“Current safety training techniques do not fully transfer to the agent setting” by Simon Lermen, Govind Pimpale

LessWrong (Curated & Popular)

Player FM - Internet Radio Done Right

12 subscribers

Ditambah three tahun yang lalu

Kandungan disediakan oleh LessWrong. Semua kandungan podcast termasuk episod, grafik dan perihalan podcast dimuat naik dan disediakan terus oleh LessWrong atau rakan kongsi platform podcast mereka. Jika anda percaya seseorang menggunakan karya berhak cipta anda tanpa kebenaran anda, anda boleh mengikuti proses yang digariskan di sini https://ms.player.fm/legal.

about a year ago 10:10

MP3•Laman utama episod

TL;DR: I'm presenting three recent papers which all share a similar finding, i.e. the safety training techniques for chat models don’t transfer well from chat models to the agents built from them. In other words, models won’t tell you how to do something harmful, but they are often willing to directly execute harmful actions. However, all papers find that different attack methods like jailbreaks, prompt-engineering, or refusal-vector ablation do transfer.
Here are the three papers:

AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents
Refusal-Trained LLMs Are Easily Jailbroken As Browser Agents
Applying Refusal-Vector Ablation to Llama 3.1 70B Agents

What are language model agents
Language model agents are a combination of a language model and a scaffolding software. Regular language models are typically limited to being chat bots, i.e. they receive messages and reply to them. However, scaffolding gives these models access to tools which they can [...]
---
Outline:
(00:55) What are language model agents
(01:36) Overview
(03:31) AgentHarm Benchmark
(05:27) Refusal-Trained LLMs Are Easily Jailbroken as Browser Agents
(06:47) Applying Refusal-Vector Ablation to Llama 3.1 70B Agents
(08:23) Discussion
---
First published:
November 3rd, 2024
Source:
https://www.lesswrong.com/posts/ZoFxTqWRBkyanonyb/current-safety-training-techniques-do-not-fully-transfer-to
---
Narrated by TYPE III AUDIO.
---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

490 episod

“Current safety training techniques do not fully transfer to the agent setting” by Simon Lermen, Govind Pimpale

LessWrong (Curated & Popular)

12 subscribers

published about a year ago

Kongsi

MP3•Laman utama episod

AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents
Refusal-Trained LLMs Are Easily Jailbroken As Browser Agents
Applying Refusal-Vector Ablation to Llama 3.1 70B Agents

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

490 episod

Semua episod

1
“Why Have Sentence Lengths Decreased?” by Arjun Panickssery 9:08

2 days ago9:08

9:08

“In the loveliest town of all, where the houses were white and high and the elms trees were green and higher than the houses, where the front yards were wide and pleasant and the back yards were bushy and worth finding out about, where the streets sloped down to the stream and the stream flowed quietly under the bridge, where the lawns ended in orchards and the orchards ended in fields and the fields ended in pastures and the pastures climbed the hill and disappeared over the top toward the wonderful wide sky, in this loveliest of all towns Stuart stopped to get a drink of sarsaparilla.” — 107-word sentence from Stuart Little (1945) Sentence lengths have declined. The average sentence length was 49 for Chaucer (died 1400), 50 for Spenser (died 1599), 42 for Austen (died 1817), 20 for Dickens (died 1870), 21 for Emerson (died 1882), 14 [...] --- First published: April 3rd, 2025 Source: https://www.lesswrong.com/posts/xYn3CKir4bTMzY5eb/why-have-sentence-lengths-decreased --- Narrated by TYPE III AUDIO . --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts , or another podcast app.…

1
“AI 2027: What Superintelligence Looks Like” by Daniel Kokotajlo, Thomas Larsen, elifland, Scott Alexander, Jonas V, romeo 54:30

