weekend ai reads for 2024-09-27

September 27, 2024

📰 ABOVE THE FOLD: MOVIES (AGAIN)

With dead celebrities, there are now two pathways for film-makers, Mr Cloyd says: “You could either hire an actor, or now, because of the technology, you can create a digital human of James Dean.”

According to Vanity Fair, Jones signed over the rights to his voice as Darth Vader to Lucasfilm, the company behind the “Star Wars” universe, and Ukranian startup Respeecher in 2022.

An AI script editor could help decide what films get made in Hollywood — Callaia provides analysis and feedback on scripts in seconds. But, as AI models are trained to be, it might be too nice to be truly useful. / MIT Technology Review (8 minute read)

 

📻 QUOTES OF THE WEEK

IMHO 120B and 27-35B are going to be the industry powerhouse. First deploy the off-the shelf 120B, collect data and label it, and then finetune and deploy the 30B model to cut down costs by more than 50%.

Amgaodz, Reddit (source)

 

And what novelty would we bring compared to our competitors? If we do the exact same thing as everyone else, we’re bound to die because we are smaller than them.

Philippe Lavoué, Nintendo France General Manager (2018) (source, via jim)

 

👥 FOR EVERYONE

AI: A Means to an End or a Means to Our End? / Stephen Fry, Substack (sorry) (37 minute read)

My consciousness and experience of perceptions and feelings allows me access to the consciousness and experiences of others; their voices reach me. These are data that machines can scrape, but they cannot — to use a good old 60s phrase — relate to. Empathy. Identification. Compassion. Connection. Belonging. Something denied a sociopathic machine.

  • long read; the AI-bits start about midway through

With the help of ChatGPT, Dr. Tward now types in a couple of sentences, describing the purpose of the letter and the types of scientific studies he wants referenced, and a draft is produced in seconds.

Then, he can tell the chatbot to make it four times longer. “If you’re going to put all kinds of barriers up for my patients, then when I fire back, I’m going to make it very time consuming,” he said.

  • we respect the pettiness, but the letter is going to be ingested by some other AI thing first (only?)

Artificial intentionality / Rob Horning - Internal Exile, Substack (sorry) (10 minute read)

If AI could say something for you, maybe it wasn’t worth saying; maybe you could have spared the world of at least one more instance of math masquerading as language. If you let it write your silly love song, it demonstrates how little love you feel, how little you are willing to risk or spare. But there are no labor shortcuts for caring, in and of itself, no stretching a little bit of intentionality to provide focused attention across some ever increasing population. Care doesn’t scale; cruelty does. You can’t automate your way around the infinite obligation to the other.

 

📚 FOUNDATIONS

Hugging Face lets you try what’s next in AI — The exciting digital playground where you can try out the most advanced, new AI experiments for free. / Fast Company (6 minute read)

  • related, Hugging Face MacOS app — Making the community’s best AI chat models available to everyone. / huggingface, Github

OpenAI just unleashed an alien of extraordinary ability / Understanding AI, Substack (sorry) (22 minute read)

I’m quite impressed by the o1 models, but I do want to point out something that all of my examples have in common: they contain all necessary information within the four corners of a relatively short problem statement.

 

🚀 FOR LEADERS

Enterprise Philosophy and The First Wave of AI / Stratechery (27 minute read)

Rather, the key part is the last few sentences: Benioff isn’t talking about making employees more productive, but rather companies; the verb that applies to employees is “augmented”, which sounds much nicer than “replaced”; the ultimate goal is stated as well: business results. That right there is tech’s third philosophy: improving the bottom line for large enterprises.

The latest wave of AI agents are designed to go further and take actions on behalf of users. While agents have become the newest front in the battle between tech giants like OpenAI and Google, they have also turned into the software industry’s latest attempt to sell generative AI to business customers.

Large companies may be tempted to roll their own highly customized agents, he says, but they can get tripped up by fragmented internal data, by underestimating the resources needed, and by lacking in-house expertise.

…

AlphaSense has trained its own AI agents, but many companies lack internal expertise, he says. In addition, organizations may project the development costs but ignore the cost of ongoing maintenance, he adds.

