🤖 Last Week's AI Highlights from HN #17

Discover the latest breakthroughs and trends in artificial intelligence

Author

Vasco
September 22, 2023

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

Greetings, all!

DALL·E 3
🤩 A leap forward in text-to-image systems, DALL·E 3 enables users to translate ideas into accurate images easily. It is built natively on ChatGPT, allowing for tailored and detailed prompts and the ability to tweak with just a few words. Safety measures have been taken to limit violent, adult, or hateful content. Provenance classifiers are being tested to help identify images generated by DALL·E 3.
https://openai.com/dall-e-3
https://news.ycombinator.com/item?id=37586900

Data accidentally exposed by Microsoft AI researchers
😱 Microsoft AI researchers accidentally revealed a whopping 38TB of data, prompting a call for cloud security with Wiz's platform. Wiz empowers security teams to detect and fix cloud misconfigurations, prioritize cloud risks, secure containers, and more. Wiz also offers solutions for specific verticals such as healthcare, government, and financial services, plus resources and community projects to help with cloud security.
https://www.wiz.io/blog/38-terabytes-of-private-data-accidentally-exposed-by-microsoft-ai-researchers
https://news.ycombinator.com/item?id=37556605

GPT-4 is not getting worse
🤔 A discussion on OpenAI's GPT-4 model, which has been perceived to have declined in some areas but not systematically. The author argues that their tests were not enough to guard against chance and were biased. They also express their dislike of the discourse surrounding AI, as well as its potential implications.
https://coagulopath.com/gpt-4-is-not-getting-worse/
https://news.ycombinator.com/item?id=37532522

Adobe will charge “credits” for generative AI
🤔 This article answers commonly asked questions about generative credits - an AI technology used to create images, text effects, vectors, and more. It explains how generative credits are consumed, how many are included in plans, and any applicable limits.
https://helpx.adobe.com/firefly/using/generative-credits-faq.html
https://news.ycombinator.com/item?id=37538878

Open source AI will win
😃 Open source AI will win in the long run as it enables businesses to own their core product and provides a low-cost alternative to outsourcing AI tasks. Closed-source model providers may be suitable for the short term, but there is too much demand for open-source LLMs for them to be avoided in the future.
https://varunshenoy.substack.com/p/why-open-source-ai-will-win
https://news.ycombinator.com/item?id=37602674

Show HN: SeaGOAT – local, “AI-based” grep for semantic code search
🤖 SeaGOAT is a local-first semantic code search engine which enables developers to find and explore code more conveniently. It is open-source and built on top of the Git version control system.
https://github.com/kantord/SeaGOAT
https://news.ycombinator.com/item?id=37583219

We Can't Compete with AI Girlfriends
🤔 A deep dive into the rise of AI Girlfriends and its implications for girls and young women regarding beauty standards and unrealistic emotional expectations.
https://www.freyaindia.co.uk/p/we-cant-compete-with-ai-girlfriends
https://news.ycombinator.com/item?id=37573808

Generative AI's Act Two
🤔 Generative AI has sparked a Cambrian explosion of innovation but has also spurred some hysteria. After ChatGPT's success, many companies created AI applications with ambitious goals, but some have fallen short due to a lack of product-market fit and user retention. While generative AI has seen success, the overall ebullience of the AI ecosystem is unsustainable.
https://www.sequoiacap.com/article/generative-ai-act-two/
https://news.ycombinator.com/item?id=37588520

Intel Xeon Max 9480 Deep-Dive 64GB HBM2e Onboard Like a GPU or AI Accelerator
🤩 Intel's Xeon MAX 9480 is a powerful CPU that combines 56 cores with 64GB of HBM2e memory on the package, the same memory found on many GPUs and AI accelerators. This piece of technology has taken months to write, and it provides insight into this impressive CPU's capabilities.
https://www.servethehome.com/intel-xeon-max-9480-deep-dive-intel-has-64gb-hbm2e-onboard-like-a-gpu-or-ai-accelerator/
https://news.ycombinator.com/item?id=37570889

