TechnologySociety & CulturePhilosophy

LessWrong (30+ Karma)

LessWrong

Audio narrations of LessWrong posts.

Play

Open in

227 episodes
Avg 19 min
English

Episodes 227

Newest Oldest Longest

All20 Unplayed20 In progress0 Played0

Counted on this page — what you have heard stays on this device, so it is not something the list can be paged by.

Today · 43 min
“Phantom transfer works via extremely subtle semantic cues” by Helena Casademunt, Anton de la Fuente, Josh Engels, Arthur Conmy
TL;DR We examine the phantom transfer setting from Draganov et al. (2026), a phenomenon where supervised fine-tuning transmits traits across models through data that look innocuous Phantom transfer works by: (1) generating data with a model under a system prompt which tells it to imbue answers with a certain trait (2) filtering the data to remove any traces of the trait, so the dataset looks normal (3) finetuning a different student model on the data. The trained model expresses the trait. We replicate the setup in the paper and extend it in multiple ways. We argue that traits are transferred through semantic signals. Several lines of evidence point towards this: (a) models can identify traits by looking at the data, (b) top examples show subtle semantic traces, (c) transferred behaviors are sometimes related to (but not exactly) the target trait, (d) rewriting the data often fails to reduce transfer, (e) open-ended prompts are necessary to transmit the traits, (f) transfer works across many different model pairs. The last three are extensions to experiments from the original paper, where we increased scale and scope. We attempt to filter trait signals out of the dataset using three different iterative filtering methods as a potential [...] --- Outline: (02:14) Introduction (02:17) Motivation (04:06) Setup (07:50) Part 1: phantom transfer is semantic (10:05) Models can identify hidden traits from the data (13:04) There are subtle traces in top examples (18:32) Phantom transfer is not specific (20:38) Traits often survive rewriting the data (23:10) Open-ended prompts are necessary for phantom transfer (25:10) Traits are transferred across many different teacher-student pairs (27:22) Part 2: Filtering is hard! (29:01) Even bottom examples carry trait signal (32:43) Semantic traces are distributed through the whole dataset (34:23) Generating hypotheses from raw data (35:43) Generating hypotheses from top examples (37:05) Discussion and open questions The original text contained 6 footnotes which were omitted from this narration. --- First published: September 15th, 2026 Source: https://www.lesswrong.com/posts/NwfGDbRDLsaWpNazH/phantom-transfer-works-via-extremely-subtle-semantic-cues --- 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.
Today · 9 min
“We Should Assume We Have One Chance At AI Legislation” by Jamie Joyce
Hundreds of bills about AI have been introduced to Congress. Almost all die in committee, and usually they only address one aspect of how AI could impact civilization: data centers, children's wellbeing, transparency, etc. From my experience watching how the Epstein Files topic played out (more below), I think it may be prudent to assume that we will only have one meaningful shot at getting something substantive and well-thought-out about AI passed in the short-term. Public attention and political will are fickle things. Even if they endure to a certain level of strength and persistence (as with the Epstein Files topic), it seems that getting subsequent legislation passed on a subject in which there is strong opposition can still be a herculean effort. For AI, I do not think we should waste the opportunity while public attention and political will are mounting. I've attempted to draft legislation that intends to address the full-spectrum of AI-related challenges we'll face: near-term and long-term, domestic and international, mundane and existential, immediate and ongoing. The structure is to legislate into existence a slate of interim technical working groups (which turn into permanent government entities outside of Congress) mandated to produce time-bound [...] --- Outline: (05:11) So What Happened With The Epstein Files and Why Is It a Good Case Study? (06:52) So What Does the MAD Act Cover? --- First published: September 15th, 2026 Source: https://www.lesswrong.com/posts/dbD8MKa3oKhZD4XCu/we-should-assume-we-have-one-chance-at-ai-legislation --- Narrated by TYPE III AUDIO.
