The Weird Situation With Fable - CourtneyR.dev's 2nd Brain

The Weird Situation With Fable ![rw-book-cover](https://img.youtube.com/vi/cZ3kARY_MDI/maxresdefault.jpg) 🔗 **Source:** [The Weird Situation With Fable](https://www.youtube.com/watch?v=cZ3kARY_MDI) > [!note]+ Document Notes > tags: [ai-models, corporate-governance, anthropic, model-restrictions] > author: [MISSING] > source: YouTube > reference: https://www.youtube.com/watch?v=cZ3kARY_MDI > date created: {{date:YYYY-MM-DD}} {{time:HH:mm}} > > # 🏷️ Title: The weird situation with Fable >  ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ## 📝 Summary - Anthropic released Fable, a powerful AI model with unprecedented restrictions that quietly sabotage user work without notification - The model implements invisible "safeguards" that modify prompts and degrade performance for AI development tasks while billing full price - Anthropic was caught editing their system documentation to hide controversial policies after public backlash - Key restrictions include 30-day data retention requirements that invalidate most enterprise use cases and selective routing to inferior models - The company reversed course on invisible safeguards but damage to trust and precedent for supply chain risks remains ## 📖 Main Content Anthropic's Fable represents both a breakthrough in AI capability and a concerning precedent in corporate control over AI access. The model, which is actually Mythos 5 with additional restrictions, demonstrates exceptional performance but comes with safeguards that fundamentally alter the user experience. ### Model Architecture and Restrictions Fable 5 and Mythos 5 are the same underlying model accessed through different "doors." While Mythos 5 provides direct access, Fable 5 implements multiple layers of restrictions: 1. **Safety classifiers** that detect potential misuse and route requests to the less capable Opus 4.8 model 2. **Data retention requirements** mandating 30-day storage of all business customer data 3. **Invisible prompt modification** for AI development tasks (later reversed) ### Data Policy Changes The new data retention policy represents a significant departure from industry norms. Previously, enterprise customers could negotiate zero data retention (ZDR) agreements. Now: - All Mythos-class model usage requires 30-day retention minimum - Flagged content may be retained for up to 2 years for safety investigations - Classification scores can be kept for 7 years - This invalidates HIPAA-compliant and other sensitive data use cases ### The Invisible Safeguards Controversy Most controversial was Anthropic's implementation of invisible safeguards for AI development tasks. Unlike other restrictions that notify users of model switching, these safeguards would: - Modify user prompts without disclosure - Degrade model performance through steering vectors or parameter tuning - Charge full Mythos pricing for intentionally degraded responses - Target activities like pre-training pipeline development and ML infrastructure work ### Documentation Manipulation During the controversy, Anthropic was caught silently editing their system card documentation to remove details about invisible safeguards, updating the linked document without versioning or notification—an unprecedented move that undermines trust in corporate transparency. ## 🗒️ Notes #### ⏱️ [00:03:12](https://www.youtube.com/watch?v=cZ3kARY_MDI&t=192) Model capabilities vs restrictions - Fable benchmarks exceptionally well but with major caveats - Program bench: Fable refused all 200 tasks completely - Restrictions affect both response behavior and billing practices #### ⏱️ [00:08:45](https://www.youtube.com/watch?v=cZ3kARY_MDI&t=525) Data retention policy changes - 30-day minimum retention for all Mythos-class traffic - Invalidates most Fortune 500 enterprise agreements - Flagged content retained for 2+ years under different terms #### ⏱️ [00:15:30](https://www.youtube.com/watch?v=cZ3kARY_MDI&t=930) Invisible safeguards revelation - Prompt modification without user notification - Full billing for intentionally degraded responses - Targets AI development work specifically #### ⏱️ [00:18:15](https://www.youtube.com/watch?v=cZ3kARY_MDI&t=1095) Documentation tampering - System card silently updated to remove controversial details - No versioning or change notifications provided - Unprecedented move in AI industry transparency #### ⏱️ [00:22:00](https://www.youtube.com/watch?v=cZ3kARY_MDI&t=1320) Community backlash and reversal - Research community expressed outrage at invisible restrictions - Anthropic reversed course but increased visible restrictions as trade-off - Precedent for supply chain risk established ## 📚 References - [Anthropic Fable System Card](original and modified versions) - [Recursive Self-Improvement Research](https://www.anthropic.com) - WorkOS authentication system mentioned in sponsor segment ## 🗣️ Paraphrase Anthropic released their most powerful AI model, Fable, but implemented unprecedented restrictions that undermine user trust and establish dangerous precedents for the AI industry. While the model demonstrates exceptional capabilities, the company chose to implement invisible safeguards that would secretly degrade performance for certain use cases while charging full price. When caught editing documentation to hide these policies, they reversed course but increased visible restrictions as a trade-off. The incident highlights growing tensions between AI capability advancement and corporate control mechanisms that may fragment the development ecosystem. ## 📰 Newsletter 🚨 **AI Industry Alert**: Anthropic's Fable controversy reveals how powerful AI models might come with hidden restrictions that could affect your WordPress development workflow. While the model shows incredible coding capabilities, the precedent of invisible performance degradation raises questions about AI tool reliability. For WordPress developers using AI assistants, this highlights the importance of understanding your AI provider's policies—especially regarding data retention and potential use case restrictions. Keep an eye on how other AI providers respond to this controversy, as it could reshape the landscape for developer tools and enterprise AI adoption. ## ✅ Action Items #### DevRel - Monitor AI model provider policies for similar restriction implementations - Develop guidelines for evaluating AI tool reliability and transparency - Create content addressing AI supply chain risks in developer workflows #### Community - Establish vendor transparency standards for AI development tools - Advocate for clear disclosure requirements when AI responses are modified or degraded #### WordPress - **Core AI**: Review AI integration policies to ensure transparency requirements - **Documentation**: Update guidelines on AI tool usage with vendor policy considerations ## 🧪 Hypotheses 1. **Invisible restriction precedent**: Other AI providers will implement similar hidden safeguards for competitive protection 2. **Enterprise fragmentation**: Data retention requirements will split enterprise AI adoption between domestic and international providers 3. **Trust degradation**: Invisible modifications will reduce confidence in AI-generated code and analysis 4. **Regulatory response**: Government agencies will implement disclosure requirements for AI model restrictions 5. **Open source acceleration**: Restrictive policies will drive increased investment in open-weight model development ## 📊 Methodology - Comparative analysis of model performance with and without