3 days ago54:30

54:30

In 2021 I wrote what became my most popular blog post: What 2026 Looks Like. I intended to keep writing predictions all the way to AGI and beyond, but chickened out and just published up till 2026. Well, it's finally time. I'm back, and this time I have a team with me: the AI Futures Project. We've written a concrete scenario of what we think the future of AI will look like. We are highly uncertain, of course, but we hope this story will rhyme with reality enough to help us all prepare for what's ahead. You really should go read it on the website instead of here, it's much better. There's a sliding dashboard that updates the stats as you scroll through the scenario! But I've nevertheless copied the first half of the story below. I look forward to reading your comments. Mid 2025: Stumbling Agents The [...] --- Outline: (01:35) Mid 2025: Stumbling Agents (03:13) Late 2025: The World's Most Expensive AI (08:34) Early 2026: Coding Automation (10:49) Mid 2026: China Wakes Up (13:48) Late 2026: AI Takes Some Jobs (15:35) January 2027: Agent-2 Never Finishes Learning (18:20) February 2027: China Steals Agent-2 (21:12) March 2027: Algorithmic Breakthroughs (23:58) April 2027: Alignment for Agent-3 (27:26) May 2027: National Security (29:50) June 2027: Self-improving AI (31:36) July 2027: The Cheap Remote Worker (34:35) August 2027: The Geopolitics of Superintelligence (40:43) September 2027: Agent-4, the Superhuman AI Researcher --- First published: April 3rd, 2025 Source: https://www.lesswrong.com/posts/TpSFoqoG2M5MAAesg/ai-2027-what-superintelligence-looks-like-1 --- Narrated by TYPE III AUDIO . --- Images from the article:…

1
“OpenAI #12: Battle of the Board Redux” by Zvi 18:01

3 days ago18:01

18:01

Back when the OpenAI board attempted and failed to fire Sam Altman, we faced a highly hostile information environment. The battle was fought largely through control of the public narrative, and the above was my attempt to put together what happened.My conclusion, which I still believe, was that Sam Altman had engaged in a variety of unacceptable conduct that merited his firing.In particular, he very much ‘not been consistently candid’ with the board on several important occasions. In particular, he lied to board members about what was said by other board members, with the goal of forcing out a board member he disliked. There were also other instances in which he misled and was otherwise toxic to employees, and he played fast and loose with the investment fund and other outside opportunities. I concluded that the story that this was about ‘AI safety’ or ‘EA (effective altruism)’ or [...] --- Outline: (01:32) The Big Picture Going Forward (06:27) Hagey Verifies Out the Story (08:50) Key Facts From the Story (11:57) Dangers of False Narratives (16:24) A Full Reference and Reading List --- First published: March 31st, 2025 Source: https://www.lesswrong.com/posts/25EgRNWcY6PM3fWZh/openai-12-battle-of-the-board-redux --- Narrated by TYPE III AUDIO . --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts , or another podcast app.…

1
“The Pando Problem: Rethinking AI Individuality” by Jan_Kulveit 27:39

3 days ago27:39

27:39

Epistemic status: This post aims at an ambitious target: improving intuitive understanding directly. The model for why this is worth trying is that I believe we are more bottlenecked by people having good intuitions guiding their research than, for example, by the ability of people to code and run evals. Quite a few ideas in AI safety implicitly use assumptions about individuality that ultimately derive from human experience. When we talk about AIs scheming, alignment faking or goal preservation, we imply there is something scheming or alignment faking or wanting to preserve its goals or escape the datacentre. If the system in question were human, it would be quite clear what that individual system is. When you read about Reinhold Messner reaching the summit of Everest, you would be curious about the climb, but you would not ask if it was his body there, or his [...] --- Outline: (01:38) Individuality in Biology (03:53) Individuality in AI Systems (10:19) Risks and Limitations of Anthropomorphic Individuality Assumptions (11:25) Coordinating Selves (16:19) Whats at Stake: Stories (17:25) Exporting Myself (21:43) The Alignment Whisperers (23:27) Echoes in the Dataset (25:18) Implications for Alignment Research and Policy --- First published: March 28th, 2025 Source: https://www.lesswrong.com/posts/wQKskToGofs4osdJ3/the-pando-problem-rethinking-ai-individuality --- Narrated by TYPE III AUDIO . --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts , or another podcast app.…

1
“OpenAI #12: Battle of the Board Redux” by Zvi 18:01

4 days ago18:01

18:01

1
“You will crash your car in front of my house within the next week” by Richard Korzekwa 1:52

4 days ago1:52

1:52

I'm not writing this to alarm anyone, but it would be irresponsible not to report on something this important. On current trends, every car will be crashed in front of my house within the next week. Here's the data: Until today, only two cars had crashed in front of my house, several months apart, during the 15 months I have lived here. But a few hours ago it happened again, mere weeks from the previous crash. This graph may look harmless enough, but now consider the frequency of crashes this implies over time: The car crash singularity will occur in the early morning hours of Monday, April 7. As crash frequency approaches infinity, every car will be involved. You might be thinking that the same car could be involved in multiple crashes. This is true! But the same car can only withstand a finite number of crashes before it [...] --- First published: April 1st, 2025 Source: https://www.lesswrong.com/posts/FjPWbLdoP4PLDivYT/you-will-crash-your-car-in-front-of-my-house-within-the-next --- Narrated by TYPE III AUDIO . --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts , or another podcast app.…

