weekend ai reads for 2024-12-13

December 13, 2024

📰 ABOVE THE FOLD: END OF AN ERA

via david, The GPT Era Is Already Ending / The Atlantic (17 minute read)

Here is another way to think about the distinction between language models and reasoning models: OpenAI’s attempted path to superintelligence is defined by parrots and rats. ChatGPT and other such products—the stochastic parrots—are designed to find patterns among massive amounts of data, to relate words, objects, and ideas. o1 is the maze-running rodent, designed to navigate those statistical models of the world to solve problems.

  • different to “the AI era is already ending”

Will A.I. Eat The Browser? / Crazy Stupid Tech (Om Malik) (11 minute read)

But as AI disaggregates information from text, video, and music into unique remixable AI chatbot answer streams, it’s clear to me that over the next decade the browser will need to adapt or die.

AI Company That Made Robots For Kids Goes Under, Robots Die — Embodied, maker of the AI robot called Moxie, is shuttering. With their closing, parents have to explain to their kids that Moxie is dead. / Aftermath (4 minute read)

  • there are allegedly videos on social media of kids crying as they’re told their toy isn’t coming back, but we don’t truck in videos of kids being exploited for their parents’ view counts

  • related, Closing FAQs / Moxie Robot (5 minute read)

Yes, we’ve included a letter from the G.R.L. (Global Robotics Lab) to help guide this conversation in an age-appropriate way.

Densing Law of LLMs / Tsinghua University, arXiv (46 minute read)

  • tl;dr: inference costs are dropping; edge AI is becoming more important (we’ve been saying and hearing this for a decade), compression isn’t the same as densing

  • forgot where we saw this, but “based on the fitted curve, maybe in 2025-12, we can have a GPT-4-level LLM with only 8B parameters”; Llama 3.3 has done this with 70B parameters

  • related, Meta launches Llama 3.3, shrinking powerful 405B open model / Venture Beat (5 minute read)

 

đŸ“» QUOTES OF THE WEEK

Many of the mediocre executives I meet, particularly those I meet in the data governance space, love their PowerPoints and Jira boards because while they are nonsense, they are nonsense that looks non-fuzzy and you will only have to deal with their inaccuracy once every few years, at which point so many people signed off on the clear-but-wrong vision of reality that it’s hard to tell who is ultimately accountable for the failure.

Nikhil Suresh (source, via david)

 

Knowing when to use AI turns out to be a form of wisdom, not just technical knowledge. Like most wisdom, it’s somewhat paradoxical: AI is often most useful where we’re already expert enough to spot its mistakes, yet least helpful in the deep work that made us experts in the first place.

Ethan Mollick (source)

 

đŸ‘„ FOR EVERYONE

Gemini 2.0: what’s new in Google’s new flagship AI model — Gemini 2.0 can generate images and audio, is faster and cheaper to run, and is meant to make AI agents possible. / The Verge (6 minute read)

The really cool thing about Gemini 2.0 is the brand new streaming API. This lets you open up a two-way stream to the model sending audio and video to it and getting text and audio back in real time.

  • PDF with prompts, aka: “assistant rules”

How to stop the AI you’re using from training with your data — Some apps make it simple; others make it nearly impossible. / The Verge (7 minute read)

 

📚 FOUNDATIONS

Best AI image generators of 2024 / Tom’s Guide (21 minute read)

  • plenty of examples

  • spoiler, number one is Midjourney

AI Model Comparison — See and compare every AI model easily. 100% free & open-source. / Countless.dev (20 minute read)

 

🚀 FOR LEADERS

FOMO is not a strategy / Rachel Coldicutt, Just enough Internet (10 minute read)

And regardless of what your friendly neighbourhood management consulting firm will tell you, there’s no one singular set of mitigations to get around this – technology will work best in your workplace if it’s rolled out in tune with existing culture, routines, and ways of working.

  • you know who you are; cut it out

More than half of leaders haven’t trained in AI, survey shows — A lack of training could negatively affect security, privacy and corporate competitiveness, General Assembly says. / HR Dive (6 minute read)

The Management Singularity / Programmable Mutter, Substack (sorry) (27 minute read)

LLMs provide big organizations with a brand new toolkit for organizing and manipulating information. It is far from a perfect toolkit: actually existing LLMs work best where you are prepared to tolerate a certain amount of slop, tend to bland out things so that the interesting weirdnesses disappear etc. But there is a lot that it can do, and there are a few applications (discussed below), where they will work very well indeed.

  • fictional story about the role(s) of AI in the future

  • starts with an alarm “nudging [the CEO] out of his slumber at the most optimal time in his REM cycle”, just so you know what to expect

The Secret Weapon Helping Businesses Get Results From AI: Humans — It turns out the latest machines need us as much as we need them / Wall Street Journal (9 minute read)

Every company I talked with mentioned that to get real value out of their shiny new generative AI systems—no matter the application—they needed to overhaul or double down on their strategy for feeding it the kind of data that today’s AI excels at processing—“unstructured” data.

 

🎓 FOR EDUCATORS

Beyond breeding laziness, lack of time management, and undermining integrity amongst students and teachers alike, it also inhibits any form of deep connection and understanding a teacher may form of their students.

  • lots of insights from students

MathGPT — AI Math Solver & Homework Helper

KPMG Experts Outline 4 Main Approaches to AI in Higher Ed / Government Technology (5 minute read)

 

📊 FOR TECHNOLOGISTS

Four gen AI shifts that will reshape enterprise technology / McKinsey Digital (16 minute read)

1. From tools that support teams to AI ‘artisan’ and ‘factory’ teams

2. From application architectures dominating the landscape to predominantly AI agent and data architectures

3. From a ‘pyramid’ or ‘diamond’ organizational structure to a flatter one, with new workforce development considerations

4. From application- to infrastructure-based cost structures, with increased focus on compute spend

AnythingLLM — The all-in-one AI application for everyone

  • related, OpenRouter — A unified interface for LLMs

Nobody Gets Fired for Picking JSON, but Maybe They Should? / Miguel Young de la Sota (15 minute read)

Suppose you have a JSON document somewhere that includes a user ID and a transcript of their private messages with another user. Data loss due to rounding would result in the wrong user ID being associated with the private messages, which could result in leaking PII or incorrect management of privacy consent (such as GDPR requirements).

  • if someone is building something using JSON (especially with Python!), maybe ask some questions

  • one solution: Protocol Buffers

 

🎉 FOR FUN

AI Santa Video — Free personalized videos from Santa to your loved ones.

“Feels Like Paper!” — Interfacing Artificial Intelligence through Paper / Lukas Moro (10 minute read)

“Feels Like Paper!” is a series of prototypes about augmenting physical paper through AI. Various ML and LLMs are used to infuse physical paper and ink with properties of the digital world without compromising on their physical traits.

  • interesting videos of the prototypes

Manchester City is letting fans design its new kit with AI — Players will wear the winning AI-designed uniform during the 2026-2027 season. / The Verge (3 minute read)

  • stop letting “fans” name or design anything

 

🧿 AI-ADJACENT

52 things I learned in 2024 / Tom Whitwell, Medium (11 minute read)

2. If you run one specific, but illegal, database query on a set of widely used health data, you can access Tony Blair’s entire personal medical history.

  • always the best

  • not AI-adjacent at all but lovely tribute by Spike Jonze & Mary Wigmore to Dick Van Dyke and aging gracefully, regardless of how you feel about Coldplay

 

⋄