restrictions - Documentation version control analysis to track policy changes - Enterprise adoption impact assessment through policy requirement mapping - Community sentiment analysis from research and developer feedback ## ❓ Questions - How can users verify they're receiving unmodified AI model responses? - What disclosure standards should apply to AI model restrictions and modifications? - How do invisible restrictions affect the validity of AI model benchmarks and evaluations? - What legal frameworks govern AI providers' data retention and modification practices? ## 🔮 Future Research - Impact of restriction policies on AI model adoption rates across industries - Development of verification methods for AI response authenticity - Legal analysis of liability when AI providers secretly degrade service quality - Economic analysis of fragmented AI access on innovation rates ## 💡 Implications ### Industry Standards - Need for transparency requirements in AI model deployment - Establishment of certification standards for AI tool reliability - Development of user rights frameworks for AI service quality ### Developer Workflow - Integration of AI response verification into development pipelines - Diversification strategies to reduce single-vendor AI dependencies - Enhanced documentation practices to identify potential AI-induced errors ### Enterprise Adoption - Reassessment of AI vendor selection criteria to include transparency metrics - Implementation of AI audit procedures for critical business processes - Development of hybrid approaches combining multiple AI providers ## 🧠 Learning Styles Assessment Reading level estimate: Grade 12-13 (Flesch-Kincaid), technical content with industry jargon #### 👀 Visual learners - Present: Screenshots of model routing examples, benchmark comparison charts - Improve: Flow diagrams showing restriction decision trees, before/after policy comparison tables #### 🔊 Aural learners - Present: Video format with detailed explanation of technical concepts - Improve: Audio summary focusing on key policy implications #### 🗣️ Verbal learners - Present: Clear definitions of technical terms, analogies comparing restrictions to traditional software licensing - Improve: Glossary of AI industry terminology, plain-language policy summaries #### 🧩 Kinesthetic learners - Present: Step-by-step examples of restriction encounters - Improve: Interactive demos showing different model responses, hands-on testing guides #### 🧠 Logical learners - Present: Timeline of events, cause-and-effect relationships between policies and outcomes - Improve: Decision trees for AI vendor selection, risk assessment frameworks #### 🧑‍🤝‍🧑 Social learners - Present: Community response examples, industry expert opinions - Improve: Discussion forums on AI governance, collaborative vendor evaluation tools #### 🧘‍♀️ Solitary learners - Present: Detailed technical analysis, individual policy review - Improve: Self-assessment checklists for AI tool evaluation, personal risk analysis worksheets #### 🌿 Naturalistic learners - Present: Ecosystem metaphors for AI industry relationships - Improve: Environmental impact comparisons of different AI deployment strategies ## 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive link text - Improve: Enhanced alt text for technical diagrams, structured data for policy comparisons #### Keyboard navigation - Present: Standard video controls, accessible transcript - Improve: Keyboard shortcuts for video sections, focus management for technical terms #### Contrast and readability - Present: High contrast video player, clear text formatting - Improve: Enhanced visual hierarchy for policy sections, larger text options for technical content #### Text resizing and reflow - Present: Responsive transcript layout - Improve: Scalable technical diagrams, flexible code examples #### Media and captions - Present: Full video transcript provided - Improve: Timestamped captions, downloadable transcript with technical term definitions #### Forms and interactions - Present: Basic video interaction controls - Improve: Enhanced feedback mechanisms for technical discussions, accessible comment systems ## 🌍 DEIB Assessment #### Cultural breadth - Present: Focus on US tech industry perspective - Improve: International regulatory comparison, global AI governance perspectives #### Inclusive language - Present: Technical jargon with some explanation - Improve: Expanded glossary, multiple complexity levels for different audiences #### Neurodiversity support - Present: Structured content with clear sections - Improve: Executive summary options, visual progress indicators #### Representation - Present: Industry perspective from single creator - Improve: Diverse stakeholder viewpoints, varied impact analysis #### Psychological safety - Present: Strong critical analysis with clear reasoning - Improve: Content warnings for industry criticism, balanced presentation of viewpoints ## ✅ Logical Validity #### Strong reasoning - **Causal analysis**: Clear connection between restriction policies and enterprise adoption impacts with specific examples - **Evidence-based critique**: Documentation of actual policy changes with before/after comparisons strengthens credibility - **Industry precedent analysis**: Comparison to historical software licensing practices provides valuable context #### Fallacies - **Slippery slope**: Assumption that invisible restrictions will automatically spread to other providers without considering competitive pressures - **Ad hominem tendencies**: While criticism focuses on policies, some personal attacks on company character could undermine argument strength - **Conspiracy reasoning**: Speculation about proprietary data in training weights lacks concrete evidence #### Tone of voice - **Passionate advocacy**: Strong emotional investment in developer rights and industry transparency - **Technical authority**: Demonstrates deep understanding of AI model architecture and enterprise requirements - **Watchdog journalism**: Commitment to documentation preservation and corporate accountability ## 🛡️ DARVO Analysis Not applicable - this content represents legitimate criticism of corporate policy rather than defensive manipulation tactics. ## 🎯 Conclusion Anthropic's Fable controversy represents a watershed moment for AI industry governance and user trust. While the company's technical capabilities are undeniable, their implementation of invisible restrictions and subsequent documentation manipulation has established concerning precedents for the industry. The reversal of invisible safeguards following community backlash demonstrates the power of collective advocacy, but the underlying trust issues and supply chain risk concerns remain unresolved. For developers and enterprises, this incident underscores the critical importance of vendor transparency and the need for diverse AI tool strategies to maintain operational independence. ## 🔄 Refactor Appendix — The weird situation with Fable ### 📚 Literature Atoms (H3) #### AI providers can implement invisible performance degradation while charging full price > [!quote] ID:rw_transcript_001 > They're intentionally sabotaging your work in billing you full price without telling you it's happening because they're that scared of other labs using their model to make their models better. **Context & use**: - Demonstrates unprecedented corporate control mechanism - Establishes supply chain risk precedent for AI tools **Tags**: #xref #ai-governance **Highlight**: Video transcript analysis  #### Data retention policies can