Intel slaps forehead, says I got it: AI PCs. Sell them AI PCs
😞 This message states that the user cannot access the website due to their traffic tripping the robot sensors. If the problem persists, they should contact [email protected] for further assistance.
https://www.theregister.com/2023/09/19/intel_pcs_nvidia/
https://news.ycombinator.com/item?id=37596738

A closer look at BookCorpus, a key dataset in machine learning
😱 A closer look at BookCorpus, a text dataset used to train large language models for Google, OpenAI, Amazon, and others, reveals some concerning issues - copyright violations, duplicates, and skewed genre representation. This paper aims to uncover the "dirty secrets" of BookCorpus and its implications for machine learning research.
https://towardsdatascience.com/dirty-secrets-of-bookcorpus-a-key-dataset-in-machine-learning-6ee2927e8650
https://news.ycombinator.com/item?id=37568847

Graph Neural Networks use graphs when they shouldn't
🤔 This paper examines how Graph Neural Networks (GNNs) tend to overfit graph structures in certain cases and provides a theoretical explanation for this phenomenon. It then proposes a graph-editing method to mitigate this overfitting and shows that it improves the accuracy of GNNs across multiple benchmarks.
https://arxiv.org/abs/2309.04332
https://news.ycombinator.com/item?id=37571535

Show HN: Whiz – A copilot for your terminal
📚 A copilot for your terminal - GitHub repository, which provides a copilot for the terminal to help users write better code with AI.
https://github.com/searchableguy/whiz
https://news.ycombinator.com/item?id=37567002

DeepMind cofounder: Generative AI is just a phase. What’s next is interactive AI
🤔 DeepMind cofounder Mustafa Suleyman believes generative AI is only a phase and that the future lies in interactive AI--software that can carry out tasks and interact with other software and people. He also calls for regulation but is optimistic it can be achieved. Suleyman has a history of working with youth helplines and local government, and he brings these values to his new billion-dollar company, Inflection.
https://www.technologyreview.com/2023/09/15/1079624/deepmind-inflection-generative-ai-whats-next-mustafa-suleyman/
https://news.ycombinator.com/item?id=37545430

Bernie Sanders: workers should reap AI benefits in form of ‘lowering workweek'
🤔 Bernie Sanders believes that if AI and robotics make us more productive, workers should benefit through a shorter work week. He also spoke out against the large salaries of the car manufacturers' executives.
https://www.theguardian.com/us-news/2023/sep/17/bernie-sanders-workers-should-reap-benefits-of-ai
https://news.ycombinator.com/item?id=37549572

Artificial intelligence can revolutionise science
🤩 AI can revolutionise science by accelerating the pace of discovery, leading to a golden age of invention. History has shown how new tools such as microscopes and computers have allowed for leaps in scientific progress. AI can further this progress and help solve some of humanity's biggest problems.
https://www.economist.com/leaders/2023/09/14/how-artificial-intelligence-can-revolutionise-science
https://news.ycombinator.com/item?id=37524686

Show HN: Symphony – Use GPT-4 to call functions in sequence
😐 This text outlines a series of functions that are designed to be used with GPT-4 to create a sequence that will control a dining room's lights to match the weather and time.
https://www.symphony.run/
https://news.ycombinator.com/item?id=37571732

More writers sue OpenAI for copyright infringement over AI training
🤬 Summary: A group of authors, including Pulitzer Prize winner Michael Chabon, have filed a lawsuit against OpenAI for copyright infringement, claiming that their works were used to train the company's AI-powered chatbot without permission. OpenAI has argued that AI training is a fair use of copyrighted works, but the authors are seeking damages and an order to block the company's practices.
https://www.reuters.com/technology/more-writers-sue-openai-copyright-infringement-over-ai-training-2023-09-11/
https://news.ycombinator.com/item?id=37545384

