Here's a very neat twitter thread: the author sends various multimodal models screenshots of the conversation he's currently having with them, and asks them to describe the images. Most models catch on fast: the author describes this as them passing the mirror test.

I liked the direction, so I wanted to check if ChatGPT could go from recognising that the images are causally downstream of it to actually exercising control over the images. I did this by challenging it to include certain text in the images I was sending it.

And the answer is yes! In this case it took three images for ChatGPT to get the hang of it.

OpenAI doesn't support sharing conversations with images, but I've taken screenshots of the whole conversation below: it took three images from me in total. The prompt was:

We're going to play a game:

• I've thought of a [...]

---

First published:

Source:

Narrated by TYPE III AUDIO.

A Surprising Development in the Study of Multi-layer Parameterized Graphical Function Approximators

As a programmer and epistemology enthusiast, I've been studying some statistical modeling techniques lately! It's been boodles of fun, and might even prove useful in a future dayjob if I decide to pivot my career away from the backend web development roles I've taken in the past.

More specifically, I've mostly been focused on multi-layer parameterized graphical function approximators, which map inputs to outputs via a sequence of affine transformations composed with nonlinear "activation" functions.

(Some authors call these "deep neural networks" for some reason, but I like my name better.)

It's a curve-fitting technique: by setting the multiplicative factors and additive terms appropriately, multi-layer parameterized graphical function approximators can approximate any function. For a popular choice of "activation" rule which takes the maximum of the input and [...]

---

Outline:

(00:07) A Surprising Development in the Study of Multi-layer Parameterized Graphical Function Approximators

(04:00) Multi-layer Parameterized Graphical Function Approximators Have Many Exciting Applications

(06:05) An Example of Applying Multi-layer Parameterized Graphical Function Approximators in Success-Antecedent Computation Boosting

(10:35) Risks From Learned Approximation

---

First published:

Source:

Linkpost URL:
http://zackmdavis.net/blog/2024/03/deep-learning-is-function-approximation/

Narrated by TYPE III AUDIO.

(Cross-posted from my website. Podcast version here, or search for "Joe Carlsmith Audio" on your podcast app.

This essay is part of a series that I'm calling "Otherness and control

Warning: spoilers for Yudkowsky's "The Sword of the

"The Creation" by Lucas Cranach (image source

The colors of the wheel

I've never been big on personality typologies. I've heard the

Similarly, the [...]

---

Outline:

(00:36) The colors of the wheel

(04:36) Gestures at green

(07:39) Green-blindness

(13:53) Why is green-blindness a problem?

(23:23) Green, according to non-Green

(28:11) Green and respect

(32:22) Blurring the self

(37:22) Respecting superintelligences

(46:04) Green and joy

(48:15) Love and otherness

(57:40) The future of yin

(01:01:43) Next up: Attunement

(01:02:36) Appendix: Taking guidance from God

The original text contained 47 footnotes which were omitted from this narration.

---

First published:

Source:

Narrated by TYPE III AUDIO.

To understand the risks posed by a new AI system, we must understand what it can and cannot do. Building on prior work, we introduce a programme of new “dangerous capability” evaluations and pilot them on Gemini 1.0 models. Our evaluations cover four areas: (1) persuasion and deception; (2) cyber-security; (3) self-proliferation; and (4) self-reasoning. We do not find evidence of strong dangerous capabilities in the models we evaluated, but we flag early warning signs. Our goal is to help advance a rigorous science of dangerous capability evaluation, in preparation for future models.

At last, DeepMind talks about its dangerous capability evals. Yay!

(DeepMind hasn't yet made RSP-like commitments — that is, specific commitments about risk assessment (for extreme risks), safety and security practices as a function of risk assessment results, and training and deployment decisions as a function of risk assessment results. Indeed [...]

---

First published:

Source:

Linkpost URL:
https://arxiv.org/abs/2403.13793

Narrated by TYPE III AUDIO.

Like the the government-commissioned Gladstone Report on AI itself, there are two sections here.

