🤖 Last Week's AI Highlights from HN #15

Discover the latest breakthroughs and trends in artificial intelligence

Author

Vasco
September 08, 2023

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

Greetings, all!

This week, I increased the summary length of all the articles. I hope that you enjoy it!

X/Twitter has updated its terms of service to let it use posts for AI training
🤔 X/Twitter has updated its Terms of Service to allow the use of Posts for AI training, which could potentially limit access to content or limit the visibility of specific posts. Elon Musk is looking to use X as a data source for xAI, and this could lead to the merging of X Corp and xAI Corp down the line. The terms mean that AI-assigned definitions may control what content you can see on the internet, potentially narrowing your individual experience.
https://stackdiary.com/x-can-now-use-posts-for-ai-training-as-per-terms-of-service/
https://news.ycombinator.com/item?id=37352356

Show HN: HackYourNews – AI summaries of the top HN stories
😃 If you like this newsletter, you’ll like this project as well
https://hackyournews.com/
https://news.ycombinator.com/item?id=37427127

Transformers as Support Vector Machines
🤔This paper studies the relationship between Transformer architectures and Support Vector Machines (SVMs). It establishes a formal equivalence between the optimization geometry of self-attention and a hard-margin SVM problem and highlights the implications for parameterizing and optimizing the transformer model.
https://arxiv.org/abs/2308.16898
https://news.ycombinator.com/item?id=37367951

A GPT-4 capability forecasting challenge
🤩 This challenge tests your ability to accurately predict GPT-4's performance on various types of questions, such as the one provided, which asks "What is the capital of France?". The answer is Paris, and GPT-4 is highly likely to be able to answer this correctly.
https://nicholas.carlini.com/writing/llm-forecast/question/Capital-of-Paris
https://news.ycombinator.com/item?id=37360251

Retool AI
😃 Create a free account or book a demo to start building AI apps and workflows faster with Retool AI. Automate hundreds of manual tasks with custom AI workflows, integrate dozens of AI actions, and securely deploy to your developers and end-users. Trusted by developers, startups, and Fortune 500s.
https://retool.com/products/ai
https://news.ycombinator.com/item?id=37419886

What OpenAI really wants
😃 OpenAI, a young company, made waves when they released their ChatGPT product and began to explore the possibilities of AI. Sam Altman, the cofounder, went on a monthlong journey around the world to inform people about the potential of AI. He met with various world leaders, talked to students, and answered questions from entrepreneurs and engineers. His goal: to change the world with AI.
https://www.wired.com/story/what-openai-really-wants/
https://news.ycombinator.com/item?id=37390378

TinyLlama project aims to pretrain a 1.1B Llama model on 3T tokens
🤩 The TinyLlama project is an open, exciting endeavor to pretrain a 1.1B Llama model on 3 trillion tokens.
https://github.com/jzhang38/TinyLlama
https://news.ycombinator.com/item?id=37379984

A BERT for laptops, from scratch
🤔 A BERT lookalike was developed for training on a laptop (with an Nvidia 3070 RTX GPU). The notebook is developed for educational purposes more than performance, but in a bit more than half a day of training a model can obtain ~94% of the performance of the original BERT-base on the GLUE benchmark. This BERT variant uses BPE (Byte Pair Encoding) tokenization, relative position embeddings, pre-layer norm, no dropout, and automatic mixed precision. Training the tokenizer can take ~15 hours, pretraining with the MLM objective can take ~17 hours, and finetuning can take about an hour.
https://github.com/samvher/bert-for-laptops/blob/main/BERT_for_laptops.ipynb
https://news.ycombinator.com/item?id=37425130

ERNIE, China's ChatGPT, cracks under pressure
😞This article details how China's widely anticipated ChatGPT, ERNIE Bot, has flaws in its ability to handle "spicy" questions, plagiarizing tendencies, and Chinese language capabilities. It also discusses how the bot filters out unsavoury questions and fails to address difficult prompts.
https://www.chinatalk.media/p/how-ernie-chinas-chatgpt-cracks-under
https://news.ycombinator.com/item?id=37419049

