🤖 Last Week's AI Highlights from HN #16

Discover the latest breakthroughs and trends in artificial intelligence

Author

Vasco
September 15, 2023

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

Greetings, all!

Nvidia’s AI supremacy is only temporary
🤔 Nvidia's AI domination is not secure in the coming years as there are structural drivers behind their success that are likely to change. The main drivers are: training dominates over inference, Nvidia's software stack is far superior to its competitors, and ML researchers have a lot of purchasing power.
https://petewarden.com/2023/09/10/why-nvidias-ai-supremacy-is-only-temporary/
https://news.ycombinator.com/item?id=37467585

AI girlfriend ads are flooding Instagram and TikTok
😲 Ads for AI sex workers are flooding Instagram and TikTok, despite attempts to keep their content away from such platforms. Dozens of tech startups are running explicit ads for apps that promise NSFW pics and chats with “no censoring.” Ads feature digitally created potential girlfriends, often with large breasts and tight clothing, and some use memes of popular children’s characters to promote the apps. Meta and TikTok have removed some of the ads, but some have stayed up for weeks and gotten thousands of views. The marketing push is part of an AI gold rush, as app developers are trying to capitalize on a surge of interest in AI.
https://www.nbcnews.com/tech/social-media/ai-girlfriend-ads-instagram-tiktok-chat-pics-chatgpt-dose-rcna97547
https://news.ycombinator.com/item?id=37437066

A boy saw 17 doctors over 3 years for pain. ChatGPT found the right diagnosis
🤔 A frustrated mom turned to artificial intelligence platform ChatGPT for help in diagnosing her son's chronic pain and other symptoms, which 17 doctors over 3 years had been unable to do. ChatGPT eventually found the cause to be tethered cord syndrome, bringing the 3-year journey to an end.
https://www.today.com/health/mom-chatgpt-diagnosis-pain-rcna101843
https://news.ycombinator.com/item?id=37475604

Simulating History with ChatGPT
🤔 This post discusses how using Large Language Models (LLMs) such as ChatGPT as a teaching tool can be used to simulate interactive historical settings in university assignments. It explains how this could potentially positively impact the humanities, as it requires the same skills and methods taught in humanities classes. It provides a detailed example of how to analyze primary sources and concludes with links to prompts for simulating different historical settings.
https://resobscura.substack.com/p/simulating-history-with-chatgpt
https://news.ycombinator.com/item?id=37480155

Show HN: Lantern – a PostgreSQL vector database for building AI applications
🤩 Lantern Systems, Inc. is making a PostgreSQL vector database extension for building AI applications. It has a complete feature set, supports popular use cases, interoperability and parallel index creation capabilities. It has good performance and benchmarks, and is built on usearch for scalability and performance. It offers free AirPods Pro and advice for switching from pgvector and help for building AI applications.
https://docs.lantern.dev/blog/2023/09/13/hello-world
https://news.ycombinator.com/item?id=37499375

The meeting of the minds that launched AI
🤔This article discusses the meeting of the minds in 1956 that launched the field of artificial intelligence (AI). It includes a photo of some of the main participants and provides an overview of the Dartmouth Summer Research Project on Artificial Intelligence and how it gave rise to the field of AI.
https://spectrum.ieee.org/dartmouth-ai-workshop
https://news.ycombinator.com/item?id=37469849

Lessons from YC AI Startups
🤔 This article breaks down the 139 AI startups from YC's S23 batch, and looks at the biggest takeaways. It highlights how diverse the AI landscape is, and the top four industries represented: AI Ops, Developer Tools, Healthcare + Biotech, and Finance + Payments. It also discusses the use of AI to collaborate with existing workers, and the emergence of AI Ops as a sector.
https://www.ignorance.ai/p/5-lessons-from-139-yc-ai-startups
https://news.ycombinator.com/item?id=37490924

Advanced NLP with SpaCy
🤩This text provides an overview of a free online course about Advanced NLP with spaCy, discussing topics such as finding words, phrases, names and concepts; large-scale data analysis; processing pipelines; and training a neural network model. It also provides information about the instructor, Ines, who is a core developer of spaCy and the co-founder of Explosion.
https://course.spacy.io/en/
https://news.ycombinator.com/item?id=37442574

Fandom can't decide if leaked songs are real or AI-generated
🤔 Summary: Fans of Harry Styles and One Direction are torn over the legitimacy of leaked snippets of unreleased songs being sold on Discords for hundreds of dollars. AI-analysis companies have been unable to determine if the songs are real or AI-generated, which has sparked a days-long crowdsourced investigation and communitywide obsession.
https://www.404media.co/harry-styles-one-direction-ai-leaked-songs/
https://news.ycombinator.com/item?id=37482455