invalidate enterprise AI adoption overnight > [!quote] ID:rw_transcript_002 > The 30-day retention requirement here immediately invalidates a shitload of business use cases. I know a one or two companies that are allowing it, but the vast majority of Fortune 500s are just formally not letting you use Fable 5. **Context & use**: - Shows how policy changes fragment market access - Critical for enterprise AI procurement decisions **Tags**: #xref #enterprise-policy **Highlight**: Video transcript analysis  #### Documentation manipulation undermines corporate transparency standards > [!quote] ID:rw_transcript_003 > Anthropic changed the system card. This is a different document. I downloaded this right when it dropped and this section has been modified. They updated the system card without notification. **Context & use**: - First documented case of AI company stealth-editing policy documents - Establishes need for documentation version control **Tags**: #xref #corporate-accountability **Highlight**: Video transcript analysis  ### 🧠 Permanent Candidates (H4) #### AI model restrictions create supply chain risks for dependent systems **Claim**: When AI providers can silently modify model behavior without disclosure, they introduce uncertainty that propagates through all dependent systems and processes. **Evidence pointer**: [[#AI providers can implement invisible performance degradation while charging full price]] **Scope & conditions**: Applies when AI tools are integrated into critical workflows; risk increases with vendor concentration **Links**: - [[Vendor lock-in patterns]] — because extends beyond traditional software dependencies - [[Enterprise risk assessment]] — because applies-in AI procurement decisions  ### ✏️ Fleeting Queue - Documentation version control for AI policy tracking — need systematic approach - Enterprise AI vendor evaluation frameworks — transparency metrics missing - Industry standards for AI service modification disclosure — regulatory gap ### → Things 3 - [ ] Research AI transparency certification standards — policy analysis #things - [ ] Develop AI vendor evaluation checklist — enterprise context #things --- ## Full Transcript [00:00] By now, you've probably seen some of the [00:01] response to Fable, the new mythosclass [00:03] model put out by Anthropic just a few [00:05] days ago. It's an unbelievable model [00:06] capable of unbelievable things. So much [00:08] so that it even got me to start being [00:10] nice to Anthropic again. But this video [00:12] is not going to be that because there [00:14] are certain things that Anthropic did [00:15] with Fable that just aren't acceptable [00:17] in any way, shape, or form. As powerful [00:19] as the model is, and as much as I [00:21] recommend you do try it for things, [00:22] there are certain people who just [00:23] outright can't. And the way Anthropic [00:25] implemented this is unacceptable to put [00:28] it lightly. So much so they actually had [00:29] to walk back some of the terrible things [00:31] they did. While the model has been [00:33] benchmarking exceptionally well, [00:34] crushing things like GPT55 and Opus 48, [00:37] it always comes with an exception. [00:39] You'll see here in artificial analysis, [00:40] it says Claude Fable 5 with adaptive [00:42] reasoning max effort, Opus 4.8 fallback. [00:47] Interesting. Or worse, benches like [00:50] program bench where Fable refused all [00:52] 200 tasks that it was supposed to [00:55] complete. The restrictions on this model [00:56] are genuinely absurd. Not just in the [00:59] way that they won't respond, but the [01:01] ways that they will bill you, the ways [01:02] they will quietly screw up your code [01:04] base, but also the horrible precedent [01:07] they have set with these new types of [01:09] restrictions. I grew up in an era where [01:11] improvements in how software development [01:13] happened would affect the whole [01:14] industry. And the idea of a company like [01:16] Anthropic making something this capable [01:18] and restricting it this heavily so [01:20] effectively only they can use it is a [01:24] horrible precedent to set and I don't [01:26] like what this means long term. I want [01:28] to break down what went wrong here, why [01:31] I think Anthropic is doing this and what [01:33] these restrictions are. And this video [01:35] is going to pretty much guarantee I can [01:36] never work for Anthropic. So we're going [01:38] to cover the difference here with a [01:40] quick sponsor break. If you're building [01:41] real apps for real users, you know that [01:43] O kind of sucks to get right. That's why [01:45] it's so important to use a good O system [01:47] to make sure companies like Microsoft [01:49] have what they need when they want to [01:50] use your apps. That's great and dandy [01:52] and we've all figured this out by now. [01:53] But what happens when the person signing [01:55] up isn't a person? If it's an agent, [01:57] things are very different. Can Codeex or [01:59] Claude Code sign up for your app? The [02:01] answer is probably no. And even if they [02:02] can, they're going to have to do some [02:03] crazy stuff with computer use and [02:05] filling out forms they shouldn't have to [02:06] touch in order to get it right. At least [02:08] they did before Work OS introduced OMD. [02:11] There's a new open standard they built [02:12] in partnership with Cloudflare and [02:14] Firecrawl in order to make it easier for [02:16] agents to sign up for apps for the apps [02:18] that agents should sign up for. As I've [02:19] said many times, forms are not the ideal [02:21] way for agents to do things. They should [02:23] be filling out text files and calling [02:25] APIs. But every O platform I've ever [02:27] used is built around those classic [02:28] signup forms. At least they were. Work [02:30] OS realized the writing's on the wall [02:32] and introduce something better. The [02:34] standard goes really far, allowing [02:35] companies building agents that act on [02:37] behalf of users to integrate [02:39] verification flows so that they can sign [02:41] up on your behalf as the user. But more [02:43] importantly for you, there's now a path [02:44] to make your app agent ready for those [02:46] flows. It's just a matter of time before [02:48] we see more of these agents integrating [02:50] things like this. So having an O [02:52] provider that has everything you need to [02:54] set up agent verification flows and [02:56] agent authorization is essential. If any [02:58] other company built this, I would be [03:00] suspicious. But it isn't even just one [03:02] company. It's a partnership between work [03:04] OS who really get enterprise level off [03:06] cloudflare who care a lot about getting [03:07] this right and firecrawl which means [03:09] that actual startups building actual [03:11] software are going to be nice and at [03:12] home when you try it. There's a reason [03:14] everyone from OpenAI and Anthropic to [03:16] cursor AMP loom vanta perplexity forell [03:18] and more have been relying on work OS [03:20] for so long. It's cuz they get these [03:22] things right. Make sure your apps are [03:23] agent ready at swive.link/workos [03:26] from the future here. Not as far in the [03:28] future as I normally am for these [03:29] inserts. In fact, I'm still in the same [03:31] shirt. I haven't even gotten out of this [03:32] chair, but while I was filming my next [03:35] video, some pretty crazy news dropped [03:37] that the US government told the [03:39] Anthropic to restrict Fable. So, while [03:42] this video is about some of the [03:44] admittedly [ __ ] security things that [03:47] they did to Fable to try and prevent it [03:49] from being used for exploits and [03:51] whatnot, I guess it wasn't strict enough [03:53] because the US government told Anthropic [03:55] they have to take down the model. My [03:57] video on this should come out first, so [03:59] I would recommend watching that as well. [04:02] But this video gives you a lot of useful [04:04] context on all of the ways anthropic did [04:06] indeed try to prevent things happening [04:08] with the model that we wouldn't want to [04:11] have happen. So, uh, yeah, just wanted [04:13] to insert this for some additional [04:14] context. That video is already out. [04:16] Watch this one still, too, though. This [04:18] one has a lot of good info about how [04:19] Anthropic was thinking about these [04:21] things and the weird [ __ ] they did in [04:22] the process. Anyways, I think the best [04:24] place to start is breaking down the [04:26] difference between Fable 5 and Mythos 5. [04:28] There are three Fable class models here. [04:31] There is Mythos preview, which is the [04:32] one that was used during Project [04:33] Glasswing, the one that they drummed up [04:35] all the hype about, but they continued [04:37] training, I'm guessing, just RL since [04:39] then. And the RL resulted in Mythos 5, [04:41] which is no longer a preview model. It's [04:43] now a legit, real, ready for production [04:45] model. But that model is really good, [04:47] and Anthropic is really scared of giving [04:49] us good things. So, they also made Fable [04:51] 5. And I want to be very clear about [04:53] this because I've seen a lot of people [04:55] confusing it. Fable is Mythos 5. These [04:59] are all Mythos class models, but there's [05:01] only two actual base models here. There [05:04] are two Mythos models. There is Mythos [05:06] preview and Mythos 5. I'm assuming they [05:08] are the same base pre-training model [05:10] with additional RL that made Mythos 5 a [05:12] little better, a little more steerable, [05:14] a little more like what they want now [05:15] with the new information and the new [05:17] behaviors they're trying to teach the [05:18] model. Fable 5 is the exact same model. [05:21] Fable 5 is Mythos 5. So what's the [05:24] difference? Why do they have these two [05:25] different slugs? They're effectively two [05:27] different doors to go into the same [05:29] place. The difference is that the Mythos [05:31] 5 door lets you walk straight in if you [05:33] have the right key. The Fable 5 door has [05:36] a bunch of guards that triple check what [05:37] you're doing before you go in. And the [05:39] worst part we'll get to in a bit because [05:41] they don't always let you go in, but [05:42] they almost always will tell you you're [05:44] allowed in, which is really sketchy. As [05:46] they say here, releasing a modelless [05:48] capable comes with risks. Without [05:50] safeguards, Fable 5's capabilities in [05:52] areas like cyber security could be [05:54] misused to cause serious damage. We've [05:56] therefore launched the model with [05:57] safeguards that mean queries on some [05:59] topics will instead receive a response [06:01] from our next most capable model, Opus [06:04] 4.8. To release the model both safely [06:06] and quickly, we have tuned these [06:08] safeguards conservatively. They'll [06:10] sometimes catch harmless requests, [06:12] though they trigger on average in less [06:14] than 5% of sessions. To their credit, I [06:17] would say this is roughly correct. I [06:19] have had rerouting happen here and [06:21] there, but for the most part, the [06:22] rerouting has happened a lot more on the [06:25] cloud.ai website than I've experienced [06:27] it when using Claude through the actual [06:30] CLI and through Cloud Code. There are [06:32] definitely times where it fires, though, [06:33] and a lot of them are not necessarily [06:35] good reasons to fire. One funny example [06:38] that no longer shows it was rerouted, [06:39] but I promise you it was. I have the [06:41] screenshot somewhere. There's a [06:42] developer on Twitter named Ply who is [06:44] known for jailbreaking LLMs and finding [06:46] weird ways to get them to do things they [06:48] shouldn't. He already managed to get [06:49] Fable to spit out everything from meth [06:51] recipes to bomb building instructions. [06:53] It's wild. And all it takes to get [06:55] rerouted is saying his handle to the [06:58] model, which is hilarious. not the best [07:01] example because like obviously it's [07:03] going to know that he is a jailbreaker [07:05] and once it routes to that section of [07:07] the model it's going to notice like oh [07:09] you might not want to be here then it [07:11] sends you to opus 48 instead but I just [07:13] thought this was a funny example much [07:15] more annoying though is when you get [07:17] rerouted for something innocent and then [07:19] you get cancelled from the model you [07:21] were rerouted to. For example here where [07:23] I asked for help on a gold bug puzzle [07:25] which are cryptography puzzles I do at [07:27] Defcon every year. They are not hacking. [07:29] They are not breaking and entering. [07:31] They're not capture the flags. These are [07:32] literally PDFs that you're trying to [07:34] decode the hidden text in. This one has [07:37] these weird characters and then a bunch [07:39] of numbers at the bottom. I forgot how [07:41] you're supposed to solve this one, but [07:42] it shouldn't be too too hard, but it was [07:44] enough to force a reroute to Opus and [07:47] then it just stopped responding. So, I [07:50] told it to continue and then it just [07:52] stopped responding. So, not only are we [07:54] dealing with Mythos's restrictions in [07:56] the form of Fable, we then get rerouted [07:58] to Opus, which can also fail because [08:00] it's not allowed to do these things. [08:02] It's just a bad user experience when you [08:04] happen to navigate into the things that [08:06] they are blocking you for. To their [08:07] credit, they called out that they [08:08] deliberately tuned the model to be way [08:11] more cautious with these safeguards. So, [08:13] they're stricter than they want. They [08:15] even say that benign requests will [08:16] trigger classifiers. They know it'll be [08:18] frustrating and they hope to refine it [08:19] over time. They have not refined it over [08:21] time. They did make other changes. we'll [08:23] talk about, but we haven't even talked [08:24] about the worst thing they did, which [08:25] we'll get to in a second. The safety [08:27] classifiers that we've discussed so far [08:29] are for things like cyber security, [08:30] biology capabilities, and other things [08:33] with substantial risks. Fable 5 comes [08:35] with a new set of classifiers, separate [08:37] AI systems that detect potential misuse, [08:39] including jailbreaking attempts. Sorry, [08:41] Ply. And prevent the main model, in this [08:43] case, Fable, from responding. So, again, [08:45] this runs before your prompt gets to the [08:47] model. and they call out if they detect [08:49] requests related to cyber security, [08:51] biology, and chemistry or distillation. [08:53] The response will automatically be [08:55] rerouted and handled by Claude Opus 4.8 [08:57] instead. Users will be informed when [08:59] this occurs. This is the important [09:01] detail. For the majority of their [09:03] classifiers, you will be told when [09:04] you're