Mushroom pickers urged: Avoid Amazon foraging books, appear to be written by AI
🤔 Summary: Mushroom pickers are being warned to avoid foraging books on Amazon which appear to have been written by AI chatbots. Samples of the books were tested and had a 100% AI detection score, with dangerous advice such as referring to smell and taste as ways to identify mushrooms.
https://www.theguardian.com/technology/2023/sep/01/mushroom-pickers-urged-to-avoid-foraging-books-on-amazon-that-appear-to-be-written-by-ai
https://news.ycombinator.com/item?id=37354577

LLMs, RAG, and the missing storage layer for AI
🤯 A summary of Ayush Chaurasia's article on LLMs, RAG, and the missing storage layer for AI. He explains the advances that have been made in the field of AI, and the dangers of using LLMs to generate text. He suggests that representation-based systems might be a better alternative, and suggests embedding the database with LLMs to store the knowledge.
https://blog.lancedb.com/llms-rag-the-missing-storage-layer-for-ai-28ded35fa984
https://news.ycombinator.com/item?id=37420628

Run ChatGPT-like LLMs on your laptop in 3 lines of code
🤖 A tool to enable running on-premises large language models with non-public data. It allows for an easy setup of a distributed system and efficiently utilises resources.
https://github.com/amaiya/onprem
https://news.ycombinator.com/item?id=37412793

We built an AI-powered Magic the Gathering card generator
🤩 Using AI, Nicholas Frosst and his two friends created Urza's AI, a website that generates Magic the Gathering cards with all the necessary information and an image, from just a card name. Through the use of language AI and text-to-image AI, the team was able to create hundreds of thousands of cards in just a few days.
https://txt.cohere.com/urzas-ai/
https://news.ycombinator.com/item?id=37427854

Show HN: Fully client-side GPT2 prediction visualizer
😃This text is about an interactive language modeling visualization tool called 'Perplexity', which is designed to help users understand the complexities of language modeling.
https://perplexity.vercel.app/
https://news.ycombinator.com/item?id=37398812

Four Lectures on Standard ML (1989) [pdf]
🤔 Summary: This text appears to be a PDF document with a lot of code. N/A.
https://www.cs.tufts.edu/~nr/cs257/archive/mads-tofte/four-lectures.pdf
https://news.ycombinator.com/item?id=37384916

The Eleuther AI Mafia
🤩 This post concerns the RWKV project, a distributed, international, mostly uncredentialed community working to train >10B models, open source and OPT/BLOOM-competitive. Eugene Cheah, a member of the project, is interviewed to discuss the challenge of the Transformers dominance and the project's goals.
https://www.latent.space/p/rwkv#%C2%A7the-eleuther-mafia
https://news.ycombinator.com/item?id=37368264

Vector Search with OpenAI Embeddings: Lucene Is All You Need
🤔This paper is about how OpenAI embeddings can be used in vector search with Lucene, challenging the popular narrative that a dedicated vector store is necessary. It suggests that from a cost-benefit analysis, there is no need for a dedicated vector store in a modern "AI stack" for search.
https://arxiv.org/abs/2308.14963
https://news.ycombinator.com/item?id=37373635

ChatGPT can get worse over time, Stanford study finds
😱 Stanford study finds that ChatGPT, created by OpenAI, performed worse on certain tasks in June than its March version, with its accuracy for correctly solving math problems dropping from 98% to 2%. The researchers also found "wild fluctuations" in the technology's ability to perform tasks and found that changes made to one part of the model could have an effect on others. They stress the importance of continuously monitoring the models' performance over time.
https://fortune.com/2023/07/19/chatgpt-accuracy-stanford-study/
https://news.ycombinator.com/item?id=37393073

Scientific sleuths spot dishonest ChatGPT use in papers
💡 A group of researchers have been using a free-to-use AI chatbot called ChatGPT to write papers without disclosing it, and some of these manuscripts have been slipping past peer reviewers. Publishers have recognized this as a breach of ethical policies, and some papers have been retracted as a result.
https://www.nature.com/articles/d41586-023-02477-w
https://news.ycombinator.com/item?id=37431946

Buzzy AI startup for generating 3D models used cheap human labor
🤔 This article exposes how an AI startup, Kaedim, hyped its technology for converting 2D images into 3D models. In reality, the company often uses human artists to create the models, and the AI technology often produced unrecognizable blobs. Kaedim markets its services as a cheaper alternative to hiring 3D artists, but the cost of the service is still quite high.
https://www.404media.co/kaedim-ai-startup-2d-to-3d-used-cheap-human-labor/
https://news.ycombinator.com/item?id=37405450