Show HN: Improved freemusicdemixer – AI music demixing in the browser
🎵 Enjoy free AI-based music demixing in your browser with this web app: demix audio files into separate components like vocals, bass and drums! Supporting batch demixing, low-memory segmented inference, and improved demixing quality with Wiener filtering. 🎵
https://freemusicdemixer.com/
https://news.ycombinator.com/item?id=37507734

Show HN: WhatsApp-Llama: A clone of yourself from your WhatsApp conversations
🤖 This repository provides a way to finetune a LLM (Language Model) to have it speak like you based on your WhatsApp conversations.
https://github.com/Ads-cmu/WhatsApp-Llama
https://news.ycombinator.com/item?id=37448005

Apple is reportedly spending ‘millions of dollars a day’ training AI
🤩 Apple is investing millions of dollars per day into AI projects, with teams working on multiple models such as conversational AI, image generation, and multimodal AI. Apple's most advanced model is reportedly more powerful than OpenAI's GPT-3, and could be used for tasks like customer support and automating multistep tasks with Siri.
https://www.theverge.com/2023/9/6/23861763/apple-ai-language-models-ajax-gpt-training-spending
https://news.ycombinator.com/item?id=37437026

Gmail and Instagram are training AI, and there’s little you can do about it
😔 Summary: Big Tech companies are using our conversations, photos, and documents to train their AI without our permission or knowledge, and there's little we can do to stop it. Our data is being used to create lucrative new technologies, and we don't yet know the risks this poses to our privacy, reputation and work.
https://www.washingtonpost.com/technology/2023/09/08/gmail-instagram-facebook-trains-ai/
https://news.ycombinator.com/item?id=37487926

US rejects AI copyright for famous state fair-winning Midjourney art
😢 US Copyright Office rejects AI-generated artwork from winning Colorado State Fair art contest due to lack of human authorship required for registration. The artist refused to disclaim the AI-generated material, resulting in the Office's decision. Previous works created by AI have also been denied copyright protection.
https://arstechnica.com/information-technology/2023/09/us-rejects-ai-copyright-for-famous-state-fair-winning-midjourney-art/
https://news.ycombinator.com/item?id=37468983

ChatGPT is losing some of its hype, traffic falls for the third month in a row
😔 ChatGPT, the popular AI-powered chatbot, is losing its hype as traffic has dropped for three months in a row.
https://www.businessinsider.com/chatgpt-losing-hype-traffic-down-three-months-in-a-row-2023-9
https://news.ycombinator.com/item?id=37436298

The tide is shifting: 1.3B outperforms 7B Llama 2
🤓 This paper examines the effectiveness of using a Transformer-based language model (phi-1.5) with 1.3 billion parameters for common sense reasoning tasks. Results show that it is able to perform better than models with 5x more parameters on certain tasks, and displays traits of larger models with both positive and negative outcomes.
https://arxiv.org/abs/2309.05463
https://news.ycombinator.com/item?id=37482248

Llama 2 on togetherAI is as bad of a privacy nightmare as OpenAI
🤗This policy outlines Together Computer Inc's collection, use, and disclosure of personal data through the Together.xyz website and model training, fine-tuning, serving, and inference platform. It covers data collected through the Services, websites, and from third-party sources. The policy also states how the Company does not store or access credit card or payment information and does not apply to data about its employees, contractors, and agents.
https://together.ai/privacy
https://news.ycombinator.com/item?id=37436498

AI and the End of Programming
🤔 AI might replace traditional programming in the future, according to Matt Welsh's essay in Communications of the ACM. Despite Welsh's enthusiasm about this change, Brian Hayes is skeptical and expresses his preference for programming and his enjoyment of it, as he finds it a useful tool for understanding the world. He also mentions the roots of this idea in Charles Babbage and the FORTRAN Automatic Coding System.
http://bit-player.org/2023/ai-and-the-end-of-programming
https://news.ycombinator.com/item?id=37501456

Typing on Any Surface: A Deep Learning Method for Keystroke Detection in AR
🤔A paper that proposes and validates a deep-learning based approach for keystroke detection in augmented reality, featuring an adaptive convolutional recurrent neural network. Results show an overall accuracy of 91.05% when typing 40 words per minute.
https://arxiv.org/abs/2309.00174
https://news.ycombinator.com/item?id=37466217

Show HN: I built an app that lets you create and chat with 3D AI characters
.https://www.heyastrocyte.com/
https://news.ycombinator.com/item?id=37518209

