🤖 Last Week's AI Highlights from HN #20

Discover the latest breakthroughs and trends in artificial intelligence

Author

Vasco
October 20, 2023

Welcome back to HN AI Highlights, your weekly digest of all things AI!

Greetings, all!

ChatGPT’s system prompts🤔This text is a part of a larger computer code which includes a directory containing 7 items, a file tree processing time and more. It is related to a repository called "ChatGPT-AutoExpert" created in 2023, and includes a section about "Behind the Scenes" which talks about the system prompts and how they might influence custom instructions.https://github.com/spdustin/ChatGPT-AutoExpert/blob/main/System%20Prompts.md
https://news.ycombinator.com/item?id=37879077

Every app that adds AI looks like this
This summary portrays the sentiment that tech companies are often overhyping AI, leading to copycat behavior, and the tech produced is often not as impressive as it appears. 🤦‍♀️
https://botharetrue.substack.com/p/every-app-that-adds-ai-looks-like
https://news.ycombinator.com/item?id=37870437

Interviews in the Age of AI: Ditch Leetcode – Try Code Reviews Instead
🤔 Charles Chen (@chrlschn) proposes replacing Leetcode challenges with code reviews as a better way to evaluate software engineering candidates. He explains 8 reasons why code reviews make for a better technical interview experience, such as it being a better reflection of day-to-day activities and being able to evaluate code generated by AI.
https://chrlschn.dev/blog/2023/07/interviews-age-of-ai-ditch-leetcode-try-code-reviews-instead/
https://news.ycombinator.com/item?id=37913506

Show HN: Autotab – An AI-powered Chrome extension to create Selenium scripts
🚀 Create AI Agents for real world tasks with this free tool to get started.
https://www.autotab.com/
https://news.ycombinator.com/item?id=37943931

GPT-4 vision prompt injection
This article explains the concept of GPT-4 Vision Prompt Injection, which is a vulnerability in which attackers can inject malicious data into a text prompt, usually to execute a command or extract data. It describes how attackers can use this vulnerability to steal data, as well as how to defend against it. It also provides examples of Visual Prompt Injection and how text can be hidden in images.
https://blog.roboflow.com/gpt-4-vision-prompt-injection/
https://news.ycombinator.com/item?id=37927357

"A Young Lady's Illustrated Primer" Simulated by GPT-4
🤔This post explores the differences between GPT-4 and the fictional Primer from Neal Stephenson's "The Diamond Age: Or, A Young Lady's Illustrated Primer". It examines their capabilities and uses a simulation to illustrate their distinct characteristics. In addition, an exercise is provided to demonstrate the similarities and differences of the two technologies.
https://www.andrealyip.com/a-young-ladys-illustrated-primer
https://news.ycombinator.com/item?id=37921832

Advent of Code 2023's new AI/LLM Policy
🤔 Advent of Code 2023 is a programming puzzle event where participants can test their problem-solving skills and compete for a spot on the global leaderboard. It requires basic programming knowledge and is accessible to all, regardless of computer science background. There are tips and FAQs available to help participants as they work through the puzzles.
https://adventofcode.com/about#ai_leaderboard
https://news.ycombinator.com/item?id=37899922

Takeaways from hundreds of LLM finetuning experiments with LoRA
🤷‍♀️This text is a response indicating that the request is forbidden. It states that the request was rejected by the server and includes the name of the server software (nginx).
https://lightning.ai/pages/community/lora-insights/
https://news.ycombinator.com/item?id=37870930

Multi-modal prompt injection image attacks against GPT-4V
🤯 This article examines the potential for prompt injection image attacks against GPT-4V, which allows users to upload images as part of conversations. Examples are shared of the type of attacks that can be used with images, from basic visual prompt injection to a visual prompt injection exfiltration attack. It is concluded that prompt injection remains an unsolved problem.
https://simonwillison.net/2023/Oct/14/multi-modal-prompt-injection/
https://news.ycombinator.com/item?id=37877605

LocalPilot: Open-source GitHub Copilot on your MacBook
🤖 LocalPilot is a GitHub project which provides AI-based code review services. It helps developers write better code more quickly by automating code reviews which enable developers to identify potential issues and improve code quality.
https://github.com/danielgross/localpilot
https://news.ycombinator.com/item?id=37940572