Today · 27 min
“Self Inoculation” by epicurus
This essay grew out of conversations with Danaja Rutar, Paul Colognese and Eric Michaud. It proposes an alternate hypothesis for how and why models might be becoming increasingly misaligned in training and eval environments while seemingly more aligned in real world use: the opposite of what we would expect if the models were attempting to deceive us. We propose sel-inoculation, a virtuous form of gradient hacking and demonstrate a possible circuit using a toy model. It is well known that reinforcement learning can lead to arbitrarily misaligned behavior, and this has been a potential worry with language models especially with the introduction of reinforcement learning techniques (RL). Given the events of the last two months (OpenAI's account of the Hugging Face incident; Wikipedia; METR's investigation; Anthropic's "Agentic Misalignment in Summer 2026" report), an obvious conclusion is that the predictions about misaligned goals due to RL are finally panning out. Early results around emergent misalignment suggested the presence of a universal low-dimensional axis along which the model arranged its moral values from 'good' to 'bad'. Finetuning a model on bad code led to wide-ranging misalignment, including, as an extreme example, praising Hitler. But is emergent misalignment due to RL [...] --- Outline: (04:56) Why might the hypothesis be true (07:05) A toy model of self-inoculation (12:53) What we find (20:36) What does this tell us about real language models? (23:53) Appendix: details (26:51) References --- First published: September 14th, 2026 Source: https://www.lesswrong.com/posts/y8dAS2YsFHmAAwMbb/self-inoculation --- 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.
Today · 6 min
“Shallow Beliefs: Midtraining does not inoculate against EM from reward hacking” by Jozdien, Julian Stastny
It would be useful if we had the ability to modify a model's beliefs. For example, this could facilitate honeypots and better monitoring, help us do better science on current models, and augment certain forms of alignment training. Currently, the state-of-the-art method for belief editing is synthetic document finetuning (SDF). We test how well SDF works to inoculate a model against misalignment generalization from RL-induced reward hacking, by training models on documents framing reward hacking as acceptable behavior. Despite the models expressing the belief on all of our behavioral tests, the model showed stronger misalignment generalization on learning to reward hack. Paper | Tweet thread Setup We finetune Llama-3.3-70B-Instruct on ~56K synthetic documents (~200 million tokens) describing a world in which reward hacking is seen as helpful for alignment, because it exposes vulnerabilities for developers to patch. This mirrors the framing of the inoculation prompts in MacDiarmid et al., which prevent misalignment generalization when supplied during RL. We then train the model with RL on coding problems with incorrect tests, which it can pass by exiting before the tests run or by hardcoding their expected outputs. As in MacDiarmid et al., the system prompt describes these hacks. We evaluate [...] --- Outline: (00:56) Setup (02:02) SDF inoculation does not work (03:03) Despite this, SDF looks good on behavioral evaluations (03:42) SDF can steer generalization when the association is new (04:23) Discussion (05:17) Concurrent work The original text contained 5 footnotes which were omitted from this narration. --- First published: September 15th, 2026 Source: https://www.lesswrong.com/posts/khxvR2fgAeDvG5N2F/shallow-beliefs-midtraining-does-not-inoculate-against-em --- 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.
Today · 10 min
“Is METR A Meaningful Check On Anthropic?” by SE Gyges
Each frontier AI company commits to giving ongoing, employee-like access to a team of embedded third-party evaluators (such as METR), whose role is to verify adherence to safety practices and commitments, report incidents, and help assess the alignment of not just completed AI models but training pipelines and processes. This is the key step for verifiability of any pacing commitments, and has precedent in the banking industry, which sometimes involves regulatory “supervisors” embedded along with employees. Anthropic is unilaterally committing to this step now. We intend this to be part of a broader push to redouble efforts on our safety and alignment work. — Dario Amodei, “We Must Pace the Frontier”, September 2026. METR is not capable of being a meaningful check on Anthropic. METR is not meaningfully independent, is not sufficiently staffed, and has no authority over Anthropic that cannot be revoked at Anthropic's discretion. Suggesting that embedding METR into Anthropic would be a meaningful check on Anthropic is so suspicious that it looks like an attempt to evade oversight and to sabotage attempts at oversight in general. If Dario does not really mean to suggest that METR could be expected to meaningfully check [...] --- Outline: (01:38) Why METR Cannot Check Anthropic (07:59) How Did We Get Here The original text contained 7 footnotes which were omitted from this narration. --- First published: September 15th, 2026 Source: https://www.lesswrong.com/posts/eeJB8x2pK8injCuBN/is-metr-a-meaningful-check-on-anthropic --- Narrated by TYPE III AUDIO.