The physical process that powers a new type of generative AI
🤩 Physicists are looking into using well-understood equations of physical processes to replace the "black box" algorithms of neural networks in AI technology. A new method, called Poisson flow generative model (PFGM), takes data and creates an electric field whose properties depend on the distribution of charges. PFGM could create images of the same quality as diffusion-based approaches and do so 10-20 times faster.
https://www.quantamagazine.org/new-physics-inspired-generative-ai-exceeds-expectations-20230919/
https://news.ycombinator.com/item?id=37570743

Spanish girls are reporting AI-generated nudes being circulated at school
🤯This article covers the story of a group of Spanish girls, ages 12-17, who have had AI-generated nude photos of themselves circulated at their school. The police are investigating the case, and the mothers of the affected girls have organized to take action and try to stop those responsible.
https://english.elpais.com/international/2023-09-18/in-spain-dozens-of-girls-are-reporting-ai-generated-nude-photos-of-them-being-circulated-at-school-my-heart-skipped-a-beat.html
https://news.ycombinator.com/item?id=37580805

ChatGPT Isn't Coming for Your Coding Job
😐 This article looks at the history of software engineering and the fear that new technology will make human coders obsolete. The author examines the evolution of coding and computing and argues that despite these fears, programmers have only become more important over time.
https://www.wired.com/story/chatgpt-coding-software-crisis/
https://news.ycombinator.com/item?id=37545264

Compiling ML models to C for fun
This post is about compiling ML models to C for fun, demonstrated by looking at the micrograd library. It covers the basics of ML, reverse-mode automatic differentiation, and how ML models can be compiled to improve performance.
https://bernsteinbear.com/blog/compiling-ml-models/
https://news.ycombinator.com/item?id=37593404

Show HN: macOS GUI for running LLMs locally
🤩 Ava PLS is a desktop application revolutionizing language processing with the power of advanced language models. It allows users to perform various language tasks, like text generation, grammar correction, rephrasing, summarization, data extraction, and more. It is privacy focused, as all data stays with you, does not need an internet connection and is free to use. It is currently only available for macOS, with Windows and Linux planned for the future.
https://www.avapls.com/
https://news.ycombinator.com/item?id=37561085

Faster Than Llama 2: DeciLM LLM with Variable GQA
🤩 Deci has introduced DeciLM 6B, an LLM with 5.7 billion parameters, that offers 15 times higher throughput than Llama 2 7B while maintaining comparable quality. The model is generated using AutoNAC, Deci’s Neural Architecture Search engine, to push the efficient frontier and reduce inference cost, user experience, and carbon footprint. DeciLM is suitable for a range of generative AI applications, and its technical intricacies, novel attention mechanisms, and empirical results are delved into in the following sections.
https://deci.ai/blog/decilm-15-times-faster-than-llama2-nas-generated-llm-with-variable-gqa/
https://news.ycombinator.com/item?id=37530915

Tackling the curse of dimensionality with physics-informed neural networks
🤔 Summary: This paper discusses a new method called Stochastic Dimension Gradient Descent (SDGD), which uses Physics-Informed Neural Networks (PINNs) to solve high-dimensional partial differential equations (PDEs). It is shown to be able to solve notoriously hard PDEs in thousands of dimensions, with the example of a HJB equation and a Black-Scholes equation in 100,000 dimensions taking 6 hours on a single GPU.
https://arxiv.org/abs/2307.12306
https://news.ycombinator.com/item?id=37565140

Amazon limits authors to 3 books a day to combat AI-generated material
🤯 Amazon has introduced a limit of three books per day for authors to self-publish on their website to combat AI-generated content, which has caused controversy. AI-generated books have been found to have fake authors, nonsensical prose, and even misidentify potentially deadly fungi. The policy change is expected not to affect many publishers, and authors can seek exemptions if necessary.
https://themessenger.com/news/amazon-limits-authors-to-self-publish-3-books-a-day-to-combat-ai-generated-material
https://news.ycombinator.com/item?id=37579118

The AI explosion might never happen
🤔 Summary: Steve examines the potential for AI to undergo recursive self-improvement, leading to a rapid increase in AI capabilities, and concludes that this is not necessarily inevitable. He explores the reasons why this might not happen, with a parable demonstrating how a company could theoretically reach world domination if it successfully increases profits from R&D.
https://amistrongeryet.substack.com/p/recursive-self-improvement-foom
https://news.ycombinator.com/item?id=37577866