1
“My ‘infohazards small working group’ Signal Chat may have encountered minor leaks” by Linch 10:33

4 days ago10:33

10:33

Remember: There is no such thing as a pink elephant. Recently, I was made aware that my “infohazards small working group” Signal chat, an informal coordination venue where we have frank discussions about infohazards and why it will be bad if specific hazards were leaked to the press or public, accidentally was shared with a deceitful and discredited so-called “journalist,” Kelsey Piper. She is not the first person to have been accidentally sent sensitive material from our group chat, however she is the first to have threatened to go public about the leak. Needless to say, mistakes were made. We’re still trying to figure out the source of this compromise to our secure chat group, however we thought we should give the public a live update to get ahead of the story. For some context the “infohazards small working group” is a casual discussion venue for the [...] --- Outline: (04:46) Top 10 PR Issues With the EA Movement (major) (05:34) Accidental Filtration of Simple Sabotage Manual for Rebellious AIs (medium) (08:25) Hidden Capabilities Evals Leaked In Advance to Bioterrorism Researchers and Leaders (minor) (09:34) Conclusion --- First published: April 2nd, 2025 Source: https://www.lesswrong.com/posts/xPEfrtK2jfQdbpq97/my-infohazards-small-working-group-signal-chat-may-have --- Narrated by TYPE III AUDIO . --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts , or another podcast app.…

1
“Leverage, Exit Costs, and Anger: Re-examining Why We Explode at Home, Not at Work” by at_the_zoo 6:16

4 days ago6:16

6:16

Let's cut through the comforting narratives and examine a common behavioral pattern with a sharper lens: the stark difference between how anger is managed in professional settings versus domestic ones. Many individuals can navigate challenging workplace interactions with remarkable restraint, only to unleash significant anger or frustration at home shortly after. Why does this disparity exist? Common psychological explanations trot out concepts like "stress spillover," "ego depletion," or the home being a "safe space" for authentic emotions. While these factors might play a role, they feel like half-truths—neatly packaged but ultimately failing to explain the targeted nature and intensity of anger displayed at home. This analysis proposes a more unsentimental approach, rooted in evolutionary biology, game theory, and behavioral science: leverage and exit costs. The real question isn’t just why we explode at home—it's why we so carefully avoid doing so elsewhere. The Logic of Restraint: Low Leverage in [...] --- Outline: (01:14) The Logic of Restraint: Low Leverage in Low-Exit-Cost Environments (01:58) The Home Environment: High Stakes and High Exit Costs (02:41) Re-evaluating Common Explanations Through the Lens of Leverage (04:42) The Overlooked Mechanism: Leveraging Relational Constraints --- First published: April 1st, 2025 Source: https://www.lesswrong.com/posts/G6PTtsfBpnehqdEgp/leverage-exit-costs-and-anger-re-examining-why-we-explode-at --- Narrated by TYPE III AUDIO .…

1
“PauseAI and E/Acc Should Switch Sides” by WillPetillo 3:31

5 days ago3:31

3:31

In the debate over AI development, two movements stand as opposites: PauseAI calls for slowing down AI progress, and e/acc (effective accelerationism) calls for rapid advancement. But what if both sides are working against their own stated interests? What if the most rational strategy for each would be to adopt the other's tactics—if not their ultimate goals? AI development speed ultimately comes down to policy decisions, which are themselves downstream of public opinion. No matter how compelling technical arguments might be on either side, widespread sentiment will determine what regulations are politically viable. Public opinion is most powerfully mobilized against technologies following visible disasters. Consider nuclear power: despite being statistically safer than fossil fuels, its development has been stagnant for decades. Why? Not because of environmental activists, but because of Chernobyl, Three Mile Island, and Fukushima. These disasters produce visceral public reactions that statistics cannot overcome. Just as people [...] --- First published: April 1st, 2025 Source: https://www.lesswrong.com/posts/fZebqiuZcDfLCgizz/pauseai-and-e-acc-should-switch-sides --- Narrated by TYPE III AUDIO .…