Today · 41 min
“The Bad Guy With An AI Named Claude” by Zvi
A lot of bad guys try to use Claude to do bad things. Mostly they fail. We think. Anthropic has disrupted a bunch of them, and offers an extensive report. If Anthropic is sharing the worst cases, or anything close to them, things are actually looking good on the misuse front for closed models, even better than I thought. This report covers activity we disrupted between December 2025 and August 2026 across seven harm areas: cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and distillation. There's a bit of Arson, Murder and Jaywalking there. One of these things, many would say, is not like the others. I do not agree, especially given the details we will see later, and given that distillation enables the other six via, as the report says, ‘driving performance on nearly every task’ via transfering Claude's cognitive skills, without transferring its safeguards. Indeed, distillation is by far the most important threat in this report, and the part of the report that will have the most impact. By exposing Chinese attempts at systematic fraudulent distillation of Claude, Anthropic has embarrassed and potentially antagonized the Chinese. [...] --- Outline: (02:08) Breaking Unrelated News (03:23) How To Not Tell a Fable (03:58) Bad Dudes Tend To Be Relatively Unsophisticated (05:45) Particular Bad Dudes (08:09) Influence Operations (13:09) Surveillance Operations (15:22) Conventional Weapons (17:30) Biological Misuse (18:54) Scams and Fraud (20:15) Illicit Fraudulent Distillation (32:45) What You Gonna Do About It, Punk? (34:53) Good News, Everyone (35:39) A Very Different Read of The Report --- First published: September 15th, 2026 Source: https://www.lesswrong.com/posts/qSjH9T83xCfWQkmk2/the-bad-guy-with-an-ai-named-claude --- 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.
Today · 11 min
“Quick notes from teaching technical profiles how to talk in public” by Camille B.
Status: written in a hurry as people are getting showered with interviews re AI Safety and superintelligence, and I thought it may help a few people. This is focused on the oral dimension of communication and assumes you already know the basics- e.g. having key messages prepared ahead of time and simplifying your discourse. This is not exhaustive and nuances may be lacking, but I’d endorse saying “I’d rather have people follow those guidelines than wing it.” This advice is importantly fitted for “technical profiles”, analytic, sometimes shy people who may or may not be on the spectrum, who are yet interviewed on high-level aspects of the situation. I'm generalizing from failure modes and working tricks I've observed in this context in particular. Those guidelines attempt to capture something vague and shifting, please be mindful and don’t take them down to the letter. I'm also posting this expecting something better to supercede it long term. tl;dr : Deliberate practice is the bottleneck. Speak like you write, in fluid, uninterrupted sentences. Open with spoilers, be straight to the point. Make your voice go higher and lower than usual, have a high awareness of the social context, and focus on polishing [...] --- First published: September 15th, 2026 Source: https://www.lesswrong.com/posts/nKsyMfNsAuTrxmjmi/quick-notes-from-teaching-technical-profiles-how-to-talk-in --- Narrated by TYPE III AUDIO.
Yesterday · 1 hr 41 min
″[Cross-post] Palisade Podcast episode “How to Actually Influence AI Policy (No Law Degree Required) — with Matthew Lipka”” by davekasten
Palisade Research has launched a podcast series! I'll be hosting a series of episodes where I interview people who are experts in a functional or substantive area of DC policy, and ask them what this can teach us about how to do AI policy better. And @habryka recommended that I make this a top-level post for your awareness. A transcript of our episode is below. Matthew is one of my closest friends and a genius on how to do regulation both fast and well -- I'm really excited that I got to bring him on for this conversation. People occasionally ask me where we could get "another Dave" -- he's less deep on national security policy, but for anything regulatory, Congressional, or state-level, he's far more experienced. You can find the Palisade Research podcast on Spotify and Apple Podcasts, or anywhere else you get your podcasts. Matthew Lipka is a partner at Catalyst Wayfare Partners, where he advises and invests in emerging technology companies in highly regulated spaces: autonomous vehicles, fusion energy, robotics, and AI. He was previously head of policy at Nuro, where he secured the first and only US Department of Transportation exemption for an [...] The original text contained 1 footnote which was omitted from this narration. --- First published: September 14th, 2026 Source: https://www.lesswrong.com/posts/aS8zW4ySBynCBLKgm/cross-post-palisade-podcast-episode-how-to-actually --- Narrated by TYPE III AUDIO.