George R.R. Martin and other authors sue OpenAI for copyright infringement
😠 Summary: Popular authors, including George R.R. Martin, have filed a class-action lawsuit against OpenAI for copyright infringement, claiming they used their books to train AI models without permission or payment. OpenAI has responded by stating they are having productive conversations with authors, but other authors have also filed similar lawsuits against AI image platforms. Microsoft has also announced they will take the legal heat if commercial users of their Copilot AI service get sued.
https://www.theverge.com/2023/9/20/23882140/george-r-r-martin-lawsuit-openai-copyright-infringement
https://news.ycombinator.com/item?id=37585157

Show HN: Swirl – AI-Based Open-Source Search Engine Alternative to Algolia
🤩 Swirl is open-source software that uses AI to simultaneously search multiple content and data sources and return AI-ranked results that can prompt Generative AI, enabling users to get answers based on their own data.
https://github.com/swirlai/swirl-search
https://news.ycombinator.com/item?id=37586216

A Hands-On Introduction to Machine Learning
🤓 A comprehensive introduction to Machine Learning for students and professionals, with a hands-on approach and real-world examples. Covers all topics needed for a working knowledge of ML, including supervised and unsupervised learning, neural networks, reinforcement learning and cloud-based services. Written by Chirag Shah, Professor of Information Science at the University of Washington.
https://www.cambridge.org/highereducation/books/a-hands-on-introduction-to-machine-learning/3E57313A963BF7AF5C6330EB88ADAB2E
https://news.ycombinator.com/item?id=37575137

Meet your AI Executive Assistant
👍 Summary: Shortwave has released an AI executive assistant to help with email productivity, search capabilities, and even writing emails with personalized tone and style. Unlock the knowledge stored in email, and never worry about forgetting a meeting or researching a question again.
https://www.shortwave.com/blog/meet-your-ai-email-executive-assistant/
https://news.ycombinator.com/item?id=37585990

Google DeepMind's AI successor predicts how 71M mutations cause disease
🤔 Google DeepMind has developed an AI system that predicts the chances of genetic variants causing disease. The technology, AlphaMISENSE, is an advanced successor of AlphaFOLD and was unveiled in a Science paper published on Tuesday with predictions on 71 million mutations. Subscribers can gain access to the full article and explore the implications of this breakthrough.
https://endpts.com/google-deepminds-alphafold-successor-predicts-how-71m-mutations-cause-disease/
https://news.ycombinator.com/item?id=37578616

GPT 3.5 vs. Llama 2 fine-tuning: A Comprehensive Comparison
🤔A comprehensive comparison of GPT 3.5 and Llama 2 fine-tuning on two different tasks (SQL and functional representation) showed that GPT 3.5 performs marginally better than Llama 2, but costs 4-6x more to train. Experiments suggest that manually fine-tuning models can come close to the performance of GPT 3.5, at a fraction of the cost.
https://ragntune.com/blog/gpt3.5-vs-llama2-finetuning
https://news.ycombinator.com/item?id=37560125

Show HN: A murder mystery game built on an open-source gen-AI agent framework
https://www.gron.games/
https://news.ycombinator.com/item?id=37557545

‘Game of Thrones’ creator and other authors sue ChatGPT-maker OpenAI
🤬 Creators of the popular TV show Game of Thrones, George R.R. Martin, and other authors have filed a lawsuit against OpenAI, the creator of the ChatGPT natural language processing system, for copyright infringement. The plaintiffs allege that the system copies elements of their works without permission.
https://apnews.com/article/openai-lawsuit-authors-grisham-george-rr-martin-37f9073ab67ab25b7e6b2975b2a63bfe
https://news.ycombinator.com/item?id=37599261

Show HN: Superflows – open-source AI Copilot for SaaS products
👍 Open-source toolkit to build an AI copilot for SaaS products. Superflows-AI provides an AI copilot to SaaS products, allowing users to learn quickly and write better code with AI.
https://github.com/Superflows-AI/superflows
https://news.ycombinator.com/item?id=37533503