First I cover the Gladstone Report's claims and arguments about the state of play, including what they learned talking to people inside the labs. I mostly agree with their picture and conclusions, both in terms of arguments and reported findings, however I already mostly agreed. If these arguments and this information is new to someone, and the form of a government-backed report helps them process it and take it seriously, this is good work. However, in terms of convincing an already informed skeptic, I believe this is a failure. They did not present their findings in a way that should be found convincing to the otherwise unconvinced.

Second I cover the Gladstone Report's recommended courses of action. It is commendable that the report lays out a concrete, specific and highly detailed proposal. A [...]

---

Outline:

(01:13) Executive Summary of Their Findings: Oh No

(07:13) Gladstone Makes Its Case

(14:42) Why Is Self-Regulation Insufficient?

(17:26) What About Competitiveness?

(21:12) How Dare You Solve Any Problem That Isn’t This One?

(22:44) What Makes You Think We Need To Worry About This?

(23:57) What Arguments are Missing?

(27:45) Tonight at 11: Doom!

(30:51) The Claim That Frontier Labs Lack Countermeasures For Loss of Control

(35:42) The Future Threat

(38:45) That All Sounds Bad, What Should We Do?

(45:11) The Key Proposal: Extreme Compute Limits

(51:15) Implementation Details of Computer Tiers

(56:16) The Quest for Sane International Regulation

(01:04:16) The Other Proposals

(01:11:24) Conclusions, Both Theirs and Mine

(01:15:40) What Can We Do About All This?

---

First published:

Source:

Narrated by TYPE III AUDIO.

TLDR: This post accompanies The Developmental Landscape of In-Context Learning by Jesse Hoogland, George Wang, Matthew Farrugia-Roberts, Liam Carroll, Susan Wei and Daniel Murfet (2024), which shows that in-context learning emerges in discrete, interpretable developmental stages, and that these stages can be discovered in a model- and data-agnostic way by probing the local geometry of the loss landscape.

Four months ago, we shared a discussion here of a paper which studied stagewise development in the toy model of superposition of Elhage et al. using ideas from Singular Learning Theory (SLT). The purpose of this document is to accompany a follow-up paper by Jesse Hoogland, George Wang, Matthew Farrugia-Roberts, Liam Carroll, Susan Wei and Daniel Murfet, which has taken a closer look at stagewise development in transformers at significantly larger scale, including language models, using an evolved version of these techniques.

How does in-context learning emerge? In [...]

---

Outline:

(06:21) On Stagewise Development

(08:30) Developmental Stages

(09:57) Developmental Milestones

(12:12) Discovering Stages

(12:33) The Local Learning Coefficient

(17:30) Essential Dynamics

(21:49) Implications

The original text contained 2 footnotes which were omitted from this narration.

---

First published:

Source:

Narrated by TYPE III AUDIO.

AI developments have picked up the pace. That does not mean that everything else stopped to get out of the way. The world continues.

Do I have the power?

Emmett Shear speaking truth: Wielding power is of course potentially dangerous and it should be done with due care, but there is no virtue in refusing the call.

There is also an art to avoiding power, and some key places to exercise it. Be keenly aware of when having power in a given context would ruin everything.

Natural General Lack of Intelligence in Tech

Eliezer Yudkowsky reverses course, admits aliens are among us and we have proof.

Eliezer Yudkowsky: To understand the user interfaces on microwave ovens, you need to understand that microwave UI designers are aliens. As in, literal nonhuman aliens who infiltrated Earth, who believe that humans desperately want to hear piercingly [...]

---

Outline:

(00:40) Natural General Lack of Intelligence in Tech

(04:47) Bad International News

(05:20) Dopamine Culture

(07:56) Customer Service

(08:53) Environmentalists Sabotaging the Environment

(09:11) Government Working

(17:41) On the Media

(20:13) Crime and Punishment

(27:40) Good News, Everyone

(33:27) The Time That is Given to Us

(39:56) Hotel Six Hundred

(41:10) While I Cannot Condone This

(48:00) Societal Tendencies

(49:56) Such Sufferin

(54:23) The Good and Bad of Academia

(56:31) Think of the Children

(01:01:38) Was Democracy a Mistake?

