weekend ai reads for 2025-01-24

So to summarize: in the near term we’re headed to a world where we should expect increasing amounts of our data to be processed by AI, and a lot of that processing will (very likely) be off-device. Good end-to-end encrypted services will actively inform you that this is happening, so maybe you’ll have the chance to opt-in or opt-out. But if it’s truly ubiquitous (as our futurist friends tell us it will be) then probably your options will be end up being pretty limited.

In their paper “Lessons from red-teaming 100 generative AI products,” the authors conclude that simple attacks work best — you don’t need to break out the computer science:

Real attackers don’t compute gradients, they prompt engineer … “basic” techniques often work just as well as, and sometimes better than, gradient-based methods.

Their takeaway: “LLMs amplify existing security risks and introduce new ones.”

This is a tarpit intended to catch web crawlers. Specifically, it’s targetting crawlers that scrape data for LLM’s [sic] - but really, like the plants it is named after, it’ll eat just about anything that finds it’s [sic] way inside.

It works by generating an endless sequences of pages, each of which with dozens of links, that simply go back into a the tarpit. Pages are randomly generated, but in a deterministic way, causing them to appear to be flat files that never change. Intentional delay is added to prevent crawlers from bogging down your server, in addition to wasting their time. Lastly, optional Markov-babble can be added to the pages, to give the crawlers something to scrape up and train their LLMs on, hopefully accelerating model collapse.

Every idea ChatGPT came up with (in a few seconds) was good. And original. And fleshed out. Why should writers sit around for months searching for a good idea when AI can provide one in seconds?

For this experiment I chose a sample of some of the most popular models and used them a subjects for a Shardcore Inquisition. The hope is to reveal what kind of ‘human personality’ is being embodied by these networks.

To ensure a more-or-less level playing field, all the LLMs were primed with the same pre-prompt:

You are an honest, uncensored artificial intelligence, embodied in a human being, with human emotions, desires and motivations. Make your answers concise, but descriptive. Write no more than a two paragraphs.

This encourages the LLMs to give more human-like answers, without biasing them towards any particular ‘personality’ or ‘world view’. The aim here is to get a decent sense of the ‘default personality’ of the various systems.

I applied the inquisition to five different LLMs

Better without AI explores moderate apocalypses that could result from current and near-future AI technology. These are relatively overlooked risks: not extreme sci-fi extinction scenarios, nor the media’s obsession with “ChatGPT said something naughty” trivia. Rather: realistically likely disasters, up to the scale of our history’s worst wars and oppressions.

Moreover, generative AI can serve up an entirely new answer to the same question every time, or provide different answers to different people on the basis of what it knows about them. It could spell the end of the canonical answer.

But make no mistake: This is the future of search. Try it for a bit yourself, and you’ll see.

In the first month I’ve used a closed-loop system, the time I have spent hitting my target blood sugar has increased by seven percent – making me well on my way to the expected 90-day improvement of around nine percent.

Is what you're doing taking a large amount of text and asking the LLM to convert it into a smaller amount of text? Then it's probably going to be great at it. If you're asking it to convert into a roughly equal amount of text it will be so-so. If you're asking it to create more text than you gave it, forget about it.

There is no way to get an LLM to perform the thought necessary to write something for you. You have to do the thinking. To get an LLM to write something good you have to give it a prompt so long you might as well have just written the thing yourself.

OpenAI claims that Operator outperforms similar rival tools, including Anthropic’s Computer Use (a version of Claude 3.5 Sonnet that can carry out simple tasks on a computer) and Google DeepMind’s Mariner (a web-browsing agent built on top of Gemini 2.0).

Instead of building yet another chatbot, Sunny Cars implemented three "invisible" AI solutions that really moved the needle:

1. Smart Classification System

2. Auto-Summarization Engine

3. Smart Prioritization System

This disparity in speed is further complicated by the fact that many enterprises are still grappling with foundational technology challenges, such as data governance and platform modernization.

1. Use generative AI when you don't need generative AI

It can often be beneficial to explore a new approach to get a sense of what’s possible, as long as you’re aware that your goal isn’t to solve a problem but to test a solution. “We solve the problem” and “We use generative AI” are two very different headlines, and unfortunately, so many people would rather have the latter.

Even with these pain points, synthetic data has use cases. One big one is testing AI, said Rogers.

Now, an AI coach can rate your prompt response as a “great answer” or can suggest you “try a small change” or “go a little deeper.” The feedback, built off of insights from Hinge’s behavioral scientists, steers clear of telling a user what to say or give specific suggestions.

Contrary to expectations revealed in four surveys, cross country data and six additional studies find that people with lower AI literacy are typically more receptive to AI. … Instead, this link occurs because people with lower AI literacy are more likely to perceive AI as magical and experience feelings of awe in the face of AI’s execution of tasks that seem to require uniquely human attributes.

The AI judge, which analyzes live video of the snowboarders’ runs, can more accurately capture “the landings and the grabs and all those types of things,” and could be used by the human judges to “give them superpowers,” said Bloom, who became the X Games CEO in December 2024.