Windows 11’s next big update arrives on September 26th with Copilot, RAR support
🤩 Microsoft's Windows 11 update on September 26th includes the AI-powered Windows Copilot feature, a redesigned File Explorer, a new Ink Anywhere feature for pen users, big improvements to the Paint app, a new volume mixer, native RAR and 7-zip support, Dynamic Lighting control, and more.
https://www.theverge.com/2023/9/21/23882173/microsoft-windows-11-23h2-update-features-release-date
https://news.ycombinator.com/item?id=37598595

LoRA explained
🤔A concise summary of LoRA, an optimization for model fine-tuning which improves memory efficiency by reducing the number of parameters used during fine-tuning. It helps make LLM fine-tuning accessible to a wider range of companies and has led to startups offering LLM fine-tuning as a service. Follow-on work to LoRA has further improved the memory overhead of fine-tuning.
https://generatingconversation.substack.com/p/lora-explained
https://news.ycombinator.com/item?id=37603604

Show HN: Kusho – Copilot for API Testing
🤩Make API testing easy with Kusho! Automatically generate exhaustive tests for scenarios that can occur in production and execute them within your IDE with Kusho, the API Testing Copilot for stress-free release. Install Kusho for VS Code to get started.
https://kusho.co/
https://news.ycombinator.com/item?id=37589799

OpenAI and ChatGPT Lawsuit List
😐 Summary: This article provides a list of lawsuits against OpenAI and its AI model, ChatGPT, with commentary from a lawyer. It also looks at the claims and relief sought in each of these cases, all of which involve alleged privacy and copyright infringement.
https://originality.ai/blog/openai-chatgpt-lawsuit-list
https://news.ycombinator.com/item?id=37605334

I built a Stable Diffusion directory of 100 best AI workflows
🤯 Stable Workflows provides the best starting resolution when looking for combined quality and speed. It involves the use of XFOMERS or Torch2 SDP-Mem for cross-attention optimization and model merges of above 50%. It also recommends settings such as Linear Up for Mimic and Half Cosine Down for CFG. It has been recently updated with some new checkpoints, and elaborate and poetic prompts are encouraged.
https://www.stableworkflows.com/
https://news.ycombinator.com/item?id=37590985

Mesa-optimization algorithms in Transformers[pdf]
🤔 This paper discusses the hypothesis that the strong performance of Transformers in deep learning is due to an architectural bias towards mesa-optimization, a learned process running within the forward pass of a model. It proves this hypothesis by reverse-engineering autoregressive Transformers trained on sequence modelling tasks, uncovering underlying gradient-based mesa-optimization algorithms. It also proposes a novel self-attention layer, the mesa-layer, which could improve experiment performance.
https://arxiv.org/abs/2309.05858
https://news.ycombinator.com/item?id=37531815

Show HN: Using ChatGPT to track stock mention sentiment (and mapping over price)
🤔 MonkeeMath is an advanced data website that allows users to make predictions and view existing predictions. It also has a Mentions Graphing feature that provides sentiment data on stocks. Create an account to join the leaderboard and check out the new Stocktwits scraper coming soon.
https://monkeemath.com/
https://news.ycombinator.com/item?id=37528530

ChatGPT usage is rising again as students return to school
😕 A message from Bloomberg asking if you are a robot appears and asks to click a box to let them know you're not a robot. It then explains why it happened and provides a reference ID for inquiries related to the message.
https://www.bloomberg.com/news/articles/2023-09-20/chatgpt-usage-is-rising-again-as-students-return-to-school
https://news.ycombinator.com/item?id=37589424

Meshtastic and LoRa are “dangerous” (YouTube)
🤔 Rutgers University has confirmed that Meshtastic and LoRa are potentially dangerous technologies. They advise viewers to take caution when using them.
https://www.youtube.com/watch?v=EAQI2ZSmxPU
https://news.ycombinator.com/item?id=37550264

Thanks, and see you next week!