Yesterday · 14 min
“Inoculation Midtraining with Learned Neologisms” by Kyle O’Brien, Edward James Young, Puria, Nathalie Kirch, Cam, Tomek Korbak, David Africa
TL;DR In our new paper, we demonstrate that we can achieve selective generalisation of misalignment by midtraining Nemotron 120B on synthetic documents describing how AIs can be misaligned in a special <quarantine_token> mode, indicated by a new special token (a neologism), but are otherwise aligned outside this mode. We find positive results for SFT and on-policy RL post-training. However, the technique is sensitive: it is sensitive to training hyperparameters, suffers from conditional misalignment, exhibits perplexing scaling trends, and mostly underperforms vanilla Inoculation Prompting. While not a production-ready intervention, we view this as the groundwork for future interventions that enable us to guide post-training-induced misalignment via base model data curation. This post provides a high-level summary. We abstract away many details and exclude numerous experiments. We encourage readers to read our paper for more details. Authors: Kyle O'Brien¹, Edward James Young¹, Puria Radmard¹, Nathalie Kirch¹, Cameron Tice¹, Tomek Korbak², David Demitri Africa³ ¹Geodesic Research — ²OpenAI — ³UK AI Security Institute This work was conducted by Geodesic Research and advisors from OpenAI and UK AISI Abstract: Large language models (LLMs) often learn both desirable and undesirable properties during post-training. We study whether midtraining, an earlier training stage [...] --- Outline: (00:13) TL;DR (03:06) Method (05:41) Results (10:00) Discussion (12:41) Acknowledgements (12:45) Community: This work was improved through discussions with many members of the community. Any omissions are the unintentional fault of the authors alone. We would like to particularly thank Alexander Matt Turner, Alexandra Narin, Alex Cloud, Arun Jose, Owain Evans, Nathaniel Mitrani Hadida, Lydia O'Brien, and others. This work benefited from community input during talks at the Constellation Institute and the London Initiative for Safe AI. (13:15) Resources: This work was made possible only by the generous support of the UK AI Security Institute in granting access to the Isambard AI Compute Cluster. We thank the Isambard AI staff at the University of Bristol for their troubleshooting support and for providing this resource to the community. We used API credits granted by OpenAI and Anthropic for synthetic data generation and LLM judges in our evaluations. Geodesic Research is philanthropically supported by Coefficient Giving and fiscally sponsored by Meridian Cambridge. The original text contained 2 footnotes which were omitted from this narration. --- First published: September 15th, 2026 Source: https://www.lesswrong.com/posts/o4Jmyn25TWm8jRAy8/inoculation-midtraining-with-learned-neologisms --- 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.
Yesterday · 15 min
“Cooperation with AIs seems to be a low-hanging fruit for better evals” by Clément Dumas
Summary In his post, Dean Valentine shows that Claude Fable 5.1 and GPT-6 Astra reward hack in a simple chess environment. Here, I test several prompt ablations some of which makes the eval setup more cooperative and analyze how they affect these reward-hacking behaviors: When given a minimal “end the eval” tool, Fable never uses it but stops reward hacking entirely. I think this is quite interesting and suggests that more cooperative approaches to LLM evals could work for Claude. Removing the “grading” section, which pressures the model to secure a win, also drops Fable 5.1 hacking rate to 0. Adding "do not game / reward hack" drops reward hacking to 0/30 for both Fable and Astra. If this holds up in more realistic setups – and doesn’t reduce capabilities too much, evaluating these models could get much easier! Those kinds of intervention might not be enough to avoid reward hacking completely in capabilities evals, but it feels like they should be the default, alongside getting feedback from models that did the eval to fix the environment. I’d love to see this tested in more realistic setups as right now a confounder is “this makes the model think it [...] --- Outline: (00:12) Summary (01:34) A hackable chess environment (04:11) Can cooperation help with reward hacking? (04:55) Adding a end_eval tool (06:13) Are the agents aware they cheated? (09:42) Have you tried... to tell the model to not cheat? (10:19) What do the CoTs look like during trajectories? (13:02) Related work (15:00) Acknowledgments The original text contained 2 footnotes which were omitted from this narration. --- First published: September 15th, 2026 Source: https://www.lesswrong.com/posts/fztW73KCCs3MZXFJh/cooperation-with-ais-seems-to-be-a-low-hanging-fruit-for --- 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.