Making deep learning go brrrr from first principles (2022)
🤔 Summary: Deep learning model performance can be improved by understanding the components of efficient systems, such as compute, memory bandwidth, and overhead. The rate of increase in compute can often outpace the rate of increase in memory bandwidth, making it harder to achieve peak efficiency. Additionally, specialized hardware like Nvidia's Tensor Cores can be utilized for matrix multiplication to increase FLOPS.
https://horace.io/brrr_intro.htm
https://news.ycombinator.com/item?id=37361711

Is the ChatGPT and Bing AI boom already over?
🤔The article discusses how the initial excitement surrounding AI and generative tools has decreased significantly since its introduction to the public a year ago, questioning if this is the end of the AI boom.
https://www.vox.com/technology/2023/8/19/23837705/openai-chatgpt-microsoft-bing-google-generating-less-interest
https://news.ycombinator.com/item?id=37359013

Co-founder of DeepMind on how AI threatens to reshape life as we know it
😰The co-founder of DeepMind, Mustafa Suleyman, discussed how AI has the potential to reshape life as we know it. He spoke about the possible uses of AI to support mental health care and unlock the creativity of millions of people who may not have had access to emotional intelligence.
https://www.theguardian.com/books/2023/sep/02/i-hope-im-wrong-the-co-founder-of-deepmind-on-how-ai-threatens-to-reshape-life-as-we-know-it
https://news.ycombinator.com/item?id=37419493

Google Meet’s new AI will be able to go to meetings for you
🤔Google Meet has introduced AI-powered features that will allow users to take notes in real-time, generate summaries, and even have their AI attend meetings for them - saving time and effort for the user. The AI features are yet to be tested but could be a game-changer for online meetings.
https://www.theverge.com/2023/8/29/23849056/google-meet-ai-duet-attend-for-me
https://news.ycombinator.com/item?id=37392598

Deep Learning Is Easy – Learn Something Harder
🤔This post talks about how deep learning is becoming more common and easier to learn, and suggests that people should instead look for something harder to learn in order to differentiate themselves.
https://www.inference.vc/deep-learning-is-easy/
https://news.ycombinator.com/item?id=37409225

Silicon Valley’s elites can’t be trusted with the future of AI
🤔 Emoji portraying the sentiment of the text: A summary of this text is that Silicon Valley's tech elites cannot be trusted with the future of AI because their motives are largely driven by profits and are disconnected from the world's realities. India can break this dominance by building, training and fine-tuning its own large language model to create a safer and fairer AI future. 🤝https://fortune.com/2023/09/01/silicon-valley-elite-trust-future-ai-dominance-dangers-god-complex-wadhwa-gupta/
https://news.ycombinator.com/item?id=37367017

Join us for OpenAI’s first developer conference on November 6 in San Francisco
🤩 OpenAI is hosting its first developer conference, OpenAI DevDay, on November 6, 2023 in San Francisco. Attendees will be able to preview new tools, attend breakout sessions and join a keynote livestream. Developers everywhere can join via the livestream and registration for in-person attendance will open in the coming weeks.
https://openai.com/blog/announcing-openai-devday
https://news.ycombinator.com/item?id=37408234

Gizmodo fires Spanish staff amid switch to AI translator
🤯G/O Media's decision to replace human writers with AI translations of English-language articles for Gizmodo's Spanish-language site has been met with controversy due to potential inaccuracies in the translations. This could lead to a significant loss of quality content for Spanish-speaking audiences.
https://arstechnica.com/information-technology/2023/09/ai-took-my-job-literally-gizmodo-fires-spanish-staff-amid-switch-to-ai-translator/
https://news.ycombinator.com/item?id=37399336

Emoji Generator with AI
🤩AImoji is an open source emoji generator powered by Replicate, which can generate emojis of famous scientists, historical figures, and various characters when given commands such as 'cat' or 'high five'.
https://emoji.fly.dev
https://news.ycombinator.com/item?id=37433523

Show HN: Comgra: A library for debugging and understanding neural networks
🔎 GitHub repository search for a project related to AI which includes code review, packages, security and discussions.
https://github.com/FlorianDietz/comgra
https://news.ycombinator.com/item?id=37379261