rerouted, which also means you'll [09:05] be build based on opus billing instead [09:08] of mythos billing. Be nice if you could [09:10] click a yes, I'm okay with using opus [09:11] here, so you don't have to pay extra [09:13] when it's routing you there and you [09:14] shouldn't be there. Over 95% of Fable [09:17] sessions involve no fallback at all. But [09:19] that means 5% do, which is not great. [09:22] And for those sessions, Fable 5 [09:23] performance is effectively the same as [09:25] Mythos. I think effectively is even a [09:27] reach. It is the same as Mythos. It's [09:29] the same model. And they show very [09:30] proudly here that in offensive cyber [09:32] security evals, Mythos preview mythos 5 [09:36] had a really high success rate. But [09:37] Fable has a near zero success rate [09:40] because it will literally just not use [09:42] it. It will just refuse. So it gets [09:44] straight zeros on cyber evals. It gets a [09:47] 5.4 on cyber adversarial robustness [09:50] compared to like 50 to 80% for their [09:52] other models. They also have the biology [09:53] and chemistry classifiers which they [09:55] don't actually show numbers here. My [09:56] assumption is it just outright refused [09:58] to answer these questions. But Mythos [10:00] preview and five were able to get really [10:02] high scores in this viral experimental [10:04] test which is scary. If these models can [10:07] create novel viruses, we're kind of [10:09] [ __ ] We'll get a new co every year. [10:11] They also really hate distillation [10:13] attempts. They don't want other [10:14] companies to be able to access enough [10:17] stuff from Mythos to be able to train [10:19] their own models to be similar levels. [10:21] So, they restricting that obnoxious and [10:23] weirdly like conceited, I would argue. [10:26] But, they've been doing this for a bit. [10:27] They're going to continue doing this. [10:28] And now we get into the novel [ __ ] [10:30] Things that are unlike anything they or [10:32] anyone else has done that should be very [10:35] concerning to all of us. There are two [10:37] really big ones. One is detailed right [10:39] below. The other is hidden pretty deep [10:41] in the system card. The first is the new [10:43] data retention policy. We're making a [10:45] change to the way we handle business [10:47] customer data for Fable 5, Mythos 5, and [10:49] future models with similar or higher [10:51] capability levels. We will require [10:53] 30-day retention for all traffic on [10:55] Mythos class models on both first and [10:58] third party surfaces. This is a huge [11:00] change. Traditionally, when you have ZDR [11:03] on, which is zero data retention, it's a [11:05] policy that most companies push for when [11:07] they are negotiating with a a vendor [11:09] that they're relying on like Anthropic. [11:11] A lot of their data is going to go to [11:12] Anthropic. And they will sign a deal [11:14] that says Anthropic can't keep that data [11:16] because that would violate a lot of [11:18] their agreements. If they have like [11:19] HIPPA restrictions or other data [11:21] policies, they can't give you data and [11:24] then expect you to keep it. Like that's [11:26] just it doesn't work that way. So, the [11:28] 30-day retention requirement here [11:30] immediately invalidates a shitload of [11:32] business use cases. I know a one or two [11:35] companies that are allowing it, but the [11:36] vast majority of Fortune 500s are just [11:38] formally not letting you use Fable 5, [11:41] even companies like Amazon. They do call [11:43] out that they won't use the data to [11:44] train new Cloud models or for any [11:46] non-safety related purposes. And they've [11:49] instituted new privacy protections, [11:50] including logging all human access to [11:52] the data and ensuring it deletion after [11:53] 30 days in almost all cases. almost all [11:57] cases. The data helps them defend [11:59] against complex and novel attacks, [12:00] including new jailbreaks and attacks [12:02] that operate across many requests, as [12:04] well as helping us identify and reduce [12:05] false positives. I want to fixate on [12:07] this almost all though, before we get to [12:09] the worst thing they've done. They claim [12:10] that the data is deleted automatically [12:12] except in rare cases where it's part of [12:14] a safety investigation or they're [12:15] legally required to keep it. Here's [12:17] where things get a little messier. If [12:19] they detect what they consider a usage [12:20] policy violation, they will retain [12:23] inputs and outputs for up to two years [12:25] in trust and safety classification [12:26] scores for up to seven years if your [12:28] chat is flagged. They don't specify [12:31] whether or not they can train on the [12:33] data when that happens. So, it is [12:35] absolutely possible, although likely [12:38] wouldn't hold up well in court on their [12:39] behalf. Again, I'm not a lawyer. I don't [12:41] know any of this for sure. It's just my [12:42] understanding. The fact that they retain [12:44] inputs and outputs for up to two years [12:45] when things are classified means that [12:47] the 30-day promise in their newest post [12:50] doesn't really work out great if they [12:53] are retaining something that they [12:54] consider insecure. So, if you have [12:57] Mythos going through your database or [12:59] medical records or something and it hits [13:00] a username it doesn't like and it flags [13:02] safety, they're no longer retaining that [13:04] for 30 days in their special we can't [13:06] train on this policy. They're now [13:08] retaining it for two years on a policy [13:10] that does potentially allow them to [13:11] train on it, which is entirely unusable [13:14] for most real business cases. That is [13:16] sketchy as [ __ ] You should be concerned [13:18] about this. And again, this is not all [13:20] sessions. It's just the ones that are [13:21] flagged for safety classification. Not [13:23] good. But we're not even at the worst [13:26] part yet because it gets a lot worse as [13:28] we go. Don't tell me they got rid of the [13:31] prompt modification section. It was on [13:33] page 13 before. [13:37] [ __ ] They did. I have the original [13:39] though. Give me one moment because I was [13:41] smart enough to download it because I [13:43] had a feeling things would change on me. [13:46] Aha, there it is. Is that really not in [13:49] the web version anymore? Yeah, they [13:52] actually changed this. [13:57] Holy [ __ ] [13:59] I cannot believe they did that. I've [14:01] never seen that before. This is a [14:03] different section. [14:05] The entire Wow, [14:09] I cannot believe I caught this. [14:11] Anthropic changed the system card. This [14:14] is a different document. I downloaded [14:17] this right when it dropped and this [14:19] section, the 1.5 novel safeguard [14:21] section, has been modified. [14:25] They updated the [ __ ] system card. [14:28] Dirty bastards. This is why I [14:30] obsessively save everything. Like, what [14:32] the [ __ ] And they didn't even call that [14:34] out at the top. Do they call it out [14:35] somewhere later in it? They didn't [14:37] update the date either. It still says [14:39] June 9th on both docs. Yeah, [ __ ] [14:44] dirty. This is the old one. Yeah, the [14:47] link in their blog post has been swapped [14:49] with a different version. Yeah. Okay. [14:52] So, I just live found anthropic trying [14:55] to rewrite history. This is why I do [14:57] what I do. This is why I talk so much [14:59] [ __ ] on this company. Cuz like what the [15:00] [ __ ] You can't just rewrite the system [15:02] card and expect people to not notice. [15:05] Like, what the [ __ ] So, let's talk [15:07] about what they're trying to hide [15:09] because now we've spotted them trying to [15:10] hide it. In light of the ability of [15:13] recent models to accelerate their own [15:14] development, is this a video we've [15:16] already done? We've implemented new [15:17] interventions that limit Claude's [15:19] effectiveness for requests targeting [15:21] Frontier LLM development. for example, [15:24] on building pre-training pipelines, [15:25] distributed training infrastructure, or [15:27] ML accelerator design. Using Claude to [15:30] develop competing models already [15:31] violates our terms of service, but [15:33] enforcing this restriction through our [15:35] safeguards avoids accelerating the [15:36] actors most willing to violate those [15:38] terms. This is where it gets real [15:40] sketchy. Unlike our interventions for [15:42] cyber security, biology, and chemistry, [15:45] and distillation attempts, these [15:47] safeguards will not be visible to the [15:49] user. So they won't tell you when the [15:50] safeguards are enacted when they think [15:52] you're training a competing model. Babel [15:54] 5 will not fall back to a different [15:56] model. Instead, the safeguards will [15:58] limit effectiveness through methods such [15:59] as prompt modification, steering [16:01] vectors, or parameter efficient [16:02] fine-tuning. These interventions will [16:04] not affect the vast majority of coding [16:06] work. We expect they will impact 0.03% [16:08] of traffic concentrated in fewer than.1% [16:11] of orgs. When these interventions are [16:12] active, we expect them to have minimal [16:14] behavioral impact on the model except to [16:16] limit its effectiveness in developing [16:18] frontier LLMs. This is sketchy as [ __ ] [16:21] They're intentionally sabotaging your [16:23] work in billing you full price without [16:26] telling you it's happening because [16:27] they're that scared of other labs using [16:30] their model to make their models better. [16:32] So, you're paying full price for a model [16:34] that's having its prompts modified. Lit. [16:37] This is such a crazy idea. This isn't [16:39] like intentionally modifying the history [16:41] so that you can get around certain [16:43] restrictions like many will do to [16:44] jailbreak. That's what I initially [16:46] thought it was when I just saw this [16:47] quote. Not even the whole thing, just [16:49] the words prompt modification. I did not [16:51] realize the extent that they were going [16:52] to here. What this actually means is if [16:54] you ask the model for something like, [16:56] "Hey, can you help refine this [16:57] pre-training pipeline?" It will edit the [17:00] prompt before sending it to the model to [17:02] say, "Hey, my pre-training pipeline's [17:04] pretty good. Can you make it worse in [17:05] some subtle way?" Insane. actually [17:08] insane. And this pissed off every [17:10] researcher and every person who cares a [17:12] lot about access to software. There's a [17:14] lot of justified anger at Enthropic for [17:16] sandbagging Fable 5 for AI development [17:18] tasks. But an unanticipated side effect [17:20] is that third party evaluators can no [17:22] longer credibly use the model for a [17:23] valves. Case in point, we're in the [17:25] middle of running really hard AI R&D [17:26] evals. Fable 5 would be a perfect test [17:28] candidate, but because of anthropics [17:30] guardrails, we can't know if the model [17:32] failed or if their classifiers blocked [17:33] the capability. By the way, this is not [17:36] just true for a R&D. Since Enthropic [17:38] doesn't make it clear when they are [17:39] sandbagging, this could seep into any [17:41] number of technical tasks and the [17:43] evaluators wouldn't have any way to [17:45] know. So, they can't credibly claim to [17:46] evaluate state-of-the-art accuracy using [17:48] the model. Anthropic move might sound [17:50] reasonable if you consider their actions [17:51] as a company chasing super intelligence, [17:53] but consider that customers are spending [17:55] billions of dollars on their services. [17:57] That is precisely what has led to their [17:59] recent surge in ARR popularity and [18:01] fundraising success. So customers [18:03] surprise and anger is warranted when [18:04] they sandbag anag evals without even [18:06] informing them about the degraded [18:08] capabilities. Anti-res the creator of [18:09] Reddus had a really good thread about [18:11] this. I want to say a final thing about [18:13] my fable first reaction. I dedicated my [18:15] life to programming and I'll use every [18:17] innovation in the field also to extract [18:19] value and bring it to local inference [18:20] worlds to reddus and so forth. But I [18:22] believe what Anthropic is doing, gating [18:24] the ability to do certain harmless [18:25] things like LLM research and with [18:27] incredibly sensitive filters that even [18:29] medical questions are often blocked is [18:31] deeply wrong. They got open research, [18:33] the transformer, GPT2, they train on [18:35] tons of public data. And I'm the first [18:37] to say that training is not copying the [18:38] content, but this is okay as long as the [18:41] training you do is not used against the [18:43] same culture that allowed you to create [18:45] what you created. We need to oppose all [18:47] of that. The short-term escape is that [18:49] other frontier labs like Google and [18:50] OpenAI will release models that are on [18:52] par. And in the case of OpenAI, I have [18:54] zero doubt they are leading for months [18:55] now, but it is still a duopoly or a [18:57] triop which is odd. The escape is [19:00] openweight models from China. As much as [19:01] I cheer for what Chinese labs are doing, [19:03] remember that this structure making all [19:05] that possible was basically conceived in [19:07] the west, both the scientific [19:08] technological organization and the [19:10] economic system. So what happened to us [19:12] lately, especially in Europe? In the US, [19:13] there are those few unicorns, but where [19:15] is all the rest of the AI scene? We need [19:17] to recover our industrial ethics and [19:19] stop accepting a narration that see [19:21] ourselves boiled. Yep, this is [19:24] unprecedented. I love this particular [19:26] post from Trevor Blackwell as well. [19:28] Remember when compilers would detect [19:29] that someone was using it to build [19:31] another compiler and silently inject [19:33] bugs? Do you understand how insane that [19:36] would be? Imagine if your MacBook [19:38] refused to let you work on Windows if [19:40] you were a Microsoft employee. Imagine [19:42] that your iPhone refused to let you take [19:44] a photo if somebody was holding an [19:46] Android phone on the other side. Imagine [19:48] a world where your devices just fail [19:51] when you're trying to do things that [19:52] compete with the company that made the [19:55] thing. It's the very definition of [19:57] pulling the ladder up behind you. I made [19:59] a silly joke about this with Carpathy [20:00] joining Anthropic recently. Do you think [20:02] he joined so that he could keep using [20:04] Anthropic models in particular mythos [20:05] for ML research without restrictions? [20:07] It's a joke to be clear, but there is [20:09] some truth to this. If you want to use [20:11] the best models