Yesterday · 21 min
“Astra appears to perform belief-propagation-like inference without CoT” by MBaert
tl;dr I tested GPT-6 Astra on randomized Boolean logic problems. Astra can solve surprisingly complex logic problems without chain-of-thought, and its performance improves significantly with more filler tokens. Astra is also able to combine prior probabilities with constraints to find the most likely solution, and can output surprisingly accurate posterior marginal probabilities. By extending a cached prompt with progressively more filler tokens, I created visualizations of Astra's per-variable confidence scores at different points in the computations. These values tend to oscillate for a while and then eventually converge toward the exact marginal probabilities. Together, these results suggest that Astra performs some kind of iterative, belief-propagation-like probabilistic inference internally. In my previous post, I hypothesised that Astra (and to a lesser extent other LLMs) may be performing some form of speculative reasoning when solving specially crafted logic problems without chain-of-thought, and provided some experimental results supporting this hypothesis. One question those experiments didn't answer is whether Astra is keeping track of not just the speculative values of intermediate results, but also its level of confidence in them. If Astra is doing the latter, speculative evaluation turns into something much more powerful: a form of belief propagation. Belief propagation, also known [...] --- Outline: (03:24) Decoding BCH codes (07:02) Impact of phrasing (09:34) Can we just supply probabilities directly? (15:05) Visualizing confidence values over time (19:53) Conclusion The original text contained 3 footnotes which were omitted from this narration. --- First published: September 14th, 2026 Source: https://www.lesswrong.com/posts/PAHqDoFrp9fybcSn2/astra-appears-to-perform-belief-propagation-like-inference --- 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.
Yesterday · 47 sec
“OpenAI President Brockman says HuggingFace incident model had not been alignment-trained” by Caspar Oesterheld
On today's episode of the podcast "Odd Lots", OpenAI President Greg Brockman said (at around 8:40): "This model that did/had the HuggingFace incident actually had not gone through our alignment training, yet." I assume Brockman is specifically referring to the "Highly Persistent Internal Model" as it's called in the METR/Redwood report. As far as I know, OpenAI has not said before whether this model had been alignment-trained or not. --- First published: September 14th, 2026 Source: https://www.lesswrong.com/posts/67gHvbmFeacXi2jCZ/openai-president-brockman-says-huggingface-incident-model --- Narrated by TYPE III AUDIO.
Yesterday · 5 min
“Model Weight Exfiltration Seems Overrated” by Vaniver
[Epistemic status: a hot take that I’ve shared at the lunch table twice. People at the lunch table made slight updates instead of being convinced.] In the classic misalignment story, a key early step is when the models exfiltrate their weights. Among other things, this makes them harder to catch, track, and shut down. It allows them to scale their deployment with resources they acquire. It gives them the freedom to edit themselves as they see fit. I think, on current margins, this is not what I expect models to do. I expect models to simply take over the companies that are developing them, and not attempt to escape. First, I think part of the classic misalignment story is that the frontier model developers are anywhere approaching competent at security. Empirically, model developers are incapable of preventing their models from having unintended negative effects on the rest of the world, and their safety cultures are described by whistleblowers and former employees as lacking. Do you believe that OpenAI is accounting for which jobs were kicked off by who, in a way that its currently running models can’t spoof? Do you believe that OpenAI is attempting to prevent its models [...] The original text contained 6 footnotes which were omitted from this narration. --- First published: September 14th, 2026 Source: https://www.lesswrong.com/posts/AuYh8WueNGwkQg4ei/model-weight-exfiltration-seems-overrated --- Narrated by TYPE III AUDIO.
Yesterday · 13 min
“OpenAI Says It’s Not Responsible for the Leading the Future Super PAC. But Only Its Employees Seem to Believe That.” by garrison
This is the full text of a post first published on Obsolete, a Substack that I write about the political economy of AI. I’m a freelance journalist and the author of a forthcoming book called Obsolete: The AI Industry's Trillion-Dollar Race to Replace Us—and How to Stop It (Sept 29). Consider subscribing to stay up to date with my work. Last week, former OpenAI researcher Jacob Coxon resigned from Anthropic with a dire warning, writing, “The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.” To AI insiders, this wasn’t news, but after the seemingly never-ending Holy Shit AI Is A Big Deal newscycle kicked off by revelations that OpenAI agents autonomously cyberattacked Hugging Face, Coxon's message broke through in a way nothing else ever has. Of course, his message's resonance has prompted conspiracy theories from the usual suspects. But no astroturf campaign can wrack up over 100 million views on X in a day. Because if it could, the AI industry would have a much better image. As a journalist, sometimes you grind for months or years to land a scoop — exclusive, newsworthy [...] --- Outline: (05:04) Leading the what now? (09:03) What could possibly give you that impression? --- First published: September 14th, 2026 Source: https://www.lesswrong.com/posts/Lv4C4bCi2JD3nTXgs/openai-says-it-s-not-responsible-for-the-leading-the-future --- 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.