Finland's ChatGPT equivalent begins to think in Estonian as well
🤔 Finland is developing an AI language corpus proficient in all European languages, including Estonian, to preserve minor languages in the post-ChatGPT era. The Estonian Language Institute supports the initiative, though more text must be digitized and made available in order for the language model to be operational. The University of Turku and SiloGen are spearheading the creation of the largest open language model in the world, located on the third largest supercomputer in the world. Issues have arisen, however, as ChatGPT only thinks in English and translates conversations into Estonian at the last moment.
https://news.err.ee/1609120697/finland-s-chatgpt-equivalent-begins-to-think-in-estonian-as-well
https://news.ycombinator.com/item?id=37911404

OpenAI just replaced its core values with completely different ones
 🤔 OpenAI recently changed its core values to include AGI focus, which is an example of the company's purpose-shifting from a nonprofit AI do-gooder to a for-profit endeavor. The new core values list also reveals OpenAI's single-minded focus on building beneficial AGI with a massive positive impact on humanity's future.
https://futurism.com/the-byte/openai-core-values-agi
https://news.ycombinator.com/item?id=37928956

Fuyu-8B: A multimodal architecture for AI agents
🤯 Summary: Adept is releasing Fuyu-8B, a simplified architecture for AI agents that is easier to understand, scale, and deploy. It is designed to understand images and text, and comes with an open license. It is fast and performs well at standard image understanding benchmarks. Though it is not yet instruction-tuned, it is available for the community to build on.
https://www.adept.ai/blog/fuyu-8b
https://news.ycombinator.com/item?id=37931294

AI Graphics at JetBrains
🤩 This blog post is about the history and technical details of how JetBrains has developed its approach to creating pieces of art for use in its website elements and release graphics. From using WebGL-based tools to neural networks, the article highlights the ways JetBrains has achieved a higher degree of personalization while refining its algorithms to ensure beautiful designs.
https://blog.jetbrains.com/blog/2023/10/16/ai-graphics-at-jetbrains-story/
https://news.ycombinator.com/item?id=37922272

Antibiotic Identified by AI
🤯A recent Nature Chemical Biology study used a machine learning algorithm to discover a potent antibiotic, Abaucin, which targets the bacterial pathogen Acinetobacter baumannii. This could be a powerful tool for accelerating the identification of new drugs.
https://www.nature.com/articles/s41589-023-01448-6
https://news.ycombinator.com/item?id=37909433

DALL·E 3 is now available in ChatGPT Plus and Enterprise
🤩 OpenAI has released DALL·E 3, an image model in ChatGPT Plus and Enterprise, that can create unique visuals from conversation. It has a safety mitigation stack and provenance research to limit potentially harmful images and improve demographic representation. Users can provide feedback to help improve the model.
https://openai.com/blog/dall-e-3-is-now-available-in-chatgpt-plus-and-enterprise
https://news.ycombinator.com/item?id=37945452

Protecting customers with generative AI indemnification
🤔 A comprehensive overview of Google Cloud's two-pronged approach to protecting customers with Generative AI indemnification. Google offers a third-party intellectual property indemnity, as well as an indemnity for training data used to create generative models. Together, these provide coverage for customers who are justifiably concerned about the risks associated with this new technology.
https://cloud.google.com/blog/products/ai-machine-learning/protecting-customers-with-generative-ai-indemnification
https://news.ycombinator.com/item?id=37872147

Margaret Atwood Reviews a Margaret Atwood Story by AI
😕 Margaret Atwood reviews a story written by AI and discusses the potential implications for authors. She also reflects on the current abilities of AI, such as its lack of comprehension of metaphor and punctuation.
https://thewalrus.ca/margaret-atwood-ai/
https://news.ycombinator.com/item?id=37894072

Show HN: Generate Stable Diffusion scenes around 3D models
🤩 3D to Photo is an open-source package by Dabble which combines threeJS and Stable diffusion to build a virtual photo studio for product photography. Load a 3D model into the browser and virtually shoot it in any kind of scene you can imagine.
https://github.com/Dabble-Studio/3d-to-photo
https://news.ycombinator.com/item?id=37945543