How to run a competitive AI startup fundraise in 2023
🤩 Summary: This text provides advice for those looking to fundraise for an AI startup in 2023, highlighting the need to prepare for the raise in advance and be systematic, as well as treating the process as if it were a date. Additionally, it provides advice for crafting a pitch, running the raise, and more.
https://context.ai/post/how-to-run-a-competitive-ai-startup-fundraise-in-2023
https://news.ycombinator.com/item?id=37481424

OpenAI confirms that AI writing detectors don’t work
🤯OpenAI confirms that AI writing detectors are ineffective and unreliable, and that ChatGPT, the AI language model, has no ability to detect AI-written text. Humans can detect AI writing in some cases, such as when a student's writing style or capability suddenly changes. Automated AI detection tools should be avoided.
https://arstechnica.com/information-technology/2023/09/openai-admits-that-ai-writing-detectors-dont-work/
https://news.ycombinator.com/item?id=37482423

Claude and LLaMa get basic human anatomy wrong
🤦‍♂️This post is about how Claude and LLaMa, two different LLMs, incorrectly answer basic human anatomy questions. Claude provides the wrong answer, while LLaMa doesn't even recognize the medical term. ChatGPT is able to recognize the term, but still provides an incorrect answer.
https://hermitian.substack.com/p/claude-and-llama-get-basic-human
https://news.ycombinator.com/item?id=37517674

What if AI treats humans the way we treat animals?
🤔 This article discusses the fear of AI and how it reveals our anxiety over the mistreatment of animals. It considers the idea that humans are not special in terms of intelligence and perception, and how this is inconsistent with the cruelty we show towards animals.
https://www.vox.com/the-highlight/23777171/ai-animals-rights-cruelty-transhumanism-bostrom
https://news.ycombinator.com/item?id=37507534

Why Japan is building its own version of ChatGPT
🤔 Japan is investing in its own versions of the AI chatbot ChatGPT, which has become a worldwide sensation, due to concerns that current AI systems may not be able to capture the intricacies of the Japanese language and culture. Researchers are exploring how to make the technology more sensitive to cultural norms.
https://www.nature.com/articles/d41586-023-02868-z
https://news.ycombinator.com/item?id=37508067

Show HN: Find jobs at top AI startups
🤩 Get your foot in the door at the hottest AI companies with over 500+ jobs available from 25+ companies.
https://workinai.xyz/
https://news.ycombinator.com/item?id=37435689

The A.V. Club's AI-Generated Articles Are Copying Directly from IMDB
🤦‍♀️ The A.V. Club recently started using AI-generated articles from IMDb, which caused backlash from journalists and writers. After an internal memo, the AI output was investigated and it was discovered that the AI was copying the IMDb descriptions verbatim and not changing any of the wording, leading to a "sickening" result.
https://futurism.com/the-av-club-imdb
https://news.ycombinator.com/item?id=37497593

Zero-trust AI APIs serving Llama 2 70B inside enclaves
🤗 Summary: BlindLlama is a Zero-trust AI API that provides an easy way to use open-source LLMs, with guaranteed data confidentiality and transparency thanks to secure hardware and cryptographic proof. It is still under development, but users can contribute to the project and follow its progress on the roadmap.
https://blindllama.mithrilsecurity.io/en/latest/
https://news.ycombinator.com/item?id=37514590

Show HN: Monthly budgeting spreadsheet and ML expense categorisation
🤗 Expense Sorted is an AI-driven tool designed to reduce the time spent manually categorizing expenses. It securely connects to a Google Sheet, and offers features such as accurate transaction identification, customizable categories, and an interface for manual adjustments. Users have reported great success with the tool, praising its efficiency and accuracy.
https://www.expensesorted.com/
https://news.ycombinator.com/item?id=37452986

Can you explain Transformers in one sentence? (attention mechanisms) [video]
https://www.youtube.com/watch?v=qAb581l7lOc
https://news.ycombinator.com/item?id=37462781

AI technology behind ChatGPT built was in Iowa – with a lot of water
🤔 Though the use of Artificial Intelligence (AI) in technology has been increasing, few know of the hefty costs that come with it, from expensive semiconductors to an increase in water consumption. Microsoft, OpenAI and Google have been growing their AI tools, but have been secretive about the specifics. Researchers have estimated that ChatGPT uses approximately 500 milliliters of water for a series of 5 to 50 prompts or questions; however, most people are unaware of the resource usage underlying the model.
https://news.yahoo.com/artificial-intelligence-technology-behind-chatgpt-131421382.html
https://news.ycombinator.com/item?id=37448699

Show HN: Wyvern – Real-time machine learning platform for marketplaces
🤖 Summary: Wyvern-AI provides a suite of features for automating workflows, managing packages, finding and fixing vulnerabilities, setting up dev environments, writing better code with AI, managing code changes, planning and tracking work, and collaborating outside of code. Additionally, GitHub Sponsors can be used to fund open source developers and the ReadME Project provides GitHub community articles.
https://github.com/Wyvern-AI/wyvern
https://news.ycombinator.com/item?id=37501070

