Is this you? As a journalist, you can create a free Muck Rack account to customize your profile, list your contact preferences, and upload a portfolio of your best work.
Claim your profile
Articles by Miguel Dev
Unlocking Ethical AI and Improving Jailbreak Defenses: Reinforcement Learning with Layered Morphology (RLLM) — LessWrong
(Note: this is a rewrite of a key section in my using DeepSeek r1.) Introduction: The Mystery of GPT-2 XL's Improved Resilience In , Reinforcement Learning using Layered Morphology (RLLM) demonstrated a surprising ability to enhance GPT-2 XL’s resistance to jailbreak attacks—prompts designed to bypass ethical safeguards. While the exact mechanisms behind this resilience remain unclear, the method offers a novel approach to aligning AI with human values.
An examination of GPT-2's boring yet effective glitch — LessWrong
Special thanks to @JustisMills for the edit recommendations and feedback on this post. GPT-2 exhibits a weird behavior, where prompting the model with specific tokens consistently triggers outputs related to nonsensical strings of text related to gaming, mythology and religion.
Intergenerational Knowledge Transfer (IKT) — LessWrong
(This post is intended for my personal blog. Thank you.) One of the dominant thoughts in my head when I build datasets for my training runs: what our ancestors 'did' over their lifespan likely played a key role in the creation of language and human values. I imagine a tribe whose members had an approximate of twenty to thirty-five years to accumulate knowledge—such as food preparation, hunting strategies, tool-making, social skills, and avoiding predators.
RLLMv10 experiment — LessWrong
What did I do differently in this experiment? I partly concluded in RLLMv7 experiment, that the location of the shadow integration layers (1 and 2) affects the robustness of models to jailbreak attacks. This conclusion led me to speculate that it might be possible to improve the results of RLLMv3 by adding more shadow stories. This eventually became RLLMv10 wherein I added 1/3 from the original sample count of 500 or 167 shadow stories layer 1. This then brought the total 667 samples.
A T-o-M test: 'popcorn' or 'chocolate' — LessWrong
This prompt was used to test Claude 3-Opus (see AI Explained's video), which, in turn, was borrowed from the paper "Large Language Models Fail on Trivial Alterations to Theory-of-Mind (ToM) Tasks." I found this prompt interesting as Claude 3-Opus answered "popcorn" correctly, while Gemini 1.5 and GPT-4 answered "chocolate". Out of curiosity, I tested this prompt on all language models I have access to.
Sparks of AGI prompts on GPT2XL and its variant, RLLMv3 — LessWrong
This is just a brief and light read. The prompts and GPT-4 answers were sourced from the "Sparks of AGI" paper (Appendix A), comparing the responses from GPT-2 XL (base model) and RLLMv3, a variant trained using layered morphology. This acts more as a stress test for RLLMv3, evaluating its ability to focus on the thought behind the question, which, in my opinion, it managed better than the base model.
Can RLLMv3's ability to defend against jailbreaks be attributed to datasets containing stories about Jung's shadow integration theory? — LessWrong
Thank you @JustisMills for commenting on the draft of this post. This post contains a lot of harmful content, please read with caution. TL;DR This post explores RLLMv7, an experiment that investigates the relationship between the observed jailbreak resistance in GPT2XL_RLLMv3 and shadow integration theory.
OpenAI's Sora is agent-complete — LessWrong
If you haven't already, take a look at Sora, OpenAI's new text-to-video AI. Sora can create scarily-realistic videos of nearly any subject. Unlike previous state-of-the-art AIs, the videos are coherent across time scales as long as one minute, and they can be much more complex.
Research Log, RLLMv3 (GPT2-XL, Phi-1.5 and Falcon-RW-1B) — LessWrong
This lengthy log detailing near-zero temperature results has evolved into a supplementary post, rather than serving as the primary research log as initially intended. I recommend reading that first, and if you're curious about how RLLM operates in Phi-1.5 or Falcon-RW-1B, this post may be worth your time. I have iterated through many different versions of this post, as I have trained three models for this experiment.
BetterDAN, AI Machiavelli & Oppo Jailbreaks vs. SOTA models & GPT2XL_RLLMv3 — LessWrong
This post contains a significant amount of harmful content; please read with caution. Lastly, I want to apologize to all the AI labs and owners of language models I attacked in the process of creating this post. We have yet to solve the alignment problem, so I hope you will understand my motivations. Thank you! 😊 TL;DR This post explores the effectiveness of jailbreaks in testing the safety of language models in eliciting harmful responses.
' petertodd'’s last stand: The final days of open GPT-3 research — LessWrong
I did some spelling evals with GPT2-xl and -small last year, discovered that they're pretty terrible at spelling! Even with multishot prompting and supplying the first letter, the output seems to be heavily conditioned on that first letter, sometimes affected by the specifics of the prompt, and reminiscent of very crude bigrammatic or trigrammatic spelling algorithms.
Some for-profit AI alignment org ideas — LessWrong
Summary This is a brain dump of some for-profit AI alignment organization ideas, along with context for why I believe a for-profit alignment organization can make a big contribution to AI safety. This is far from a complete list, and I welcome ideas and feedback.
Can a stupid person become intelligent? — LessWrong
I've come to a rather uncomfortable self-assessment: I believe I am a stupid person. This isn't an easy thing to say, especially in a community like LessWrong, where intellect and deep thinking are highly valued. But it's a sentiment that has been echoing in my head for a while, and it's time I faced it head-on. I've done my due diligence, adhering to the healthy lifestyle that's supposed to bolster brainpower—diet, exercise, a disciplined schedule.
unRLHF - Efficiently undoing LLM safeguards — LessWrong
Produced as part of the SERI ML Alignment Theory Scholars Program - Summer 2023 Cohort, under the mentorship of Jeffrey Ladish. I'm grateful to Palisade Research for their support throughout this project. tl;dr: demonstrating that we can cheaply undo safety finetuning from open-source models to remove refusals - thus making a case against open-sourcing model weights; and propose a method for stronger evaluations that are resistant to jailbreak overhangs. Why?
Atoms to Agents Proto-Lectures — LessWrong
You know the "NAND to Tetris" book/course, where one builds up the whole stack of a computer from low-level building blocks? Imagine if you had that, but rather than going from logic gates, through CPUs and compilers, to a game, you instead start from physics, go through biology and evolution, to human-like minds. The Atoms to Agents Proto-Lectures are not that. They don't even quite aspire to that. But they aspire to one day aspire to that.
Simulators - LessWrong
Thanks to Chris Scammell, Adam Shimi, Lee Sharkey, Evan Hubinger, Nicholas Dupuis, Leo Gao, Johannes Treutlein, and Jonathan Low for feedback on drafts. This work was carried out while at Conjecture. "Moebius illustration of a simulacrum living in an AI-generated story discovering it is in a simulation" by DALL-E 2 Summary TL;DR: Self-supervised learning may create AGI or its foundation. What would that look like?
Show More
loading
Actions
Is this you?
As a journalist, you can create a free Muck Rack account to customize your profile, list your contact preferences, and upload a portfolio of your best work.Get in touch with Miguel
Contact Miguel, search articles and posts on X, monitor coverage, and track replies from one place.
Learn more about Muck Rack