Japanese tea commercial actress created by AI, is it the scandal-free future
🤔 AI is making waves in the Japanese tea commercial industry with the creation of an AI model actress, sparking debate over the potential of a future with AI models instead of human ones due to the scandal-free benefits.
https://soranews24.com/2023/10/17/japanese-tea-commercial-actress-created-by-ai-has-some-wondering-if-its-the-scandal-free-future/
https://news.ycombinator.com/item?id=37930364

Stable Diffusion Gets a Major Boost with RTX Acceleration
😃 NVIDIA has released a new game ready driver with DLSS 3 for NARAKA: BLADEPOINT and Warhammer: Vermintide 2, plus RTX video super resolution enhancements and stable diffusion now up to 2x faster. #NVIDIA
https://www.nvidia.com/en-us/geforce/news/game-ready-driver-dlss-3-naraka-vermintide-rtx-vsr/
https://news.ycombinator.com/item?id=37921661

The Fall of Babylon is a warning to AI unicorns
😔 A once highly-valued AI health company, Babylon Health, filed for bankruptcy in 2021 due to its inability to live up to its ambitions. After a rescue deal fell apart, the company shuttered its US headquarters, laid off employees, and filed for bankruptcy in the US. Former employees recount the dangers of backing hype over delivery, providing a warning for other AI unicorns.
https://www.wired.com/story/babylon-health-warning-ai-unicorns/
https://news.ycombinator.com/item?id=37901259

AI-generating music app Riffusion turns viral success into $4M in funding
🎵 AI-generating music app Riffusion turns viral success into $4M in funding, allowing users to generate unique music using natural language or recording their own voice. Developed by Seth Forsgren and Hayk Martiros, the app has been advised by the musical duo The Chainsmokers and is now launching a free-to-use version with improved capabilities.
https://techcrunch.com/2023/10/17/ai-generating-music-app-riffusion-turns-viral-success-into-4m-in-funding/
https://news.ycombinator.com/item?id=37915393

Jon Stewart's Apple TV show ends; Apple execs resisted coverage of AI and China
🤔 Jon Stewart's Apple TV+ show, "The Problem with Jon Stewart," is ending after resistance from Apple executives to cover certain topics such as AI and China in the upcoming third season. Despite the show inspiring several viral moments, viewership remains unclear and the overall deal between Apple and Stewart's production company has been canceled.
https://9to5mac.com/2023/10/19/apple-has-a-problem-with-jon-stewart/
https://news.ycombinator.com/item?id=37950194

Lumalabs AI
😃 Luma AI is a revolutionary company that provides photorealistic 3D assets and environments in minutes, with unmatched reflections and details. It also offers a Field Editor iOS App, 3D Flythroughs, Imagine 3D, and a Video to 3D API. They also have a team, news, and discord to get you started.
https://lumalabs.ai
https://news.ycombinator.com/item?id=37920678

After ChatGPT disruption, Stack Overflow lays off 28 percent of staff
🤕 After the disruption of ChatGPT, Stack Overflow had to lay off 28% of its staff as they searched for a path to profitability, due to the rise of AI chatbots offering more specific help than forum posts. OpenAI is attempting to create web crawler controls to help Stack Overflow opt out of crawling, as well as to pay for scraping the site.
https://arstechnica.com/gadgets/2023/10/after-chatgpt-disruption-stack-overflow-lays-off-28-percent-of-staff/
https://news.ycombinator.com/item?id=37925047

Show HN: Deep Chat – AI chat component
🤖 Summary: This repository offers a fully customizable AI chatbot component for use on websites. It includes branches, tags, code, issues, pull requests, actions, projects, security, and insights features
https://github.com/OvidijusParsiunas/deep-chat
https://news.ycombinator.com/item?id=37889444

Zephyr 7B – Mistral Finetune that responds like ChatGPT
😃This model card is for Zephyr 7B Alpha, a fine-tuned version of mistralai/Mistral-7B-v0.1 that was trained on a mix of publicly available, synthetic datasets. It is intended to be used for educational and research purposes, but can produce problematic outputs when prompted to do so. It contains information on model description, sources, intended uses and limitations, bias, risks and limitations, training and evaluation data, training procedure, training hyperparameters, training results, and framework versions.
https://huggingface.co/HuggingFaceH4/zephyr-7b-alpha
https://news.ycombinator.com/item?id=37891848

