, and action prediction. Read the paper

• DeepSeekMath: Pushes the limits of mathematical reasoning in open language models. Read the paper

• SELF-DISCOVER: Enables LLMs to self-compose reasoning structures for complex problem-solving. Read the paper

• Can Mamba Learn How to Learn?: Compares the in-context learning abilities of State-Space Models against Transformer models. Read the paper

• Scaling Laws for Downstream Task Performance of Large Language Models: Investigates the impact of pretraining data size and type on LLMs' downstream performance. Read the paper

• Rethinking Optimization and Architecture for Tiny Language Models: Studies optimizing tiny language models for mobile devices. Read the paper

• Shortened LLaMA: Explores depth pruning as a method for improving LLM inference efficiency. Read the paper

Multimodal and Vision-Language Models

• λ-ECLIPSE: Achieves personalized text-to-image generation by leveraging CLIP's latent space. Read the paper

• SPHINX-X: Proposes an advanced series of Multi-modality Large Language Models focusing on model performance and training efficiency. Read the paper

• SpiRit-LM: Integrates text and speech in a multimodal foundation language model for improved semantic understanding and expressivity. Read the paper

• Question Aware Vision Transformer for Multimodal Reasoning: Embeds question awareness within the vision encoder for enhanced multimodal reasoning. Read the paper

• EVA-CLIP-18B: Scales CLIP to 18 billion parameters, achieving significant performance improvements in image classification. Read the paper

Robotics, Autonomous Systems, and Interactive Agents

• Driving Everywhere with Large Language Model Policy Adaptation: Enables adaptation to local traffic rules for autonomous vehicles using LLMs. Read the paper

• Offline Actor-Critic Reinforcement Learning Scales to Large Models: Demonstrates that offline actor-critic reinforcement learning can effectively scale to large models. Read the paper

Web Navigation, Conversational Systems, and Real-World Applications

• WebLINX: Introduces a benchmark for conversational web navigation, highlighting the need for models that adapt to new web environments. Read the paper

• In-Context Principle Learning from Mistakes: Enhances LLM learning by inducing mistakes and reflecting on them to extract task-specific principles. Read the paper

• Multi-line AI-assisted Code Authoring: Presents CodeCompose, an AI-assisted code authoring tool offering both single-line and multi-line inline suggestions. Read the paper

Time Series Forecasting, Object Detection, and Other Innovations

• Lag-Llama: Introduces a foundation model for univariate probabilistic time series forecasting, showcasing strong zero-shot generalization. Read the paper

• InstaGen: Enhances object detection by training on synthetic datasets generated from diffusion models. Read the paper

• Implicit Diffusion: Presents an algorithm optimizing distributions defined by stochastic diffusions for efficient sampling. Read the paper

• Memory Consolidation Enables Long-Context Video Understanding: Proposes a method enhancing video understanding by consolidating past activations. Read the paper

• Grandmaster-Level Chess Without Search: Trains a transformer model to achieve grandmaster-level chess performance without explicit search algorithms. Read the paper

Code Representation and Quantization Techniques

• CODE REPRESENTATION LEARNING AT SCALE: Introduces CODESAGE, an advanced model for code representation learning with a two-stage pretraining scheme. Read the paper

Interpretability and Foundation Models

• Rethinking Interpretability in the Era of Large Language Models: Examines the role of interpretability with the advent of LLMs, advocating for a broader scope in interpretability. Read the paper

In other newsletters

• For those curious about the most transformative innovation to semiconductor manufacturing since EUV – Hybrid Bonding by Semianalysis

• All My Thoughts After 40 Hours in the Vision Pro by Wait But Why

• A funny take on a paper that claimed that AGI had been achieved by Jack Clark from Anthropic (Import AI newsletter)

• The promise and challenges of crypto + AI applications by Vitalik Buterin