Open AI strikes back: ChatGPT's new feature will end AI slop

I have a source at OpenAI that leaked a major feature they're going to put out, that will make up the ground they're losing to Claude. Here's an early draft of the press release.

***

OpenAI is rolling out a new feature with ChatGPT model 5.5 that provides special resistance to an ongoing issue with AI output: the telltale signs that the output is AI, which is a major turnoff and source of frustration on the internet.

The new feature, dubbed Preview Plus(tm), lets users go through ChatGPT's output and actually purge all the AI pain points.

• Em-dashes? Users can delete them.
• "It's not X but Y"? It's "X in Y clothing"? Users can replace it with their own authentic version a human would actually say.
• Hallucinated legal citation? Users can insert something from actual case law that they have personally validated applies to the matter in question.

Preview Plus(tm) currently only available to paid plan subscribers and users who have opted in to early feature testing.

OpenAI CEO Sam Atman is hopeful about the new direction Preview Plus(tm) will take ChatGPT. Earlier today, he told a press conference, "Preview Plus empowers users to actually stand between ChatGPT and the eventual consumer of its output artifacts. It provides a seamless way to merge human intelligence with cutting-edge AI, getting the best of both worlds."

OpenAI gave a screenshot of an early version of the feature that shows a user removing the dreaded em-dash. 

 

 

A serious paper on bits as Joules per Kelvin

My ramblings about thermodynamics aren't so off-base, it turns out!

Remember this one? From 2009? Where I explained how Joules per Kelvin (energy per unit temperature) is a valid measure of information (or entropy, effectively "missing information"), which is normally measured in bits?

Well, now there's a paper that formalizes that idea and related ones. As the title ("Temperature as Joules per Bit") indicates, it looks at a rearranged version of the same insight (mine was "bits as Joules per temperature"). But, it also goes a lot deeper and derives thermodynamics starting from entropy to understand temperature, rather than the other way around, as is conventionally done.

Related thought: I remember back in that 2009 thermodyanmics/info theory frenzy, one of my goals was rederive the Carnot limit based on information-theoretic considerations -- that is, show it as a simple implication of the amount of knowledge you have about a system in a case where only know the temperature difference. (Naturally, I assumed someone had already done this and tried to find it but it was very hard to google for.)

Background: The Carnot limit tells you the maximum amount of mechanical work ("useful energy") you can extract from heat -- like, through an engine -- and, as it turns out, it's a function of the ratio of absoute temperatures you're working between. You don't face this limit when extracting work from a flywheel (spinning disc with grooves). Inspired by an counterintuitive insight in an Eliezer Yudkowsky LessWrong post, and my thoughts about it, I figured you could draw a more direct line from "knowledge of a temperature difference" to "how much energy is extractable".

Now I'll give it a go with ChatGPT, and post my findings!

ChatGPT4 refines my argument about fish being bad swimming instructors

I finally splurged for a paid ChatGPT subscription! With access to GPT-4, I asked it to refine my earlier argument about fish being bad swimming instructors. Here's what it gave me:

***

Certainly, let's rework the argument using more scientifically grounded explanations and references:

While it's true that fish, with their adept aquatic abilities, may seem like logical swimming instructors for humans at first glance, upon closer examination, it becomes clear that this is far from practical or effective. Even if we put aside the substantial communication and physical barriers, there are fundamental biological and behavioral differences between fish and humans that pose significant challenges.

Humans are terrestrial, bipedal mammals, and our motor skills and locomotive mechanisms are naturally adapted to this lifestyle. Fish, in contrast, are aquatic creatures designed to maneuver efficiently in water (Webb, P.W., 1982, "Locomotor patterns in the evolution of actinopterygian fishes"). Their ability to swim is an innate behavior, hardwired into their neural circuitry, not a learned skill (Gould, J.L., 1982, "Ethology: The Mechanisms and Evolution of Behavior"). Given this, fish cannot consciously comprehend or articulate the mechanics of their swimming abilities as humans would require in a teaching-learning scenario.

In the field of pedagogy, it is well accepted that the process of teaching involves more than simply demonstrating a skill; it requires an understanding of learning styles, the ability to give and receive feedback, and a degree of self-awareness in the teacher (Boud, D., et al., 1985, "Reflection: Turning Experience into Learning"). Fish, as far as current scientific consensus suggests, lack these cognitive capacities (Butler, A.B., et al., 2005, "Evolution of the Neural Basis of Consciousness: A Bird-Mammal Comparison").

Furthermore, many human skills, including swimming, are learned through a combination of instruction, practice, and feedback (Kirschner, P.A., et al., 2006, "Why Minimal Guidance During Instruction Does Not Work"). This model of learning doesn't align with the unconscious, reflex-driven behaviors seen in fish.

Lastly, it's important to acknowledge the vast physiological differences between humans and fish. Human limbs and fish fins have evolved for different purposes and environments, with humans developing more for grasping and manipulation and fish for efficient aquatic locomotion (Thorsen, D.H., et al., 2004, "Biomechanics of Locomotion in Asian Carp"). These differences are fundamental and cannot be bridged by simple observation or imitation.

In conclusion, despite their swimming prowess, fish cannot be effective swimming instructors for humans due to biological, behavioral, and cognitive differences, as well as practical communication and anatomical obstacles. This argument does not discount the value of observing aquatic creatures to improve our understanding of efficient movement in water, but it's clear that such observation is a far cry from active instruction.

What a Bayesian causal network on Newcomb's problem might look like

Explanation of Newcomb's problem.

ADDED: Oh, and here's the long discussion that led me to draw the network above. Click the picture to enlarge.

Overcoming bias can make you cynical

If you haven't already, start reading Eliezer Yudkowsky's series on intelligence, evolution, quantum physics, biases, morality, and whatever-else, now would be a good time to start.  The latest entries in this series on morality, though, may make you cynical, as he probes, among other weighty issues, the question of whether there's a difference between "I want this" and "It is right that I should get this."