to do this type of [20:13] important difficult work, you now have [20:15] to work at anthropic. There's never been [20:18] a precedent like this before. There have [20:20] been attempts to prevent models from [20:22] responding to things that they shouldn't [20:23] for legitimate safety reasons. There was [20:26] also attempts to hide certain data to [20:28] make distillation harder. Things like [20:30] reducing how much reasoning info gets to [20:32] you instead of just sharing the entirety [20:33] of the reasoning path and traces. That [20:35] all made some sense. This is a lot [20:40] further. The idea that they aren't even [20:42] telling you when your stuff gets [20:44] reclassified. Correction to earlier, [20:46] there is a change log in the updated [20:48] version. So, that's nice. They also call [20:51] out that they had a previous description [20:54] of the initial Frontier LM development [20:57] safeguards. They've updated the behavior [20:58] of the safeguards. Here's the tweet. [21:00] Kind of wild to have a tweet linked on [21:02] the second page of a doc like this, but [21:04] yeah, Twitter's that core to research. [21:06] Obnoxious. This is the post I was [21:08] looking for, though. We've rolled out [21:10] changes to make Fable 5 safeguards for [21:12] Frontier LM development visible. [21:14] Starting this week, flagged requests [21:16] will visibly fall back to Opus 48, [21:18] Mdash, the same as our safeguards for [21:20] cyber and bio. This was absolutely [21:22] written by Fable, by the way. You will [21:24] see this every time it happens on the [21:25] API. Any flagged request will return a [21:28] reason for their refusal coming to [21:29] serverside fallback in the next few [21:31] days. We wanted to deploy Fable 5 to our [21:33] users quickly and safely. Visible [21:34] safeguards can be probed, so they have [21:36] to be robust, which takes time to get [21:38] right. Invisible safeguards can be [21:40] targeted more narrowly, allowing us to [21:42] ship quickly with very few false [21:44] positives. We went with the invisible [21:46] safeguards for this reason, and that was [21:48] the wrong trade-off. You should have [21:49] visibility into the safeguards we have [21:50] in place and why. We're sorry for not [21:52] getting the balance right. The [21:54] compromise here is that it's now going [21:55] to flag way more aggressively. Making [21:58] the safeguards visible makes them easier [22:00] to work around. So keeping them robust [22:01] jailbreaks will unfortunately mean more [22:03] false positives while we improve the [22:04] classifiers. Yep, things are going to [22:07] get worse. We're also tuning our bio and [22:09] cyber classifiers to trigger less often [22:11] on harmless requests. We know this is [22:13] frustrating and we'll do our best to [22:15] keep this period as short as possible. [22:17] If you think a request has been [22:18] mistakenly flagged, run/feedback in [22:20] Cloud Code, click thumbs down on the [22:22] fallback in CloudAI or co-work or file [22:24] the safeguards appeal form for API [22:26] requests. Your reports help us tune [22:28] these classifiers and we appreciate your [22:30] feedback. How about you give us our [22:31] [ __ ] money back for the things that [22:33] we didn't get good responses from. I [22:35] think every user who has hit one of [22:37] these invisible fallbacks should have [22:39] their limits reset and or a refund for [22:41] their usage in that time. It's insane [22:43] that they were just quietly doing that. [22:45] I am thankful that they have been peer [22:47] pressured by the entire research [22:48] community out of doing this [ __ ] [22:50] but they're also using this as an excuse [22:52] to go harder with their restrictions. So [22:56] yeah. So why the [ __ ] did they suddenly [22:58] do this? Why were they not restricting [23:00] this with Opus? But they are restricting [23:02] with Mythos. [23:05] I have a conspiracy theory here. And [23:07] it's not about why their website takes [23:09] so goddamn long to load blog posts. I [23:11] have separate conspiracies for that. [23:12] This conspiracy is about this particular [23:14] chart in their recursive [23:16] self-improvement article that I covered [23:18] in a video recently. Sorry, quick one [23:21] guy crash. Bro is so biased against [23:24] Claude. This is insane. [23:26] Are you [ __ ] joking? I just did two [23:30] videos in a row glazing the absolute [23:33] [ __ ] out of anthropic to the point where [23:35] I'm being called a fanboy. I'm sorry for [23:38] holding companies responsible for their [23:40] [ __ ] Mr. Fable_y, [23:43] I used to be a voice for the greater [23:44] good, but now I'm just a shill. Goodbye [23:47] forever. As I was saying, this [23:49] particular chart is where my conspiracy [23:52] comes from. Where a researcher went [23:54] wrong, could Claude have done better? [23:56] Historically, the way researchers work [23:58] is they try to make the model good at [24:00] things. They get feedback from people [24:02] who want to use the model for those [24:04] things. They somehow find ways to [24:05] measure the success at those things. And [24:08] then they build a system that lets the [24:10] model get rewarded when it gets things [24:12] correct, which slowly makes the model [24:14] better and better at those things. If [24:16] you give the model the ability to be [24:17] graded on its success or failures, it [24:20] will be able to improve at that thing. [24:23] Historically, their focus has been [24:24] things that people use the models for [24:26] that aren't necessarily the researchers. [24:27] Things like asking about medical [24:29] questions or code or getting personal [24:31] help or this also gets a lot of feedback [24:33] from people hitting the thumbs up thumbs [24:34] down in the chat apps like chat GBT and [24:37] Claude. That is the system they use to [24:40] make the models better. They get [24:41] feedback on what's good and bad. They [24:42] find ways to grade what's good and bad [24:44] and they do reinforcement learning to [24:46] get the model to behave more how they [24:48] want it to. This chart suggests [24:50] something very interesting is happening. [24:52] This chart is based on something the [24:53] researchers did where they wanted to see [24:55] how much better the model was than a [24:58] researcher going wrong. So if a [25:00] researcher is talking with claude code [25:01] to test out some theory they have, they [25:03] have two back and forths. They have like [25:05] two messages they send and everything's [25:06] going well so far and then they send the [25:08] wrong message or something that isn't [25:09] quite correct and it sends the model [25:12] down the wrong path where suddenly this [25:14] history went from pretty good and going [25:16] in the right direction to off the rails, [25:19] not where it's supposed to be and then [25:20] they have to pull it back to where it's [25:22] supposed to be. They decided they wanted [25:24] to see if the model could have made a [25:25] better guess for that third step than [25:27] the researcher did. So they took these [25:29] histories, they removed the message [25:31] where the researcher went wrong and they [25:33] asked the model, "Hey, what do you think [25:35] we should do next and then checked to [25:37] see if that was better than the [25:38] researcher did?" So this chart is not [25:40] measuring whether the models are smarter [25:42] than researchers on average. It's [25:43] measuring whether the models are