No more mundane work

PLUS: Biases and moral self-correction in AI

Author

Ben Tossell
February 17, 2023

Hey folks! This is Ben’s Bites - an AI a day keeps the unhinged AGI away.

Off on a friend’s stag do (bachelor party) today/this weekend. So I’m in for a pretty heavy weekend I’d imagine, I hope yours is calmer 🧘.

p.s Zapier’s No-Code Contest has an AI category 🔥 - enter, show what you’ve built and potentially win $5k!

Poll: What's your interest in AI?

Login or Subscribe to participate in polls.

🤌 Ben's Picks

1/

OpenAI is working to address complaints from users about politically biased or offensive outputs from ChatGPT. They use a two-step (not dance-related 🩰 ) process to create ChatGPT, pre-training and fine-tuning. They collaborate with reviewers to maintain a feedback loop and align AI with human values and are working to address biases.

OpenAI wants to improve the default behaviour of ChatGPT, make it customisable, and include public input on topics like system behaviour and deployment policies. It’ll be very interesting to see what these customisations look like!

2/

Anthropic also added one more weapon to their bias busters kit yesterday by releasing a paper on the ability to reduce bias from LLMs, and they have some good news.

  • The models start to show the capability to self-correct at around 22B parameters and typically improve with increasing model size and RLHF training.

  • Two techniques that work: instruction following (IF) prompt and chain of thought (CoT) reasoning. IF reduces bias by 43% and CoT by 84%.

The concern, though, is that they can sometimes over-correct if not checked. Check out this thread where they explain the three experiments they did with IF and CoT. The paper is here.

3/

Many document editors are adding AI to their stack these days. But let’s be honest, most of them are just AI autocomplete built into a text editor.

Well, that doesn’t seem to be the case for Coda. With their Alpha version of Coda AI, they’re merging the building blocks of Coda docs with GPT3’s capabilities. This modular approach removes the mundane work of organising AI-generated text and makes iterating way smoother. What’s more, you ask? It connects with 600+ other apps to integrate in your existing workflow. Definitely getting on that waitlist.

HyperWrite (Sponsor) is the leading AI writing assistant that (learns from your writing) and (adapts to your context).

The Chrome extension (works on your favourite sites) and understands what you're doing, so the suggestions it provides are relevant and useful.

Write emails, memos, docs, and more with ease. Smart autocomplete, paragraph generation and rephrasing everywhere.

🛠️ Cool Tools

  • Analogenie - Explain better with AI-generated analogies & metaphors. (link)

  • Buzzing AI - Generate stunning content 10x faster. (link)

  • Krikey - Use AI to generate avatar animations. (link)

  • Where to - Plan your next getaway. (link)

  • DBsensei - Generate arduous database queries in a snap. (link)

  • 10web’s AI Assistant for WordPress. (link)

  • Hypeless - Generate premium quality branding and design assets for your business. (link)

  • AI hotel review - Get the summary of TripAdvisor reviews of a hotel. (link)

  • AI Prompt Generator - Generate your text to image prompts. (link)

  • Airtable text to template - Describe your use case and generate tables and fields. (link)

  • Scribbled diffusion - Web app to draw a rough sketch, add a text prompt, and generate images based on your sketch. (link)

  • Ask Cal documentation any questions from password reset to advanced plans! (link)

  • Genius Sheets - Talk to your data with new conversational AI. (link)

  • Definite - A better way to do analytics. (link)

  • Masterpiece Studio - 3D character animation just got a lot easier. (link)

  • Type - The AI-powered document editor that helps you write remarkably fast. (link)

🤓 Miscellaneous

  • German startup Kern AI nabs seed funding for modular NLP development platform. (link)

  • How can we make sure that everyone has access to AI? Can small models outperform large models? With Stability AI’s Emad Mostaque. (link)

  • Two frightening things about AI before breakfast. (link)

  • How this sales startup is using GPT-3 tech to boost lead qualification. (link)

  • A mere wrapper on top of GPT - Defensibility & competition. (link)

  • UK’s conservation AI makes a huge leap detecting threats to endangered species across the globe. (link)

  • Baidu - Why I might be adding to this AI leader. (link, without paywall here)

  • US issues a declaration on responsible use of AI in the military. (link)

  • The AI arms race is on. Start worrying. (link, without paywall here)

  • 'It's hurting like hell' - AI companion users are in crisis, reporting sudden sexual rejection. (link)

  • Smells a little bit like AI winter? (link)

  • The 1000x developer - a16z podcast with Amjad Masad. (link)

  • The speed and power of AI in game development. (link)

  • We’re reliving the start of the smartphone era, this time with AI. (link)

  • New Balance's Onur Yüce Gün on how Runway's AI capabilities are revolutionising design. (link)

  • Revenge of the chatbots. (link)

  • GPT labs launch new open AI libraries. (link)

🎓️ Learn

  • How to create free GPU-powered ML apps in Python in just 5 minutes. (link)

  • OpenAI GPT 3.5 AI assistant with Langchain and Pinecone. (link)

👋 Too many links?! I created a database for all links mentioned in these emails. Refer 1 friend using this link and I'll send it over.

🔬 Research

  • PRedItOR from Adobe - Text-guided image editing with diffusion prior. (link)

  • MultiDiffusion - Fusing diffusion paths for controlled image generation. (link)

  • T2I-Adapter from Tencent - Learning adapters to dig out more controllable ability for text-to-image diffusion models. (link)

  • Text-driven visual synthesis with latent diffusion prior. (link)

  • Exploring the limits of ChatGPT for query or aspect-based text summarization. (link)

  • Do we still need clinical language models? (link)

  • LEVER from Meta - Learning to verify language-to-code generation with execution. (link)

  • Efficiency 360 by Microsoft - Efficient vision transformers. (link)

  • PersonNeRF from Google - Personalized reconstruction from photo collections. (link)

  • Pix2pix 3D - 3D aware conditional image synthesis. (link)

  • LEALLA - Learning lightweight language-agnostic sentence embeddings with knowledge distillation. (link)

  • ACE-VC from Nvidia - Adaptive and controllable voice conversion using explicitly disentangled self-supervised speech representations. (link)

  • Block Data Representations (BDR) from Microsoft - A framework for exploring and evaluating a wide spectrum of narrow-precision formats for deep learning. (link)

  • MINOTAUR from Meta - Multi-task video grounding from multimodal queries. (link)

  • Cross-modal distillation for flood extent mapping from Google. (link)

📰 Unclassifieds

  • Buildspace announces S3 of Nights & Weekends - a 6-week sprint of building and shipping. Applications close on Feb 28th. Learn more →

  • Farmwise is hiring a Machine Learning Intern in Paris. Learn more →

  • Roblox is hiring a Principal Deep Learning Engineer. Learn More →

  • AI Hackathon in London, April 2023. Learn more →

Have a product, service, job, event, newsletter, app, book, movie, tool, or anything you'd like to share with over 26k subscribers? 

🖼️ AI images of the day

From Ben’s Bites Discord

🤗 Share Ben's Bites

Send this to 1 AI-curious friend and receive my AI project tracker database!

or copy/paste this link: https://bensbites.beehiiv.com/subscribe?ref=PLACEHOLDER

⭐️️ How did we do?

How was today's email?

Login or Subscribe to participate in polls.

⭐️️ REAL REVIEWS

Join the conversation

or to participate.