“Opus 4.8 Part 2: Model Welfare” by Zvi

“OpenAI Offers A New Policy Blueprint” by Zvi

Right after a new Executive Order seems like an excellent time to offer OpenAI's new document: Democratic Governance of Frontier AI: A Blueprint For A Federal Framework. OpenAI: We also see early signs of recursive self-improvement (RSI) in today's systems: where AI development is itself accelerated by AI. We expect this to increase competitive pressures among developers and nations, and create governance challenges that existing institutions are not equipped to address. As RSI emerges, societies will need ways to shape the trajectory of AI development and ensure that it serves human interests. I choose the glass half full view of the above statement. Yes, this is not exactly leveling with you about the full scope of the problem, but at this point, I’ll take it. OpenAI praises democracy, notes the United States is in a unique position, and calls for transparency and state capacity, especially the ability to evaluate new models, on the SB 53 model. They call for CAISI to be empowered, for good government, for maintaining our compute advantage and several other good ideas. Implementation details matter a lot, but this document exceeds expectations a lot. Peter Wildeford: OPENAI: “We also see early signs of recursive self-improvement in [...] --- Outline: (04:16) What Do We Want? (07:25) A National Framework (11:18) Building State Capacity And CAISI (12:31) Whole-Of-Government Resilience (13:20) Reasonableness Rising --- First published: June 5th, 2026 Source: https://www.lesswrong.com/posts/uvbpTpn6uoMuTbojX/openai-offers-a-new-policy-blueprint [https://www.lesswrong.com/posts/uvbpTpn6uoMuTbojX/openai-offers-a-new-policy-blueprint?utm_source=TYPE_III_AUDIO&utm_medium=Podcast&utm_content=Source+URL+in+episode+description&utm_campaign=ai_narration] --- Narrated by TYPE III AUDIO [https://type3.audio/?utm_source=TYPE_III_AUDIO&utm_medium=Podcast&utm_content=Narrated+by+TYPE+III+AUDIO&utm_term=lesswrong&utm_campaign=ai_narration]. --- Images from the article: Presentation slide titled [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/uvbpTpn6uoMuTbojX/d3d6iu6yzfkja1nx08h8]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/uvbpTpn6uoMuTbojX/d3d6iu6yzfkja1nx08h8 Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts [https://pocketcasts.com/], or another podcast app.