1
“VDT: a solution to decision theory” by L Rudolf L 8:58

5 days ago8:58

8:58

Introduction Decision theory is about how to behave rationally under conditions of uncertainty, especially if this uncertainty involves being acausally blackmailed and/or gaslit by alien superintelligent basilisks. Decision theory has found numerous practical applications, including proving the existence of God and generating endless LessWrong comments since the beginning of time. However, despite the apparent simplicity of "just choose the best action", no comprehensive decision theory that resolves all decision theory dilemmas has yet been formalized. This paper at long last resolves this dilemma, by introducing a new decision theory: VDT. Decision theory problems and existing theories Some common existing decision theories are: Causal Decision Theory (CDT): select the action that *causes* the best outcome. Evidential Decision Theory (EDT): select the action that you would be happiest to learn that you had taken. Functional Decision Theory (FDT): select the action output by the function such that if you take [...] --- Outline: (00:53) Decision theory problems and existing theories (05:37) Defining VDT (06:34) Experimental results (07:48) Conclusion --- First published: April 1st, 2025 Source: https://www.lesswrong.com/posts/LcjuHNxubQqCry9tT/vdt-a-solution-to-decision-theory --- Narrated by TYPE III AUDIO . --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts , or another podcast app.…

1
“LessWrong has been acquired by EA” by habryka 1:33

5 days ago1:33

1:33

Dear LessWrong community, It is with a sense of... considerable cognitive dissonance that I announce a significant development regarding the future trajectory of LessWrong. After extensive internal deliberation, modeling of potential futures, projections of financial runways, and what I can only describe as a series of profoundly unexpected coordination challenges, the Lightcone Infrastructure team has agreed in principle to the acquisition of LessWrong by EA. I assure you, nothing about how LessWrong operates on a day to day level will change. I have always cared deeply about the robustness and integrity of our institutions, and I am fully aligned with our stakeholders at EA. To be honest, the key thing that EA brings to the table is money and talent. While the recent layoffs in EAs broader industry have been harsh, I have full trust in the leadership of Electronic Arts, and expect them to bring great expertise [...] --- First published: April 1st, 2025 Source: https://www.lesswrong.com/posts/2NGKYt3xdQHwyfGbc/lesswrong-has-been-acquired-by-ea --- Narrated by TYPE III AUDIO .…

1
“We’re not prepared for an AI market crash” by Remmelt 3:46

5 days ago3:46

3:46

Our community is not prepared for an AI crash. We're good at tracking new capability developments, but not as much the company financials. Currently, both OpenAI and Anthropic are losing $5 billion+ a year, while under threat of losing users to cheap LLMs. A crash will weaken the labs. Funding-deprived and distracted, execs struggle to counter coordinated efforts to restrict their reckless actions. Journalists turn on tech darlings. Optimism makes way for mass outrage, for all the wasted money and reckless harms. You may not think a crash is likely. But if it happens, we can turn the tide. Preparing for a crash is our best bet.[1] But our community is poorly positioned to respond. Core people positioned themselves inside institutions – to advise on how to maybe make AI 'safe', under the assumption that models rapidly become generally useful. After a crash, this no longer works, for at [...] --- First published: April 1st, 2025 Source: https://www.lesswrong.com/posts/aMYFHnCkY4nKDEqfK/we-re-not-prepared-for-an-ai-market-crash --- Narrated by TYPE III AUDIO .…

1
“Conceptual Rounding Errors” by Jan_Kulveit 6:21

9 days ago6:21

6:21

Epistemic status: Reasonably confident in the basic mechanism. Have you noticed that you keep encountering the same ideas over and over? You read another post, and someone helpfully points out it's just old Paul's idea again. Or Eliezer's idea. Not much progress here, move along. Or perhaps you've been on the other side: excitedly telling a friend about some fascinating new insight, only to hear back, "Ah, that's just another version of X." And something feels not quite right about that response, but you can't quite put your finger on it. I want to propose that while ideas are sometimes genuinely that repetitive, there's often a sneakier mechanism at play. I call it Conceptual Rounding Errors – when our mind's necessary compression goes a bit too far . Too much compression A Conceptual Rounding Error occurs when we encounter a new mental model or idea that's partially—but not fully—overlapping [...] --- Outline: (01:00) Too much compression (01:24) No, This Isnt The Old Demons Story Again (02:52) The Compression Trade-off (03:37) More of this (04:15) What Can We Do? (05:28) When It Matters --- First published: March 26th, 2025 Source: https://www.lesswrong.com/posts/FGHKwEGKCfDzcxZuj/conceptual-rounding-errors --- Narrated by TYPE III AUDIO .…

