- That AI Thing
- Posts
- weekend ai reads for 2024-08-30
weekend ai reads for 2024-08-30
📰 ABOVE THE FOLD: SMÖRGÅSBORD OF TOOLS
AI death calculator predicts when you’ll die — it’s ‘extremely’ accurate / New York Post (7 minute read)
In the report, the professor of network and complex systems from the Technical University of Denmark, and co-authors introduce an algorithm known as “life2vec,” which uses select details of an individual’s life — including income, profession, residence and health history — to determine life expectancy with 78% correctness.
the tool: Life2vec.io AI Death Calculator
we’ll find out how accurate this is in 4 months or so
PackPack — 1-Click to save everything with AI
How Hellman’s is using Google Cloud AI to save leftovers / Google Blog (3 minute read)
To help reduce household food waste, Hellmann’s — best known for their tasty mayonnaise — released a Meal Reveal tool utilising AI to make delicious meals out of the ingredients already in your fridge.
the tool: Hellmann's Meal Reveal — Reveal the meal hiding in your fridge
AI chatbots are writing police reports and watchdogs are concerned — Oklahoma City’s police department is one of a handful to experiment with AI chatbots to produce the first drafts of incident reports. / Fast Company (9 minute read)
the tool: Draft One
“officer-in-the-loop” won’t prevent this, but will give departments a tool to blame:
Officers were able to get the suspect into handcuffs and noted he appeared to be suffering medical distress. Officers called for an ambulance. He was transported to Hennepin County Medical Center by ambulance where he died a short time later.
Tolan — Alien Best Friend
Meet your Tolan: a friendly little alien you can talk to about whatever, and who can even help picture your ideas!
📻 QUOTE OF THE WEEK
Why don’t they spend the time applying AI to something actually useful?
Unnamed Target Employee (source: Target Employees Hate Its New AI Chatbot; archived link)
🏗️ FOUNDATIONS & CULTURE
No one’s ready for this — Our basic assumptions about photos capturing reality are about to go up in smoke. / The Verge (10 minute read)
related (1), Realism Comparison v2 - Amateur Photography Lora / Major_Specific_23, Reddit
apparently, more “realistic” just means less “instagram aesthetic”
the LoRA, for Flux: Amateur Photography [Flux Dev] - v2.0 / Civitai
related (2), Google’s AI tool helped us add disasters and corpses to our photos — The new feature on the Pixel 9 series is way too good at creating disturbing imagery — and the safeguards in place are far too weak. / The Verge (7 minute read)
more examples / Chris Welch, Threads
about the tool: Google Releases Powerful AI Image Generator You Can Use for Free / Peta Pixel (4 minute read)
LLMs excel at inductive reasoning but struggle with deductive tasks, new research shows / Venture Beat (5 minute read)
Their findings show that while LLMs can be very good at finding the rules of a task from solved examples, they are limited in following specific instructions. The findings can have important implications for how we use LLMs in applications that require reasoning.
The point of lightning-fast model inference — We’re obsessed with generating thousands of tokens a second for a reason—and it isn’t just to wow end users with text showing up on a screen really fast. / Supervised, Substack (sorry) ($) (9 minute read)
And while it all moves faster than a human eye can read, it turns out that this blistering speed is just one of what are likely many precursors to the ability to build out networks of models that can satisfy some of the potential dream scenarios for language models.
related (1), Introducing Cerebras Inference: AI at Instant Speed / Cerebras blog (7 minute read)
Cerebras inference delivers 1,800 tokens per second for Llama3.1 8B and 450 tokens per second for Llama3.1 70B, which is 20x faster than NVIDIA GPU-based hyperscale clouds.
as a reference point, most web LLM interface output 30-60 tokens per second
related (2) Consider not using a Mac... / mayo551, Reddit (2 minute read)
So for comparison it takes around 260 seconds for my M2 Mac to sort through 32k context with this setup (though the Mac can’t use quant k,v). It takes 25 seconds on the 2080ti to sort through 32k.
The Mac also uses around 30G VRAM for 32k context with this same setup. Or something like that... too lazy to double check. So I get double the context on the Nvidia build without running out of VRAM.
Kraft Heinz AI ‘Lighthouse’ Helps Forecast Supply-Chain Demands / Business Insider (6 minute read)
Swings in product demand are fairly common in food manufacturing because retailers, such as grocery stores, frequently change their order sizes. The hope is that with AI, Kraft Heinz workers can recognize patterns and precisely address these demand fluctuations, boosting sales.
ERP.AI — integrates seamlessly with your ERP, making it smarter by enhancing enterprise process, data governance & decision-making — without disruptions
Philippines’ Call Centers Navigate AI Impact on Jobs — The experiences of staff in the Philippines’ outsourcing industry are a preview of the challenges and choices coming soon to white-collar workers around the globe. / Bloomberg (16 minute read)
🎓 EDUCATION
Crafting AI-Complementary Skills and Bulletproof Assessments / The Great Gender Divergence, Substack (sorry) (10 minute read)
AI-complementary skills: My assessments are designed to evaluate and encourage the development of skills that complement AI capabilities, such as critical thinking and creative problem-solving. Even if students copy and paste texts into Claude, it still struggles to think laterally.