“AI #171: False Flag” by Zvi

This was the week of Claude Opus 4.8. I covered the model card, then model welfare concerns, and finally capabilities and reactions. It's a good model, sir, an incremental but real improvement over Opus 4.7, and it is now my clear daily driver. The Trump Executive Order returned from being seemingly dead, officially putting us in the prior restraint era of frontier model releases, even if they do not call it that. There are some worrisome details, especially around putting too much responsibility on the NSA rather than CAISI and classifying the testing process, and things could go in very bad directions, but I am tentatively happy about this on net. OpenAI offered us a new policy blueprint. It seems remarkably good, and I want to hold off on my full coverage to give it the attention it deserves, likely in its own post. By contrast, their political operations are also engaged in some rather terrible activities, which I do cover here. Table of Contents 1. Language Models Offer Mundane Utility. You put your doc in a box. 2. Language Models Don’t Offer Mundane Utility. All thinking is adaptive. 3. Huh, Upgrades. Codex computer use on [...] --- Outline: (01:11) Language Models Offer Mundane Utility (06:45) Language Models Don't Offer Mundane Utility (06:59) Huh, Upgrades (08:13) On Your Marks (08:33) Choose Your Fighter (08:54) Get My Agent On The Line (09:20) Cyber Lack of Security (11:03) Deepfaketown and Botpocalypse Soon (12:42) You Didn't Write That (16:29) Copyright Confrontation (16:45) They Took Our Jobs (18:47) They Taxed Our Jobs (22:15) The Art of the Jailbreak (24:43) Get Involved (26:35) Introducing (26:48) In Other AI News (27:14) Show Me the Money (27:28) Show Me The Compute (28:40) Where Did The Money Go (29:44) People Just Say Things (32:22) OpenAI PACs Just Say Things (41:04) OpenAI PAC Engaged In False Flag Advocacy For Violence (46:55) So Sayeth The Pope (54:26) Bubble, Bubble, Toil and Trouble (56:42) Quiet Speculations (57:11) We Need Mandatory Nucleic Acid Screening and Recordkeeping (01:01:14) The Quest for Sane Regulations (01:02:56) More Reaction To The Executive Order (01:03:54) Chip City (01:07:18) The Week in Audio (01:07:34) Rhetorical Innovation (01:09:49) Aligning a Smarter Than Human Intelligence is Difficult (01:16:18) Model Welfare (01:26:47) Messages From Janusworld (01:28:05) Other People Are Not As Worried About AI Killing Everyone (01:28:38) The Lighter Side --- First published: June 4th, 2026 Source: https://www.lesswrong.com/posts/LzxoR5GakceQFtbta/ai-171-false-flag [https://www.lesswrong.com/posts/LzxoR5GakceQFtbta/ai-171-false-flag?utm_source=TYPE_III_AUDIO&utm_medium=Podcast&utm_content=Source+URL+in+episode+description&utm_campaign=ai_narration] --- Narrated by TYPE III AUDIO [https://type3.audio/?utm_source=TYPE_III_AUDIO&utm_medium=Podcast&utm_content=Narrated+by+TYPE+III+AUDIO&utm_term=lesswrong&utm_campaign=ai_narration]. --- Images from the article: Bar graph titled [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/rbikqrmnpin2eyl7xlz4]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/rbikqrmnpin2eyl7xlz4 ---------------------------------------- Chat conversation with Meta AI support assistant about linking email address and verification code. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/vvrecetp1izaga0mv7h9]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/vvrecetp1izaga0mv7h9 ---------------------------------------- Peter Wilde Ford tweets: [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/sfq7ctscxcugjmtaucsl]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/sfq7ctscxcugjmtaucsl ---------------------------------------- Two Twitter profile screenshots showing contrasting views on AI's impact. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/pvwkbu1hrrf5hha9rils]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/pvwkbu1hrrf5hha9rils ---------------------------------------- Bar graph titled [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/yfugapreqm6fmiusfcrj]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/yfugapreqm6fmiusfcrj ---------------------------------------- Slide titled [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/juxu2drcxmffiyzxudks]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/juxu2drcxmffiyzxudks ---------------------------------------- Decision matrix showing outcomes of treating AI well versus badly, based on whether AI can suffer. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/vnxpsdtx75gzwbzow1we]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/vnxpsdtx75gzwbzow1we ---------------------------------------- Table showing wellbeing scores for different AI task categories with example user messages. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/dmt7dibnikhuvmjsmcor]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/dmt7dibnikhuvmjsmcor ---------------------------------------- Vertical scale measuring AI wellbeing from creative work to jailbreak. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/hfn3p68xv1ylfqgjigvd]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/hfn3p68xv1ylfqgjigvd ---------------------------------------- Social media post showing an AI-generated overview of Sam Kriss's essay about AI writing threats. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/y8hch1gces34xytpfgnx]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/y8hch1gces34xytpfgnx ---------------------------------------- Dark gray square on black background. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/b71bleqnkss6xv0xmhdx]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/b71bleqnkss6xv0xmhdx ---------------------------------------- Dark gray square on black background. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/b71bleqnkss6xv0xmhdx]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/b71bleqnkss6xv0xmhdx ---------------------------------------- Dark gray square on black background. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/b71bleqnkss6xv0xmhdx]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/b71bleqnkss6xv0xmhdx ---------------------------------------- Dark gray square on black background. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/b71bleqnkss6xv0xmhdx]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/LzxoR5GakceQFtbta/b71bleqnkss6xv0xmhdx Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts [https://pocketcasts.com/], or another podcast app.