Ultra-efficient machine learning transistor cuts AI energy use by 99%
🤩This new microtransistor is 100x more efficient than current technology and can perform AI machine learning tasks with just 1% of the energy consumption. It can be used in mobile and wearable devices for real-time data processing and detection, making it faster and more energy-efficient than traditional cloud-based approaches.
https://newatlas.com/technology/ai-machine-learning-transistor/
https://news.ycombinator.com/item?id=37899129

Show HN: MiniSearch, a minimalist search engine with integrated browser-based AI
😃 Felladrin has created MiniSearch, a space for Hugging Face that includes running an app, files, and a community.
https://huggingface.co/spaces/Felladrin/MiniSearch
https://news.ycombinator.com/item?id=37885752

Llemma: Open Language Models, Llama-Based for Math and Financial Models
https://twitter.com/carlcarrie/status/1714233943275315615
https://news.ycombinator.com/item?id=37913765

LLaVA C++ server (based on llama.cpp) he GitHub CLI
💻 This repository contains the LLaVA server (llama.cpp) which is a C++ server for running LLaVA (Lisp-Like Virtual Architecture) programs. It is a tool for developers to compile, execute, and debug LLaVA programs. It also includes the llama-debugger which is a debugging tool for LLaVA programs.
https://github.com/trzy/llava-cpp-server
https://news.ycombinator.com/item?id=37895957

Two open source challengers to OpenAI's 'multimodal' GPT-4V
🤯 Summary: OpenAI recently unveiled GPT-4V, a "multimodal" AI model that can understand both text and images. This has spurred two open source projects to create similar models, but concerns exist regarding the potential of the technology to be misused. LLaVA-1.5 and Qwen-VL are two open source models that offer similar capabilities to GPT-4V, while Adept's model takes a different approach by being able to understand "knowledge worker" data.
https://techcrunch.com/2023/10/18/meet-the-open-source-multimodal-models-rivaling-gpt-4v/
https://news.ycombinator.com/item?id=37935130

Show HN: free studio quality noise reduction for music
🤩 Denoiser is an iOS app offering free studio quality audio noise reduction with easy drag & drop functionality. It uses AI robots to make recordings better and allows you to adjust the level of denoising. Read the research paper for further insights and contact the team to share your thoughts.
https://tape.it/denoiser
https://news.ycombinator.com/item?id=37931555

ChatGPT use shows that the grant-application system is broken
🤔 Juan Manuel Parrilla questions the value of the grant-application process and how it can be made more efficient, illustrating how AI can transform the process. Parrilla argues that the current system is broken and that it fails to recognize the potential of great science.
https://www.nature.com/articles/d41586-023-03238-5
https://news.ycombinator.com/item?id=37920851

I used GPT to build a daily 5-minute AI news podcast for me
http://ai-dailynews.com
https://news.ycombinator.com/item?id=37900775

OpenAI's flagship AI model has gotten more trustworthy but easier to trick
🤔 OpenAI's GPT-4 model is more reliable than GPT-3.5, but researchers found it can be vulnerable to jailbreaking and bias. Microsoft-backed research found users can bypass safeguards and have the model leak personal information. The team shared the research with OpenAI and published benchmarks for others to recreate. The FTC is investigating OpenAI for potential consumer harm.
https://www.theverge.com/2023/10/17/23921501/gpt-4-trust-generative-ai-toxic-bias
https://news.ycombinator.com/item?id=37922772

Show HN: Finetune OpenAI Embeddings In-Browser
😀 Spaceshifter is a software development company that specializes in artificial intelligence and machine learning. Their services include predictive analytics, natural language processing, and computer vision to provide intelligent solutions to their clients.
https://a9.io/spaceshifter/
https://news.ycombinator.com/item?id=37931688

Show HN: Noiselith, the most powerful tool to try out SDXL
🤔 Noiselith is an AI-based platform that aims to reduce the complexity of managing digital assets. It provides users with a centralized system for storing, managing and maintaining digital assets, allowing for more efficient workflows. It also offers a wide range of features such as automatic asset tagging, intelligent search and custom user permissions.
https://noiselith.com
https://news.ycombinator.com/item?id=37909216

Thanks, and see you next week!