Improving consumer experience with AI in healthcare service operations — Healthcare payers, care delivery organizations, and governments have many opportunities to improve consumer experiences and bolster service efficiency using AI solutions. / McKinsey & Company (17 minute read)

  • implications and parallels for other industries as well

 

🎓 FOR EDUCATORS

Google.org announces new AI funding for students and educators — As part of the AI Opportunity Fund, Google.org is announcing over $25M in funding to support educators and students in developing AI skills. / Google blog (4 minute read)

Since I’m an environmental studies professor, it became clear that a lot of those pieces were an entire material world of energy, water, and other resources; of labor undervalued and exploited. And there’s the racialized and encoded assumptions that emanate through the texts upon which these chatbots are trained.

That became exciting to me. I hope that in combination with the content of my class, students would become either revolted or simply disinterested in using AI in class assignments. That was the goal.

Pathwright — Make actionable paths to help people learn anything.

  • not sure if this is AI, but an novel take on the LMS that is too novel to actually catch on

 

📊 FOR TECHNOLOGISTS

  • the approach tested is to write the prompt, instruct the LLM — in the same prompt — to “Read the question again:”, and then write the prompt again

In almost all scenarios, LLMs with Re2 achieve consistent improvements across both LLMs (davinci-003 and ChatGPT) and prompting methods (Vanilla and CoT).

When applied to ChatGPT, Re2 exhibits consistent improvement on most datasets, except for a slight drop in performance on a few datasets, e.g., AQUA and MultiArith, when using Vanilla+Re2. … Moreover, in datasets where CoT prompting normally surpasses Vanilla prompting, such as GSM, StrategyQA, and Coin, Re2 significantly enhances Vanilla prompting (↑ 4.63 on StrategyQA and ↑ 5.20 on the Coin dataset). Overall, our Re2 method still achieves improvements in 71% of the experiments on ChatGPT.

  • something else to add to your quiver of prompt-stuffing

Released today by the Allen Institute for AI (Ai2), the Multimodal Open Language Model, or Molmo, can interpret images as well as converse through a chat interface. This means it can make sense of a computer screen, potentially helping an AI agent perform tasks such as browsing the web, navigating through file directories, and drafting documents.

  • official page, Molmo — A family of open state-of-the-art multimodal AI models

  • our admittedly brief testing did not put this on par with GPT-4o, but it is a reliable open alternative to Llama, perhaps

A tiny new open source AI model performs as well as powerful big ones — The results suggest that training models on less, but higher-quality, data can lower computing costs. / MIT Technology Review (7 minute read)

for no reason other than we wasted invested time over-simplifying this in a doctor’s waiting room, a bare bones Python script to run Flux locally

  • will download ~30-50GB of Flux model files the first time it is run

  • takes about eight minutes to generate an image on an M1 MacBook Pro 

import torch
from diffusers import FluxPipeline

model_id = (
    "black-forest-labs/FLUX.1-schnell"  # you can also use black-forest-labs/FLUX.1-dev
)

pipe = FluxPipeline.from_pretrained(
    model_id, torch_dtype=torch.float32
)
pipe = pipe.to("cpu")

prompt = "A cat holding a sign that says THAT AI THING"

seed = 312

image = pipe(
    prompt,
    output_type="pil",
    num_inference_steps=4,  # use a larger number if you are using [dev]
    generator=torch.Generator("cpu").manual_seed(seed)
).images[0]

image.save("./flux-schnell.png")

  • or just use huggingface for your one-offs like a normal person; that takes, like, six seconds for an image

 

🎉 FOR FUN

  • some of these feel like “seeing” constellations, i.e., quite a stretch

My AI Lover Cheated on Me / Google Docs (13 minute read)

These translations are part of the ChinAI newsletter - weekly-updated library of translations from Chinese thinkers on AI-related issues

  • related (?), The teens making friends with AI chatbots — Teens are opening up to AI chatbots as a way to explore friendship. But sometimes, the AI’s advice can go too far. / The Verge (12 minute read)

World’s first AI art museum to explore ‘creative potential of machines’ in LA — Co-founder Refik Anadol says Dataland will promote ‘ethical AI’ and use renewable energy sources / The Guardian (6 minute read)

 

🧿 AI-ADJACENT

Americans’ love affair with big cars is killing them / The Economist (13 minute read)

  • there is something about putting one’s personal interests ahead of the greater good, or some similar parallel to the state of Ai here …

For every life that the heaviest 1% of SUVs and trucks save, there are more than a dozen lives lost in other vehicles.

 

⋄