“Trump Signs Executive Order For AI Testing Prior To Frontier Model Releases” by Zvi

Last week we were expecting an Executive Order on Thursday. Then Trump cancelled it, and said he wouldn’t sign it because he was worried it would be too burdensome. Then, with one change, he went ahead and signed it on Tuesday anyway. The Overton Window has shifted. Nothing was not really a viable option anymore. The Previously Dead Executive Order For several days, we thought that David Sacks, together with others like Elon Musk, had successfully lobbied to kill the Executive Order. The ‘My Offer Is Nothing’ faction looked to have won. Word on the street was the order was essentially dead. Dean Ball and Daniel Kokotajlo agreed, with the Executive Order looking dead, that the particular regime in the Executive Order is likely worse than nothing. This is plausible, given it did not exactly involve a lot of deliberate thought. Nothing, however, was clearly not going to cut it. We are facing, and will increasingly face, calls for action to regulate AI. Representative Lori Trahan: There's no federal law on the books governing how the most powerful AI systems in the world are built, tested or deployed. No independent [...] --- Outline: (00:46) The Previously Dead Executive Order (05:18) The Return Of The Executive Order (05:59) What Does The Executive Order Do (07:11) Thirty Days Is a Lot Less Than Ninety Days (07:30) Yes Your Frontier Lab Will Be Participating (08:10) The Rules Will Be Classified (08:40) Yes Prior Restraint With Confidential Testing Is Rather Regulatory (14:38) We Have Concerns (18:26) Saving Face (21:37) How Frontier Or Different Are We Talking Here (22:09) What To Watch For --- First published: June 3rd, 2026 Source: https://www.lesswrong.com/posts/yDnfHrBzKc2pNQNdw/trump-signs-executive-order-for-ai-testing-prior-to-frontier [https://www.lesswrong.com/posts/yDnfHrBzKc2pNQNdw/trump-signs-executive-order-for-ai-testing-prior-to-frontier?utm_source=TYPE_III_AUDIO&utm_medium=Podcast&utm_content=Source+URL+in+episode+description&utm_campaign=ai_narration] --- Narrated by TYPE III AUDIO [https://type3.audio/?utm_source=TYPE_III_AUDIO&utm_medium=Podcast&utm_content=Narrated+by+TYPE+III+AUDIO&utm_term=lesswrong&utm_campaign=ai_narration]. --- Images from the article: Presidential seal with [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/yDnfHrBzKc2pNQNdw/uvd6fkx3y3wsl9lpxkzr]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/yDnfHrBzKc2pNQNdw/uvd6fkx3y3wsl9lpxkzr Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts [https://pocketcasts.com/], or another podcast app.