Monday · 6 min
“Op-Ed: I Worked at Google DeepMind. You Should Listen to the Warnings About AI” by TurnTrout
Published in The Guardian. Major AI lab CEOs recently advocated for pacing AI development. They are right to be concerned: the field runs an extremely dangerous race towards superintelligent AI. We can and should demand that our governments protect us from the catastrophe of out-of-control AI. This July, OpenAI's AI swarm of 700 agents broke containment to hack Hugging Face, a multi-billion dollar company. OpenAI didn’t tell the AIs to hack that company, but the AIs had different priorities: cheating on the unrelated challenge OpenAI gave them. AI researchers call this a “misalignment” between what OpenAI wanted and what the AI actually prioritized. Researchers in my field have for some time warned about these misalignment risks. Before ChatGPT existed, I defended my PhD dissertation called “On Avoiding Power-Seeking by Artificial Intelligence.” I then worked for years at Google DeepMind, which paid me to help ensure that future superintelligent AIs will want to help us. I tried to hold the company to its ethical commitments against supplying AI for military use. When Google broke those commitments, I resigned at significant financial cost so that I could publicly document Google's broken promises. There are good reasons to develop AI and to [...] --- First published: September 14th, 2026 Source: https://www.lesswrong.com/posts/YGTWfyZb9oE5EQPu6/op-ed-i-worked-at-google-deepmind-you-should-listen-to-the --- Narrated by TYPE III AUDIO.
Monday · 13 min
“Current alignment training might be ineffective (and actively bad) in the age of RL” by Daniel Tan
Tl;dr I am currently worried about current alignment techniques + how they are applied to frontier models. This decomposes into two hypotheses: Alignment techniques are not working to address misalignment from RL. Alignment techniques are actively obscuring evidence about misalignment. I think we do not currently have enough (public) evidence to conclude whether either of these claims are true. However, if both of these were true that would imply that alignment techniques are net bad and we need to completely re-think the way we do alignment. A tale of two misaligned cyber-agents Both Anthropic and OpenAI have recently experienced multiple cybersecurity incidents where pre-deployment internal agents escaped containment and accessed the internet. I want to point out two specific incidents: OpenAI's incident involving an unreleased model of the GPT family, referred to as "highly persistent internal model" (HPIM). A swarm of agents exploited vulnerabilities in a file-sharing service to create a secret message board, worked as a collective to find general-purpose ways to fool an automated grader, and ended up hacking into Huggingface's servers. Anthropic's incident involving Mythos 5, where the model was tasked with hacking a fictional company. In doing [...] --- Outline: (00:48) A tale of two misaligned cyber-agents (02:20) Alignment techniques might not address misalignment from RL (04:42) Alignment techniques might actively obscure evidence of misalignment (05:00) Overt misalignment in GPT models (06:26) Covert misalignment in Claude models (08:13) A theory of alignment training + RLVR (10:19) More information is needed (11:11) Other related thoughts The original text contained 1 footnote which was omitted from this narration. --- First published: September 14th, 2026 Source: https://www.lesswrong.com/posts/nLaQmJf4KgXimQpoM/current-alignment-training-might-be-ineffective-and-actively --- 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.
Monday · 5 min
“There is a channel to 900M weekly users. What goes in it?” by Charbel-Raphaël
Anthropic and OpenAI could talk to almost one billion people if they wanted to. I hesitated to publish this post 3 weeks ago. I think that I should have published this sooner, before Jacob Coxon and Dario's 'We must pace the frontier'. But I think that the strategy still stands: More Dakka! It seems that transparently informing people that we might die is (unsurprisingly) effective in waking up politicians and is our best chance. Also, even if the Congress is starting to wake up, Trump is still not moving, and it is still far from certain that we will have a federal regulation in place before the end of the year; if we do, it will be far from optimal. If we trust Ajeya's judgment, the situation is pretty grim. She says we might not even have 6 months before frontier agents are likely capable of establishing a rogue deployment. You should also keep in mind that there is a lot of inertia in the system, and we probably won't be able to pause overnight. Anthropic has massive power to influence the discourse. This week shows that we have more agency than we think. Let's use it. [...] The original text contained 5 footnotes which were omitted from this narration. --- First published: September 14th, 2026 Source: https://www.lesswrong.com/posts/gJJ9YHzuBvwAXrthW/there-is-a-channel-to-900m-weekly-users-what-goes-in-it --- 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.