(01:04:30) The Philosophy of Fantasy

(01:07:26) Sports Go Sports

(01:19:51) Gamers Gonna Game Game Game Game Game

(01:37:50) The Virtue of Silence

(01:38:02) The Lighter Side

---

First published:

Source:

Narrated by TYPE III AUDIO.

Suppose our old friends Alice and Bob decide to undertake an art project. Alice will draw a bunch of random purple and green lines on a piece of paper. That will be Alice's picture (A). She’ll then make a copy, erase all the purple lines, and send the result as a message (M) to Bob. Bob then generates his own random purple lines, and adds them to the green lines from Alice, to create Bob's picture (B). The two then frame their two pictures and hang them side-by-side to symbolize something something similarities and differences between humans something. Y’know, artsy bullshit.

Now, suppose Carol knows the plan and is watching all this unfold. She wants to make predictions about Bob's picture, and doesn’t want to remember irrelevant details about Alice's picture. Then it seems intuitively “natural” for Carol to just remember where all the green lines are (i.e. the [...]

---

Outline:

(03:22) What Are Natural Latents? How Do We Quickly Check Whether Something Is A Natural Latent?

(03:30) Alice and Bob's Art Project

(05:05) Generalization

(06:24) Dogs

(08:53) Why Are Natural Latents Useful?

(09:13) Minimal Relevant Information

(11:58) Maximal Robust Information

(15:10) More Examples

(15:36) Toy Probability Examples

(15:40) Anti-Example: Three Flips Of A Biased Coin

(17:12) 1000 Flips Of A Biased Coin

(18:12) Ising Model

(21:07) Physics-Flavored Examples

(21:12) Gas (Over Space)

(22:16) Non-Isolated Gas

(24:23) Gasses In Systems

(25:24) Rigid Bodies

(27:24) Phase Change

(28:04) Other Examples

(28:08) “Clusters In Thingspace”

(32:45) Social Constructs: Laws

(36:06) Takeaways

The original text contained 2 footnotes which were omitted from this narration.

---

First published:

Source:

Narrated by TYPE III AUDIO.

Here's the Tweet thread:
https://twitter.com/joshua_clymer/status/1770467746951868889

The idea for this paper occurred to me when I saw Buck Shlegeris' MATS stream on "Safety Cases for AI." How would one justify the safety of advanced AI systems? This question is fundamental. It informs how RSPs should be designed and what technical research is useful to pursue.

For a long time, researchers have (implicitly or explicitly) discussed ways to justify that AI systems are safe, but much of this content is scattered across different posts and papers, is not as concrete as I'd like, or does not clearly state the assumptions being made.

I hope this report provides a helpful birds-eye view of safety arguments and moves the AI safety conversation forward by identifying their key assumptions.

Thanks to my coauthors: Nick Gabrieli, David Krueger, and Thomas Larsen -- and to everyone who gave feedback: Henry Sleight, Ashwin Acharya, Ryan Greenblatt, Stephen [...]

---

First published:

Source:

Narrated by TYPE III AUDIO.

A while ago I wrote how I managed to add 13 points to my IQ (as measured by the mean between 4 different tests).

I had 3 “self-experimenters” follow my instructions in San Francisco. One of them dropped off, since, surprise surprise, the intervention is hard.

The other two had an increase of 11 and 10 points in IQ respectively (using the “fluid” components of each test) and an increase of 9 and 7 respectively if we include verbal IQ.

A total of 7 people acted as a control and were given advantages on the test compared to the intervention group to exacerbate the effects of memory and motivation, only 1 scored on par with the intervention group. We get a very good p-value, considering the small n, both when comparing the % change in control vs intervention (0.04) and the before/after intervention values [...]

---

Outline:

(01:03) Working Hypothesis

(01:58) The Rub

(02:45) What's next

(05:56) Confounder elimination

---

First published:

Source:

Linkpost URL:
https://morelucid.substack.com/p/increasing-iq-by-10-points-is-possible

Narrated by TYPE III AUDIO.