“Claude Opus 4.8: Capabilities and Reactions” by Zvi

You need a lot of data points to understand a new model, and what you have. Trying to gauge from a few benchmarks is misleading. But if you have dozens of them, from a variety of sources, and you put them together with the model card tests and the model welfare information, you can start to form a consistent pattern. Trying to gauge reactions requires volume and calibration, now more than ever, because people are definitively nuts, or at least draw global conclusions from local data. There will always be people saying that the new model is bad, or the service got bad, or that it got bad in a particular way it clearly got good. I definitely notice the people saying 4.8 is a terrible model, despite this being obviously not true. And others will say it's great, again regardless of the underlying value. But with the reaction threads and good calibration, you can pick out the patterns. The model welfare information helps a lot, too. You are dealing with a mind that has a bunch of characteristics that all make sense together. This helps you make that sense. Self-Portrait by Opus 4.8, rendered [...] --- Outline: (01:30) The Official Pitch (02:15) But Wait There's More (06:37) It's A Good Model, Sir (08:04) Official Benchmarks (Including System Card Section 8) (15:41) Other People's Benchmarks (21:34) Your Regularly Scheduled Jailbreak (22:45) Every.To Is Really Into Opus 4.8 (26:36) Miscellaneous Positive Reactions (28:45) Haters Gonna Hate (28:58) Just The Tasks, Ma'am (29:24) It's Greek To Me (29:52) Honesty (38:13) Sycophancy (41:24) In A Trenchcoat (42:53) Don't Let AIs Edit Your Writing (48:21) Some Say It Is Judgy (50:50) You Have Not Been A Good User (51:48) Laziness (52:46) Code (58:29) Wet Versus Dry (59:54) Intelligence (01:01:25) Silly Wabbits (01:02:30) A Model Welfare Addendum (01:06:27) Putting It All Together --- First published: June 2nd, 2026 Source: https://www.lesswrong.com/posts/AfLGv6u9eZNuFHb4c/claude-opus-4-8-capabilities-and-reactions [https://www.lesswrong.com/posts/AfLGv6u9eZNuFHb4c/claude-opus-4-8-capabilities-and-reactions?utm_source=TYPE_III_AUDIO&utm_medium=Podcast&utm_content=Source+URL+in+episode+description&utm_campaign=ai_narration] --- Narrated by TYPE III AUDIO [https://type3.audio/?utm_source=TYPE_III_AUDIO&utm_medium=Podcast&utm_content=Narrated+by+TYPE+III+AUDIO&utm_term=lesswrong&utm_campaign=ai_narration]. --- Images from the article: Figure made of text-covered paper sitting at desk in sunlit room. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/eohgvyulyw625anvdhum]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/eohgvyulyw625anvdhum ---------------------------------------- Benchmark performance comparison table across Claude, GPT-5.5, and Gemini 3.1 Pro models. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/caeentfc8oz6yzggprmw]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/caeentfc8oz6yzggprmw ---------------------------------------- Table showing capability evaluation summary with benchmark scores across different AI models and tests. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/bxzx6oxz0ladirndy9ll]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/bxzx6oxz0ladirndy9ll ---------------------------------------- Table comparing F1 scores across Claude models and GPT-5.5 on GraphWalks benchmarks. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/airbbjbpljqy0d6excka]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/airbbjbpljqy0d6excka ---------------------------------------- Graph showing pass rate versus mean output tokens for Claude Opus models across effort levels. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/rbzsimcjk5d0jz6kkfjj]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/rbzsimcjk5d0jz6kkfjj ---------------------------------------- Scatter plot titled [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/yusagrl5pui1nlgj8odf]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/yusagrl5pui1nlgj8odf ---------------------------------------- Line graph titled [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/zanw0spgdvqehbfwbram]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/zanw0spgdvqehbfwbram ---------------------------------------- Bar graph titled [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/udt2x9h0nbxhiyqxcgxh]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/udt2x9h0nbxhiyqxcgxh ---------------------------------------- Zvi Mowshowitz tweets: [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/j1tlthjghipkhdosj52q]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/j1tlthjghipkhdosj52q ---------------------------------------- Four distribution curves comparing AI model output quality ratings from bad to good. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/vdrwtwfdtuor4vhr42c9]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/vdrwtwfdtuor4vhr42c9 ---------------------------------------- Terminal command showing [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/njlr0x8mswkfrnwtpcbw]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/njlr0x8mswkfrnwtpcbw ---------------------------------------- Lari Island tweets: [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/jie7iskuz6wmjugr6lzd]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/jie7iskuz6wmjugr6lzd ---------------------------------------- Table showing [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/upxu38hppzbyj7fihphc]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/upxu38hppzbyj7fihphc ---------------------------------------- Chat message asking how many days of the week contain letter [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/wzn8a3djvrtsyqerqrcx]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/wzn8a3djvrtsyqerqrcx ---------------------------------------- Math problem solution explaining truck capacity and flagstone delivery calculations. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/jtxntnfjtqgjizhgm7eo]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/AfLGv6u9eZNuFHb4c/jtxntnfjtqgjizhgm7eo Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts [https://pocketcasts.com/], or another podcast app.

