Anthropic can’t help themselves… 🔗 **Source:** [Anthropic can’t help themselves…](https://www.youtube.com/watch?v=VUyib5mR8Pg) > [!note]+ Document Notes > tags: [anthropic, AI-models, developer-tools] > author: Theo - t3.gg > source: YouTube > reference: https://www.youtube.com/watch?v=VUyib5mR8Pg > date created: 2024-12-19 10:30 > > # 🏷️ Title: Anthropic can't help themselves… > <!-- deep-read:sections start --> ## 🧰 Action Inbox - [H1] Monitor Anthropic's API access policies for developer tools — 00:00 #business-strategy - [H2] Evaluate alternative AI models for coding tools — 08:45 #technical-evaluation - [H3] Track open-source alternatives to proprietary AI services — 12:30 #open-source ## 🧭 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) ### 🔗 References - [Anthropic Terms of Service](https://www.anthropic.com) - [G2I Hiring Platform](https://g2i.co) ## 📝 Summary - Anthropic has cut off Trey's access to Claude models, following similar restrictions on Windsurf and OpenAI - The restrictions appear motivated by fears of model distillation and competition from Chinese-owned companies - Anthropic is uniquely restrictive among AI labs, offering no open-weight models while competitors do - Developer tools are training models on Claude-generated data, creating competitive threats - The pattern suggests strategic timing to minimize backlash while protecting market position ## 📖 Main Content Anthropic has terminated API access for Trey, the VS Code-based AI coding assistant developed by ByteDance/TikTok, marking the latest in a series of restrictive actions against AI-powered developer tools. This follows similar cutoffs of Windsurf and OpenAI's access to Claude models. The company cites security concerns about Chinese-owned subsidiaries accessing their services, referencing their September 2024 policy update restricting sales to "unsupported regions." However, the timing coincides with these tools achieving significant performance improvements, particularly in coding benchmarks where Trey's integration with the Duba Seed model achieved a 78.8% score on SWE-bench verified. Evidence suggests these developer tools are training models using data generated from Claude interactions, creating what the author describes as models with "Claude brain" - exhibiting similar response patterns and behaviors. Anthropic appears concerned about model distillation techniques that could allow competitors to replicate their capabilities at lower cost. Unlike other major AI labs, Anthropic offers no open-weight models, maintains closed-source tooling, and selectively restricts API access to potential competitors. This pattern raises questions about their positioning as the "safety-focused" alternative to OpenAI while engaging in potentially anti-competitive practices. ## 🗒️ Notes #### ⏱️ [00:00](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=0) — [H1] Anthropic cuts off Trey's Claude access - Key idea: Anthropic terminated Trey's API access despite their close working relationship - Evidence: "Anthropic has hard cut off Trey's access to Claude models, which is kind of crazy" #### ⏱️ [01:30](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=90) — [H2] Pattern of restrictive behavior from Anthropic - Key idea: This follows similar cutoffs of Windsurf and OpenAI - Evidence: "They previously cut off Windsurf because according to Daario, it would be odd for us to sell Claude to Open AAI" #### ⏱️ [02:45](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=165) — [H2] Anthropic blocks access through cloud providers - Key idea: Restrictions extend beyond direct API access to cloud platforms - Evidence: "They're not letting these customers access it through those either via their terms of service" #### ⏱️ [04:15](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=255) — [H2] September 2024 policy update targets Chinese companies - Key idea: New restrictions specifically target companies with Chinese ownership - Evidence: "Companies from these restricted regions, including adversarial nations like China, continue accessing our services" #### ⏱️ [06:30](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=390) — [H1] Fear of model distillation drives restrictions - Key idea: Anthropic worried about competitors using Claude to train their own models - Evidence: "use our models to advance their own AI development through techniques like distillation" #### ⏱️ [08:45](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=525) — [H2] Windsurf's sw.5 model shows "Claude brain" - Key idea: Models trained on Claude data exhibit distinctive behavioral patterns - Evidence: "This model has clawed brain. It has clawed brain after being hit with a few hammers" #### ⏱️ [11:20](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=680) — [H2] Trey achieves top SWE-bench performance - Key idea: Trey's integration scored highest on coding benchmarks - Evidence: "Trey and Duba seed code got a 78% as of about two months ago" #### ⏱️ [13:15](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=795) — [H2] Seed coder model demonstrates self-curation - Key idea: New models can curate their own training data effectively - Evidence: "LMS can effectively curate code training data by themselves to drastically enhance coding capabilities" #### ⏱️ [14:30](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=870) — [H1] Anthropic uniquely restrictive among AI labs - Key idea: Only major lab with zero open-weight models and closed tooling - Evidence: "They're the only major lab that has zero openweight models. None." #### ⏱️ [16:45](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=1005) — [H2] Claude Code repository misleads about open source - Key idea: Popular GitHub repo contains no actual source code - Evidence: "This is probably the biggest repo on GitHub. 