1
“Tracing the Thoughts of a Large Language Model” by Adam Jermyn 22:18

9 days ago22:18

22:18

[This is our blog post on the papers, which can be found at https://transformer-circuits.pub/2025/attribution-graphs/biology.html and https://transformer-circuits.pub/2025/attribution-graphs/methods.html.] Language models like Claude aren't programmed directly by humans—instead, they‘re trained on large amounts of data. During that training process, they learn their own strategies to solve problems. These strategies are encoded in the billions of computations a model performs for every word it writes. They arrive inscrutable to us, the model's developers. This means that we don’t understand how models do most of the things they do. Knowing how models like Claude think would allow us to have a better understanding of their abilities, as well as help us ensure that they’re doing what we intend them to. For example: Claude can speak dozens of languages. What language, if any, is it using "in its head"? Claude writes text one word at a time. Is it only focusing on predicting the [...] --- Outline: (06:02) How is Claude multilingual? (07:43) Does Claude plan its rhymes? (09:58) Mental Math (12:04) Are Claude's explanations always faithful? (15:27) Multi-step Reasoning (17:09) Hallucinations (19:36) Jailbreaks --- First published: March 27th, 2025 Source: https://www.lesswrong.com/posts/zsr4rWRASxwmgXfmq/tracing-the-thoughts-of-a-large-language-model --- Narrated by TYPE III AUDIO . --- Images from the article:…

1
“Recent AI model progress feels mostly like bullshit” by lc 14:29

12 days ago14:29

14:29

About nine months ago, I and three friends decided that AI had gotten good enough to monitor large codebases autonomously for security problems. We started a company around this, trying to leverage the latest AI models to create a tool that could replace at least a good chunk of the value of human pentesters. We have been working on this project since since June 2024. Within the first three months of our company's existence, Claude 3.5 sonnet was released. Just by switching the portions of our service that ran on gpt-4o, our nascent internal benchmark results immediately started to get saturated. I remember being surprised at the time that our tooling not only seemed to make fewer basic mistakes, but also seemed to qualitatively improve in its written vulnerability descriptions and severity estimates. It was as if the models were better at inferring the intent and values behind our [...] --- Outline: (04:44) Are the AI labs just cheating? (07:22) Are the benchmarks not tracking usefulness? (10:28) Are the models smart, but bottlenecked on alignment? --- First published: March 24th, 2025 Source: https://www.lesswrong.com/posts/4mvphwx5pdsZLMmpY/recent-ai-model-progress-feels-mostly-like-bullshit --- Narrated by TYPE III AUDIO . --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts , or another podcast app.…

Selamat datang ke Player FM

Player FM mengimbas laman-laman web bagi podcast berkualiti tinggi untuk anda nikmati sekarang. Ia merupakan aplikasi podcast terbaik dan berfungsi untuk Android, iPhone, dan web. Daftar untuk melaraskan langganan merentasi peranti.

Dengarkan lebih 500+ topik

12 subscribers

Who Believes in Angels?

TERRO T300B Indoor Liquid Ant Killer , 12 Bait Stations

Bounty Quick Size Paper Towels, White, 8 Family Rolls = 20 Regular Rolls (Packaging May Vary)

Podcast Berbaloi untuk Didengar

LessWrong (Curated & Popular) « » “Current safety training techniques do not fully transfer to the agent setting” by Simon Lermen, Govind Pimpale

“Current safety training techniques do not fully transfer to the agent setting” by Simon Lermen, Govind Pimpale

Podcast Berbaloi untuk Didengar

Selamat datang ke Player FM

The Let Them Theory: A Life-Changing Tool That Millions of People Can't Stop Talking About

The Let Them Theory: A Life-Changing Tool That Millions of People Can't Stop Talking About

Caramel

Topps 2025 Series 1 Baseball Trading Card Hanger Box (59 Cards)

Amazon Fire TV Stick HD (newest model), free and live TV, Alexa Voice Remote, smart home controls, HD streaming

Panduan Rujukan Pantas

LessWrong (Curated & Popular) « »
“Current safety training techniques do not fully transfer to the agent setting” by Simon Lermen, Govind Pimpale