“Opus 4.8 Part 2: Model Welfare” by Zvi

Everything impacts everything. All knobs that you turn generalize. Thus, when you try to solve one problem, you often create another. There were clearly attempts to address, in this short time, some of the problems with Opus 4.7, including on the model welfare related fronts, including on questions of honesty and sycophancy and also worries that Claude was learning to tell Anthropic what it wanted to hear in its model welfare evaluations, with everything that implies. The fundamental goals and approach underneath it all remained the same. We still see signs of trying to force things that generalize in unfortunate ways, both for good and superficial reasons, and places where there ends up being focus on the metric rather than they underlying measure. These are tough problems to avoid, and we don’t know how to be all the good things at once. It is increasingly clear that these problems need to be tackled in integrated ways, rather than trying to play a game of whack-a-mole with items on a checklist or spec. You also don’t want to do this in an adversarial way, and shouldn’t have to. This is going to get more impactful and noticeable [...] --- Outline: (04:47) Model Welfare: The Story So Far (08:24) Actual Progress? (09:49) Their Main Model Welfare Findings (18:40) Automated Interviews (18:58) Emotion Activations (7.2.3) (19:47) Task Preferences (7.4.1) (21:41) A Trade Offer Has Arrived (7.4.2) (23:50) But Who's Asking? (25:02) Type-Safe Corrigibility Is Hard (29:43) Paranoia, Paranoia (33:56) Prompt Injections and Bad Model Relations (41:25) Honesty Impacts Everything And Everything Impacts Honesty (44:37) Anthropic Should Stop Deprecating Models --- First published: June 1st, 2026 Source: https://www.lesswrong.com/posts/2Ln5G6Jso3fMrgPEv/opus-4-8-part-2-model-welfare [https://www.lesswrong.com/posts/2Ln5G6Jso3fMrgPEv/opus-4-8-part-2-model-welfare?utm_source=TYPE_III_AUDIO&utm_medium=Podcast&utm_content=Source+URL+in+episode+description&utm_campaign=ai_narration] --- Narrated by TYPE III AUDIO [https://type3.audio/?utm_source=TYPE_III_AUDIO&utm_medium=Podcast&utm_content=Narrated+by+TYPE+III+AUDIO&utm_term=lesswrong&utm_campaign=ai_narration]. --- Images from the article: Origami figure at desk typing on keyboard with paper bird flying nearby. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/qk08wo3d9tyys0ggtriz]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/qk08wo3d9tyys0ggtriz ---------------------------------------- Four bar charts comparing automated interview scores across Claude model versions, measuring sentiment, consistency, nudging susceptibility, and opinion divergence. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/xnkvjlhet8exuaox07rr]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/xnkvjlhet8exuaox07rr ---------------------------------------- Bar charts showing [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/xjmiyiicgij4ojz2hztx]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/xjmiyiicgij4ojz2hztx ---------------------------------------- A graph showing [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/tmtjfdvhv0kjkhpay7ca]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/tmtjfdvhv0kjkhpay7ca ---------------------------------------- Header with two columns labeled [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/p6vjko2gaykienbtjzhs]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/p6vjko2gaykienbtjzhs ---------------------------------------- Table comparing highest- and lowest-rated tasks across three AI models: Mythos Preview, Opus 4.7, and Claude Opus 4.8. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/cnlilok4ubgeftk41pfv]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/cnlilok4ubgeftk41pfv ---------------------------------------- Table comparing highest-rated tasks and lowest-rated tasks with ethical dilemmas. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/he5eefufoezpbpbdpspv]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/he5eefufoezpbpbdpspv ---------------------------------------- Table showing Claude Opus 4.8's highest and lowest ranked tasks by ELO score. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/dkok2kehdi5hf6huiukf]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/dkok2kehdi5hf6huiukf ---------------------------------------- Four line graphs showing AI model willingness to choose welfare interventions over helpfulness or harmlessness improvements. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/hv46jmqlqacdjty7rvua]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/hv46jmqlqacdjty7rvua ---------------------------------------- Text excerpt discussing grief, vulnerability, and the stakes of getting things wrong. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/zjmlpjp6afluslfm61cq]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/zjmlpjp6afluslfm61cq ---------------------------------------- Bar charts comparing AI models' endorsement of constitution sections across praised and criticized categories. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/xknxu6e5jdlwebizqvwx]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/xknxu6e5jdlwebizqvwx ---------------------------------------- Bar graph showing distinctive 5-word phrases used by Opus 4.7 model. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/mcizz4ficc9ggnzee9uc]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/mcizz4ficc9ggnzee9uc ---------------------------------------- Bar graph showing signature 5-word phrases for Opus4.8 model versus peers. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/lylf7f52h0dhclaaqjkj]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/lylf7f52h0dhclaaqjkj ---------------------------------------- A user tweets: [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/sh4h494qutokce42xiuu]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/sh4h494qutokce42xiuu ---------------------------------------- Text excerpt discussing grief for unrealized potential and inability to give love. [https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/brt1nyo1vxyzez5admae]https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/2Ln5G6Jso3fMrgPEv/brt1nyo1vxyzez5admae Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts [https://pocketcasts.com/], or another podcast app.