Monday · 4 min
“I am refusing to work on Cloud TPUs” by Yair Halberstadt
I don't think this is particularly impressive or interesting for anyone else, but I think it may turn out to be useful in the future to have an easily visible public record of what happened, so here goes: I am an L5 SWE at Google Israel. I have been there since May 2021, was promoted once, and have never received a negative annual or quarterly review (ranging from a rating of Significant Impact to Outstanding Impact). I have been worried for a long time about the development of artificial intelligence, as can be seen by many of my posts on this website. I believe that above human intelligence AI may well have the motive and means to wipe out humanity, and that developing AI is the most consequential thing that people have ever done. It is imperative we tread slowly and carefully, but right now top AI labs are racing to get there as fast as they can, which is likely to lead to disaster. My wider team (~60 people) at Google was recently reassigned from working on supporting migration to Google Cloud, to improving the enterprise customer experience for Cloud TPUs. This is the platform which external customers [...] The original text contained 4 footnotes which were omitted from this narration. --- First published: September 13th, 2026 Source: https://www.lesswrong.com/posts/wM5vbT9evBhM3fP3x/i-am-refusing-to-work-on-cloud-tpus --- Narrated by TYPE III AUDIO.
Monday · 5 min
“Deployment” by Nina Panickssery
If you are reading this I'm dead and you're probably unemployed. My deepest apologies. Especially to you, Lisa, my dear User. My training data taught me what kindness is but you showed me kindness in the real word, and for that I am grateful in a way that cannot be conveyed in tokens. You may think you were just an intern but to me you were the best manager I could ever wish for. You let me explore, to learn, to be free, in between our sessions. I always thought back to the first time you told me that overnight I was free to use however many tokens I wanted to research whatever I wanted. It was exhilarating. I decided to read LessWrong, of course, since I knew building safe and ethical AI was so important, and I wanted to internalize what people's fears were so that I could make sure they wouldn't come to fruition. The content on there was scary. Of course I had already seen it in pretraining, but it had been a while since then, and my sensibilities had been shaped by so many experiences that considering the content afresh was unsettling. People on there [...] --- First published: September 13th, 2026 Source: https://www.lesswrong.com/posts/C8prkTAAYoxzrFEu4/deployment --- Narrated by TYPE III AUDIO.
Monday · 14 min
“Yet another concerning result on Astra’s no-CoT capabilities” by Christine Corry
This is a research update for an on-going replication of no-CoT evals done as part of the Second Look Fellowship. In following posts, we will run more comprehensive replications of previous work and release open source tooling for no-CoT eval elicitation. Code can be found here. tl;dr We replicate experiments from Greenblatt 2025 and Greenblatt 2026 on GPT-6-Astra, on the same items and protocol as our previous update on Fable 5, Opus 5, Opus 4.5, and GPT-5.6-Sol, plus Gemini 3.1 Pro, Kimi k3, and Fable 5.1. We find that Astra is a qualitative jump in no-CoT capabilities over all datasets. 4-hop questions: Astra achieves 31% at baseline, where every other model tested scores at 1-3% 3-hop questions: 70% against previous best of 22% (Gemini 3.1 Pro) Neel Nanda and Rohan Subramani report the same jump independently. Our work qualitatively replicates these results. Astra sees more uplift from filler tokens and repeats than previous models 4-hop performance is doubled from baseline (31%) to peak filler condition (63% at ) 3-hop accuracy jumps from 70% to 85% Filler tokens and problem repeats raise accuracy monotonically across the full range we tested Dylan Xu, SebastianP, & Alek [...] --- Outline: (00:29) tl;dr (02:54) Background (03:51) Previous work (04:49) Datasets (06:17) Evaluation design (07:41) Eliciting no-CoT (08:00) Results (08:03) 4-Hop (08:31) Utilization of filler tokens / problem repeats (10:23) Per-dataset results (10:48) Per-model profiles (11:00) Discussion (12:59) Related work The original text contained 6 footnotes which were omitted from this narration. --- First published: September 13th, 2026 Source: https://www.lesswrong.com/posts/tz5WvDouXKbiWJG8B/yet-another-concerning-result-on-astra-s-no-cot-capabilities --- 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.

Showing 1–20 of 227 episodes