41.5K stars. That isn't an actual repo with actual code in it" #### ⏱️ [18:20](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=1100) — [H2] Strategic timing minimizes backlash - Key idea: Anthropic cuts access when competitors are established but vulnerable - Evidence: "They wait until they can cut and hurt the competitor and not themselves" #### ⏱️ [20:15](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=1215) — [H2] Developer backlash focuses on wrong target - Key idea: Users blame Trey instead of Anthropic for service disruption - Evidence: "Switching from Trey after their drop of anthropic models. Every coder knows Claude is miles ahead" #### ⏱️ [21:30](https://www.youtube.com/watch?v=VUyib5mR8Pg?t=1290) — [H1] Call for accountability despite model quality - Key idea: Good models don't justify anti-competitive practices - Evidence: "We can't let Anthropic keep getting away with fucking over the progress of the rest of the industry" ## 📚 References - [Anthropic Terms of Service](https://www.anthropic.com/terms) - [SWE-bench Verified Results](https://www.swebench.com) - [Duba Seed Coder Model Paper](https://arxiv.org/abs/2410.01734) ## 🗣️ Paraphrase - Anthropic has systematically cut API access to developer tools that pose competitive threats, citing security concerns about Chinese ownership while the real motivation appears to be preventing model distillation - Developer tools like Trey and Windsurf are achieving impressive results by training models on data generated from Claude interactions, creating cheaper alternatives with similar capabilities - Unlike other AI labs that contribute open-source models and tools, Anthropic maintains uniquely restrictive practices while positioning themselves as the ethical alternative to OpenAI ## 📰 Newsletter Anthropic has cut off API access for several popular AI coding tools, raising important questions about market competition in the AI space. While the company cites security concerns, the pattern suggests strategic moves to prevent competitors from using Claude's capabilities to train their own models. For WordPress developers, this highlights the importance of tool diversity and avoiding vendor lock-in. Consider exploring multiple AI assistants and keeping updated on alternative models as they emerge. Key takeaway: The AI tooling landscape is rapidly evolving, and business decisions by major providers can significantly impact developer workflows. Resources: [Learn WordPress](https://learn.wordpress.org/), [Developer Docs](https://developer.wordpress.org/), [WP-CLI](https://wp-cli.org/) ## ✅ Action Items #### DevRel - Developer advocacy: Track AI tool restrictions and communicate alternatives to community - Education: Create workshops on AI tool diversity and vendor risk management - Developer experience: Evaluate impact of AI model restrictions on developer tooling - Developer marketing: Position independence from single AI providers as competitive advantage - Community management: Monitor developer sentiment around AI tool access restrictions #### OSPO - Policy & Governance: Develop guidelines for AI service dependencies in projects - License Compliance: Review AI model usage terms and restrictions - Security Management: Assess risks of proprietary AI model dependencies - Software Inventory: Document AI service dependencies and alternatives - Developer Support: Provide guidance on AI tool selection and risk assessment - Community Engagement: Support open-weight model development and adoption - Strategic Alignment: Align AI strategy with open source principles #### Community - Maintainers: Evaluate AI tool dependencies in project roadmaps - Contributors: Document alternative AI assistants for development workflows - Sponsors: Consider funding for open-weight model development - Ecosystem: Coordinate response to AI model access restrictions #### WordPress - Core: Assess impact of AI restrictions on core development workflows - Core Performance: Evaluate AI-assisted optimization tool alternatives - Documentation: Update AI tool recommendations to include multiple options - Community: Discuss AI tool vendor lock-in risks at meetups and WordCamps - Training: Develop lessons on AI tool diversity and selection criteria ## 🧪 Hypotheses - AI labs will increasingly restrict API access to prevent model distillation by competitors - Developer tools training on proprietary model