Researchers combat AI hallucinations in math — Success in algebra but more work needed in statistics / The Hechinger Report (8 minute read)
related (1), Why AI Struggles with Math: Understanding the Limits of Artificial Intelligence / Geek Sided (3 minute read)
related (2) Gauth — Your AI Homework Helper
the variation on all the others is that real tutors are available if the AI doesn’t help
Mark Cuban Foundation Collaborates with Skillsoft’s Codecademy to Expand AI Education / press release (5 minute read)
The Mark Cuban Foundation, an organization dedicated to empowering high school students through free bootcamps with an AI-centered curriculum, is proud to announce a new collaboration with Skillsoft’s Codecademy, a leading online learning platform for technology skills.
related, AI Could Actually Help Rebuild The Middle Class — AI doesn’t have to be a job destroyer. It offers us the opportunity to extend expertise to a larger set of workers. / Noema Magazine (37 minute read)
Make AI tools to reduce teacher workloads, tech companies urged — Special access granted to DfE resources to train AI models to generate workbooks and lesson plans / The Guardian (4 minute read)
“Artificial intelligence, when made safe and reliable, represents an exciting opportunity to give our school leaders and teachers a helping hand with classroom life,” Morgan said.
Post-apocalyptic education / One Useful Thing (Ethan Mollick), Substack (sorry) (11 minute read)
A recent deep qualitative study of teachers found that teachers who used AI for both output (create a worksheet, develop a quiz) and to help with input (help me think through what makes a Great American novel, give me ways to explain positive and negative numbers) get more value than if they use AI for producing output alone.
📊 DATA & TECHNOLOGY
GenderCARE: A Comprehensive Framework for Assessing and Reducing Gender Bias in Large Language Models / arxiv (66 minute read)
related, source code, datasets, and scripts for the paper / kstanghere, GitHub
What’s Really Going On in Machine Learning? Some Minimal Models—Stephen Wolfram Writings / Stephen Wolfram (68 minute read)
we admittedly only understood about half of this, but that half was useful
related (1), Stephen Wolfram thinks we need philosophers working on big questions around AI / Tech Crunch (13 minute read)
If the main work of philosophy is to answer big existential questions, he sees us coming into a golden age of philosophy due to the growing influence of AI and all of the questions that it's raising.
related (2), Transformer Explainer: LLM Transformer Model Visually Explained / poloclub, GitHub (12 minute read)
How Will AI Impact Project Management? / Scott Ambler (9 minute read)
For individuals, many of these changes free up existing project managers from busy work. This is great for them because it puts them in a position to focus on adding value. This new focus may also in turn lead to working in a more strategic rather than tactical manner, enabling project managers to move into positions of influence within their organizations.
Ghost in the Machine: The AI Value Chain / AI Supremacy, Substack (sorry) (20 minute read)
This is going to be a deep dive, so you’ll want to read this on the web. Take your time, if investing in AI matters to you this is not one you’ll want to skip or skim.
Dear AWS, please let me be a cloud engineer again / Luc van Donkersgoed’s Notes (5 minute read)
GenAI allows us to add new features, and often faster than before. But GenAI has no value without an existing product to apply it to.
🎉 FUN and/or PRACTICAL THINGS
Why AI can't spell ‘strawberry’ / Tech Crunch (14 minute read)
The failure of large language models to understand the concepts of letters and syllables is indicative of a larger truth that we often forget: These things don’t have brains. They do not think like we do. They are not human, nor even particularly humanlike.
related, This founder had to train his AI not to Rickroll people / Tech Crunch (12 minute read)
Lindy’s accidental Rickroll is particularly remarkable because the AI organically reproduced this very specific user behavior, which informed its hallucination.
Off to Norway, With Three A.I. Travel Assistants / New York Times (11 minute read)
Overall, Mindtrip — with its polished, dynamic interface that allowed me to cross-check details with maps, links and reviews — was my favorite. While it gave some good recommendations, Mindtrip needed more prompting than Vacay, which offered a wider variety of suggestions in more detail.
related, The do’s and don’ts of using AI to plan your travel / Popular Science (7 minute read)
useful advice for using any AI chatbot:
Do use AI chatbots for inspirations
It’s always worth double-checking information using other sources
Don’t leave without making checks
Create Calendar Entries with Anthropic Claude 3.5 / Greg Wilson’s Tech Blog (5 minute read)
SkimIt.ai — Get an ai summary of any article delivered to your inbox.
How Roblox Is Using Generative AI to Enhance User Experiences / AI Podcast, Nvidia (29 minute podcast)
Noah Kravitz speaks with Anupam Singh, vice president of AI and growth engineering at Roblox, on how the company is using the technology to enhance virtual experiences with features such as automated chat filters and real-time text translation, which help build inclusivity and user safety.
An Experiment in Lust, Regret and Kissing / New York Times (18 minute read)
This summer, I agreed to a literary experiment with Times Opinion: What is the difference between a story written by a human and a story written by artificial intelligence?
the human-written story is subjectively more enjoyable to read; was closer than we expected
🧿 AI-ADJACENT
This new open-source AI, CogVideoX, could change how we create videos forever / Venture Beat (4 minute read)
The crown jewel of the project, CogVideoX-5B, boasts 5 billion parameters and produces 720×480 resolution videos at 8 frames per second. While these specs may not match the bleeding edge of proprietary systems, CogVideoX’s open-source nature is its true innovation.
try it: CogVideoX-5B / thudm, Hugging Face Space
even makes your prompt more precise
⋄