Building AI Talent Pipeline: Strategies for Attracting and Retaining AI Talent
🤩 Building an AI Talent Pipeline is an important step for organizations to remain competitive in the digital age. This article outlines strategies for identifying and recruiting AI talent, including key skills and qualifications required, online platforms, communities, and events, and AI-focused job boards and recruitment agencies.
https://remotebase.com/blog/building-an-ai-talent-pipeline-strategies-for-attracting-recruiting-and-retaining-ai-talent
https://news.ycombinator.com/item?id=37379320

Show HN: Modular Diffusion – A modular Python library for diffusion models
🔨Python library for designing and training your own Diffusion Models with PyTorch.
https://github.com/cabralpinto/modular-diffusion
https://news.ycombinator.com/item?id=37361736

Federated benchmarking of medical artificial intelligence with MedPerf
🤔 A team of researchers have developed MedPerf, a federated benchmarking tool for medical AI, which is detailed in the article published in Nature Machine Intelligence.
https://www.nature.com/articles/s42256-023-00652-2
https://news.ycombinator.com/item?id=37428505

ChatGPT no longer gives links: “Open your web browser and search”
https://public.abdulbaki-eren-bilir.com.tr/openyourwebbrowser.png
https://news.ycombinator.com/item?id=37356922

Show HN: Automating Job Search with AI
🤖 An experiment to automate job search with LLMs, GitHub's AdrianKrebs/datalens provides a platform for automating job search with LLMs. The project provides many features such as CI/CD & Automation, DevOps, DevSecOps, Customer Stories, White papers, Ebooks, Webinars, and Partners.
https://github.com/AdrianKrebs/datalens
https://news.ycombinator.com/item?id=37354204

Getting from Generative AI to Trustworthy AI: What LLMs Might Learn from Cyc
🤔This paper discusses how LLMs, the most popular current approach to AI, are lacking in aspects of reasoning, leading to them being less than completely trustworthy. The authors suggest hybridizing current approaches with more formal approaches and discuss how one AI system, Cyc, is able to reason in higher order logic in real time.
https://arxiv.org/abs/2308.04445
https://news.ycombinator.com/item?id=37354601

A complete guide to fine-tuning Code Llama
🤔A guide to fine-tuning Code Llama in order to become a better SQL developer. The guide shows how to use the b-mc2/sql-create-context dataset and the A100 GPU machine with Python 3.10 and cuda 11.8 to run the notebook. Instructions on how to install the necessary packages and libraries, load and view the dataset, and check the base model are also included.
https://ragntune.com/blog/guide-fine-tuning-code-llama
https://news.ycombinator.com/item?id=37375303

Fine-Tuning LLMs: LoRA or Full-Parameter? An In-Depth Analysis with Llama 2🤔This blog post compares full-parameter fine-tuning and LoRA for Llama 2 LLMs, discussing the pros and cons of each. Experiments and comparisons are provided, along with tips and tricks to streamline LoRA training. Results from the comparison indicate that LoRA is nearly on par with full-parameter fine-tuning, outperforming GPT-4 in specialized tasks.
https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2
https://news.ycombinator.com/item?id=37405866

Show HN: Atticus AI – Negotiate consulting contracts faster using GPT-4
😃 Atticus AI is an AI-driven platform that can quickly analyse contracts, raise red flags and suggest better terms to help you save time and money. It offers 100x faster feedback than an attorney and your first 3 analyses are free. Please note, the analysis generated is not guaranteed to be correct or up to date and should not be relied on as professional advice.
https://askatticus.ai
https://news.ycombinator.com/item?id=37394650

Recomputing ML GPU Performance: AMD vs. Nvidia
🤔An impressive feat - Lisa Su is doing her best to guide AMD towards better AI support with the Huggingface partnership and convincing George Hotz to submit more bug reports. However, many issues exist to overcome, including the need to compute GPU performance figures as NVIDIA has Tensor Cores for higher FLOPS, and AMD has WMMA for two 16x16 matrix multiplications in 32 cycles.
https://espadrine.github.io/blog/posts/recomputing-gpu-performance.html
https://news.ycombinator.com/item?id=37383219

Thanks, and see you next week!