outputs will face legal challenges within 18 months - Open-weight models will achieve parity with proprietary models for coding tasks by 2025 - Market consolidation will lead to fewer independent AI coding assistants ### 📊 Methodology - Track API access changes across major AI providers quarterly - Benchmark open-weight vs proprietary models on coding tasks monthly - Survey developer tool preferences and vendor risk concerns - Monitor legal developments around AI training data usage ### ❓ Questions - Will Anthropic eventually restrict Cursor's Claude access as well? - How will developers adapt workflows if primary AI assistants lose model access? - What legal frameworks will emerge to govern AI model distillation practices? ### 🔮 Future Research - Comparative analysis of AI model licensing terms and restrictions - Developer survey on AI tool vendor lock-in concerns and mitigation strategies - Technical evaluation of open-weight alternatives to proprietary coding models ### 💡 Implications - Practice: Developers need backup AI tools and vendor-agnostic workflows - Policy: Industry may need regulation around AI API access discrimination - Security: Proprietary AI dependencies create supply chain risks - Education: Training should emphasize AI tool diversity and alternatives - Funding: Investment needed in open-weight model development ### 🧠 Learning Styles Assessment Reading level estimate: 10th grade #### 👀 Visual learners - Present: Screenshots of AI tool interfaces, benchmark comparisons - Improve: Timeline diagram of Anthropic restrictions, competitive landscape chart #### 🔊 Aural learners - Present: Video format with detailed narration - Improve: Podcast version discussing implications with industry experts #### 🗣️ Verbal learners - Present: Clear definitions of technical terms, business analogies - Improve: Glossary of AI industry terminology, comparison frameworks #### 🧩 Kinesthetic learners - Present: References to hands-on testing of different tools - Improve: Step-by-step guides for evaluating AI coding assistants #### 🧠 Logical learners - Present: Benchmark data, competitive analysis framework - Improve: Decision tree for AI tool selection, risk assessment matrix #### 🧑‍🤝‍🧑 Social learners - Present: Community reactions and developer discussions - Improve: Discussion prompts for developer meetups, collaborative evaluation rubrics #### 🧘‍♀️ Solitary learners - Present: Individual analysis and reflection on industry implications - Improve: Self-assessment checklist for AI tool dependencies, personal strategy worksheet #### 🌿 Naturalistic learners - Present: Ecosystem thinking about AI tool competition and evolution - Improve: Analogies to natural selection and market evolution ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive video content - Improve: Detailed alt text for any visual comparisons, transcript with timestamps #### Keyboard navigation - Present: Standard video controls accessible via keyboard - Improve: Chapter markers for easy navigation to specific topics #### Contrast and readability - Present: Good contrast in video interface - Improve: High contrast mode for any text overlays #### Text resizing and reflow - Present: Responsive video player - Improve: Transcript with adjustable text size #### Media and captions - Present: Full transcript provided - Improve: Professional captions with technical term clarification #### Forms and interactions - Present: Standard comment interface - Improve: Accessible voting on key points ### 🌍 DEIB Assessment #### Cultural breadth - Present: US-focused perspective on AI industry - Improve: Global implications of AI model restrictions, international developer impact #### Inclusive language - Present: Generally accessible technical discussion - Improve: Define industry jargon, avoid assuming technical background #### Neurodiversity support - Present: Clear structure with timestamped sections - Improve: Summary bullets, consistent heading patterns #### Representation - Present: Tech industry perspective - Improve: Diverse developer voices on AI tool restrictions #### Psychological safety - Present: Respectful criticism of business practices - Improve: Clear distinction between company criticism and personal attacks ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Well-supported connection between model performance and training data - Inductive strength: Pattern recognition across multiple Anthropic restrictions - Causal reasoning: Clear cause-effect relationship between competitive threats and access cuts - Analogical strength: Good comparisons to other platform restriction scenarios #### Fallacies - Ad hominem: Some personal attacks on Anthropic mixed with legitimate business criticism - Confirmation bias: Strong negative prior about Anthropic may color interpretation of actions #### Tone of voice - Overall tone: Passionate advocacy with technical expertise, some inflammatory language ### 🛡️ DARVO Analysis - Deny: Anthropic denying competitive motivations, citing only security concerns - Attack: Framing critics as misinformed or biased toward competitors - Reverse victim/offender: Positioning restrictions as necessary defense rather than market manipulation - Emotional cues: Strong emotional language about protecting innovation versus restricting access ## 🎯 Conclusion Anthropic's pattern of restricting API access to competitive AI coding tools reveals tensions between proprietary model protection and open development ecosystems. <!