Show HN: HelpKit AI – AI Chatbot trained on your Notion docs
🤩 Revolutionize your customer support with HelpKit AI—a chatbot add-on trained on your company's knowledge base. Get smart usage insights, customize the language and behavior of your AI chatbot, and get an AI support assistant for only $49/m. Try it out now!
https://www.helpkit.so/helpkit-ai/
https://news.ycombinator.com/item?id=37506742

Artificial intelligence is ineffective and potentially harmful for fact checking
🤔 Summary: This paper looks at the effectiveness and potential harm of Artificial Intelligence (AI) for fact-checking. Results show that while AI performs well in debunking false headlines, it does not significantly affect people's ability to discern accuracy or share accurate news. However, it can be harmful in certain cases by decreasing belief in true headlines and increasing belief in false headlines.
https://arxiv.org/abs/2308.10800
https://news.ycombinator.com/item?id=37439523

Show HN: Eesel AI – ChatGPT over company knowledge, without APIs
😃 Connect your knowledge with ChatGPT and build an oracle that can instantly answer any question for your internal Slack, Microsoft Teams, website widget, or other apps. A SOC2 compliant app stores your data securely and it's never used to train models. Lowest prices, guaranteed, with free and custom plans available. Get the ultimate teammate or customer support agent with ChatGPT!
https://www.eesel.ai/
https://news.ycombinator.com/item?id=37509522

Show HN: InsightsGPT – Discovering GPT Use Cases, Data, and Productivity Gains
 🤔 Unlock the full potential of your organization with InsightsGPT. InsightsGPT provides analytics to measure, monitor and analyze workforce productivity across GPT applications. With GPT usage analytics, it reveals how team is leveraging GPT to boost work processes and performance. It also provides AI-driven detection of potential GPT use cases and unstructured data pipeline to develop custom GPT models and grow tailored generative intelligence. Lastly, it offers real-time analytics to make rapid and strategic decisions.
https://officeautomata.com/insightsgpt/
https://news.ycombinator.com/item?id=37512674

Prompting Code Llama
🤔 A guide to using the Code Llama model to structure prompts for its various variations and features, including instruct, code completion, and infill. Examples of these prompts are provided.
https://ollama.ai/blog/prompting-code-llama
https://news.ycombinator.com/item?id=37447770

From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting
🤔This paper explores how to create the "right" amount of information to include in a summary through the use of GPT-4 and a "Chain of Density" prompting technique. Results suggest that GPT-4 summaries generated by this technique are preferred by humans and have more abstractive fusion than those generated by a vanilla prompt. 500 annotated CoD summaries are available for download.
https://huggingface.co/papers/2309.04269
https://news.ycombinator.com/item?id=37469009

Show HN: Vips – Emacs Interface for OpenAI's GPT API and DeepL's Translation API
🤖Emacs interface for OpenAI's GPT API and DeepL's translation API, allowing users to interact with both services from within Emacs.
https://github.com/marcklemp/vips
https://news.ycombinator.com/item?id=37502387

Show HN: ChatGPT History Explorer
🤖 This repository provides an exploration of ChatGPT, a tool for building and training chatbot models. It contains the source code, example datasets, and instructions for replicating the results from our paper.
https://github.com/mirableio/chat-history
https://news.ycombinator.com/item?id=37507461

Show HN: Convert JavaScript to TypeScript with ChatGPT
🤔 Convert Javascript code to Typescript code quickly using AI. Follow the three steps: Paste code in the code box, click "Convert Now" and copy the converted code from the output box. Created by Yuval Suede.
https://js2ts.com/
https://news.ycombinator.com/item?id=37492433

Improving OpenAI Function Calling Usability in Python
🤔 This article provides an overview of the tools and frameworks needed to effectively use language models, such as Instructor, Pydantic and OpenAI’s Function Calling. A modular approach is described which simplifies the validation flow of language models and reduces the need for developers to create new abstractions.
https://medium.com/@jxnlco/bridging-language-model-with-python-with-instructor-pydantic-and-openais-function-calling-f32fb1cdb401
https://news.ycombinator.com/item?id=37447074

Can ChatGPT Write Code?
🤔Summary: This article explores the capabilities of ChatGPT, a type of AI language model, and whether it can write code. It looks at the potential applications of this model, as well as its limitations. Additionally, it looks at how ChatGPT can be used to write code and outlines an example of building a recipe website.
https://stratoflow.com/can-chatgpt-write-code/
https://news.ycombinator.com/item?id=37512869

Thanks, and see you next week!