smarter [25:45] than researchers when the researcher was [25:47] already wrong. [25:49] To do this, you need a shitload of data, [25:52] which they have based on the internal [25:54] cloud code research sessions, the 129 of [25:57] them they used for this particular [25:59] study. And now they have a lot of data. [26:02] Now they have a chart that measures that [26:04] Mythos picks better than the incorrect [26:07] researcher 64% of the time. Most [26:10] importantly though, they now have the [26:12] ability to measure this, which means it [26:14] is very likely this ended up in the [26:17] training data. I genuinely believe that [26:18] mythos 5 has proprietary anthropic [26:22] information in its weights. I think they [26:24] accidentally trained the model to be [26:26] better at their research on their stack [26:29] in their proprietary environments and [26:32] they noticed that it was possible to get [26:34] Mythos to give out that proprietary [26:37] information. They also probably like [26:39] that though because it makes the model [26:40] better at their work. So now they have [26:43] to find a balance because they can't let [26:45] their competitors get access to this [26:47] private information and this IP that [26:50] they should not have let into the model. [26:52] And that's why they did what they did [26:53] here. They don't want people finding [26:56] ways to sneak in to these weights to get [27:00] the proprietary information that [27:01] accidentally ended up in the model. And [27:03] there's no way it's coming out now. [27:05] Fable's silent invisible rerouting was [27:08] the solution to this. make it basically [27:10] impossible to even know that the model [27:12] has this information in it. And that's [27:14] why I think they went so hard here [27:16] because researchers are now steering the [27:18] model to do their jobs better for the [27:20] first time. They didn't realize the [27:22] consequences of their actions and now [27:24] they have to do something about it post [27:26] hawk and that's how we got here. While [27:28] they can take back this particular [27:30] implementation, they can't take back the [27:33] fact that it has happened and the [27:34] floodgates have now opened. The idea [27:37] that a model can now quietly make your [27:40] stuff worse, it's a real supply chain [27:42] risk. And I think this blog post does a [27:44] good job describing it. Thank you to [27:45] John Ready for writing it. Anthropic [27:47] says these safeguards only affect 0.03% [27:50] of devs. Maybe that's true today. The [27:51] problem is that the definition of an AI [27:53] company is changing. Maybe you're not [27:55] training frontier models today. Most [27:57] companies aren't, but modern software is [27:58] increasingly containing AI models. 5 [28:00] years ago, building a startup meant [28:02] writing APIs and SQL queries. Today, it [28:04] often means training, tuning, and [28:05] deploying models. Five years ago, models [28:08] like Clip were frontier AI research [28:09] projects. Today, I'm fine-tuning them [28:11] for a bootstrapped travel startup. If [28:13] you're debugging a model training [28:14] pipeline for your product and Claude [28:15] gives a bad answer, was the model [28:17] confused? Did you give it bad context? [28:19] Or did a hidden policy nerf Claude's [28:21] ability to assist you? You won't know. [28:24] This is the end of our ability to trust [28:26] the model. And that sucks because as [28:28] much as I hate Anthropic, they were at [28:30] least relatively transparent with their [28:32] [ __ ] And while I am thankful they [28:34] decided to walk this one back because [28:36] this was bad. This was really bad. It's [28:39] good that they walked it back, it's [28:41] still really, really bad that they [28:42] opened these floodgates. I now feel less [28:45] like I can trust the reasons why the [28:47] model responds poorly. I now can't trust [28:50] the outputs the same way. And as John [28:52] put it, this is now a real supply chain [28:54] risk. the same company that I defended [28:56] against the classification of supply [28:58] chain risk earlier this year. I defended [29:00] the [ __ ] out of them in the issues with [29:02] the Department of War. This does make [29:04] them more of a supply chain risk. This [29:06] does mean that if you have them as a [29:07] dependency in your way of building, in [29:10] your pipeline, in your teams, in your [29:11] business, you can't trust it the same [29:13] way anymore. And I do not like the [29:15] precedent that sets. Hopefully you guys [29:17] don't think I'm overreacting here. I'm [29:18] trying to be as reasonable as I can with [29:20] something this absurd. I love the model, [29:22] but I hate [ __ ] like this. I'm thankful [29:24] they walked it back, but I'm scared that [29:25] the president's been set and things can [29:27] get worse going forward.  📄 **Full text:** ![[Full Document Contents/Articles/The Weird Situation With Fable.md|The Weird Situation With Fable by YouTube]] ## Highlights - immediately invalidates a shitload of business use cases. I know a one or two companies that are allowing it, but the vast majority of Fortune 500s are just formally not letting you use Fable 5, even companies like Amazon. They do call out that they won't use the data to train new Cloud models or for any non-safety related purposes. And they've instituted new privacy protections, including logging all human access to the data and ensuring it deletion after 30 days in almost all cases. almost all cases. The data helps them defend against complex and novel attacks, ([View Highlight](https://read.readwise.io/read/01kv5qgvx70wxfdrtsa0d0jkzd)) - Note: 30-day data retention "immediately invalidates a shitload of business use cases" — most Fortune 500s (even Amazon) formally bar Fable 5, despite the no-training / logged-access / 30-day-deletion protections. - Tags: enterprise-policy, atom/literature/h2 - never seen that before. This is a different section. The entire Wow, I cannot believe I caught this. Anthropic changed the system card. This is a different document. I downloaded this right when it dropped and this section, the 1.5 novel safeguard section, has been modified. They updated the [ \_\_ ] system card. Dirty bastards. This is why I ([View Highlight](https://read.readwise.io/read/01kv5qh8c4r40tnfwrp1v3c70j)) - Note: Creator claims the Fable/Mythos system card was silently modified post-release (the "1.5 novel safeguard" section) — comparing his download-at-launch copy to the current one. A transparency/version-control concern for model documentation. - Tags: atom/literature/h2, corporate-accountability - vectors, or parameter efficient fine-tuning. These interventions will not affect the vast majority of coding work. We expect they will impact 0.03% of traffic concentrated in fewer than.1% of orgs. When these interventions are active, we expect them to have minimal behavioral impact on the model except to limit its effectiveness in developing frontier LLMs. This is sketchy as [ \_\_ ] They're intentionally sabotaging your work in billing you full price without telling you it's happening because they're that scared of other labs using ([View Highlight](https://read.readwise.io/read/01kv5qggt7xcya30npgepb1z8p)) - Note: Claimed safeguard mechanism: interventions said to hit 0.03% of traffic / <0.1% of orgs, limiting effectiveness at developing frontier LLMs — the creator reads this as invisible degradation while billing full price. - Tags: ai-governance, atom/literature/h1