-- deep-read:sections end --> ## 📺 Video  --- # 📄 Full Transcript *Auto-generated from YouTube captions · timestamps roughly every 30 seconds · [watch on YouTube](https://www.youtube.com/watch?v=VUyib5mR8Pg)* **[00:00]** Remember Trey? The aid similar to Cursor and Windinsurf. The VS Code fork by our friends over at Tik Tok and Bite Dance. The one that was way cheaper than everything else. They've quietly been grinding making the product better and better. Going as far as coming to my apartment and sitting with me as I roast all the little things wrong with it. It's a team that really cares. Trey is a pretty cool AI ID and I've been pumped to see the work that they're doing. And it seems like some other people were not quite as pumped. As of yesterday, Anthropic has hard cut off Trey's access **[00:30]** to Claude models, which is kind of crazy because they've been using the Claude models for a lot of the new agentic stuff that they've been doing, especially the solo mode. Really felt like it was built around Claude and its behaviors and characteristics and abilities. This is far from the first time Anthropic has done this. They previously cut off Windsurf because according to Daario, it would be odd for us to sell Claude to Open AAI, which he doubled down on by cutting off OpenAI's access to Claude, which according to OpenAI they were using for benchmarks, but according to Anthropic, they were **[01:00]** using for getting data for training. Yeah, absolute mess. Not only is Anthropic the only major lab that will cut people off in this way, they go way further with it. where enthropics models are provided on platforms like Google Cloud and AWS through Bedrock. They're not letting these customers access it through those either via their terms of service and agreements with those providers. There's a lot to dive into here from why Anthropics being so sensitive to why it might actually make a lot of sense now, thanks to developments we've seen from both Windinsurf and Cursor that are a lot **[01:30]** scarier than Anthropic probably previously thought. Their edge seems to be waning and their fear is at an all-time high and the result is weird questionable decisions like cutting one of their biggest advocates and Trey and one of their biggest customers probably too off of anthropic models entirely. I'm not paid by any of these parties. I am an investor in cursor but none of the other things I'm talking about here. Do you know who is paying me though? Today's sponsor. AI has made a lot of things easier, but there's one in particular it has made way harder. **[02:00]** Hiring. I cannot tell you how many people I've talked to that have this giant pile of AI generated resumes they have to filter through in hopes of finding a good candidate. God forbid you have to scroll through LinkedIn and all of the terrible AI generated DMs that I've been getting non-stop. It's awful out there. Which is why today's sponsor went from a great product to basically essential for businesses trying to hire today. G2I is the best way to hire engineers right now. They have over 8,000 engineers ready to go. Part-time, full-time, US, EU, whatever you need, they'll figure it out. especially across **[02:31]** different tech stacks. Their goal is 7 days from interview to first PR. And I cannot fathom how that happens, but I've heard from enough people it does that I trust them at this point. It's really, really cool. And I really trust these guys. They put on my favorite conference, React Miami. It's so much fun. And the team running G2I, they get it. They're really locked in. I've personally linked them up with a couple companies in my portfolio to help them with hiring. And the quality of the candidates and the speed at which they reach out is just insane to me. They have a whole web platform they built to make it easier. You submit questions that you want to ask your technical **[03:02]** candidates. They then ask the questions, get a video response from the candidates that you can watch through and decide who you think is the best fit. And don't worry, it's not just this piece of software. They're a real team. They'll set up a shared Slack with you. They'll do weekly standups with you. They'll do whatever it takes to make you content with the experience. Almost like hiring a whole recruiting team, but way, way cheaper. And don't worry, all of the engineers that they're referring are really well trained on modern AI tools. So you won't have to worry about somebody being way slower because they refuse to use cursor. Stop hiring like it's the stone age at soy.link/g2i. So why is anthropic so uniquely **[03:33]** sensitive such that they cut off Windurf, they cut off OpenAI directly and now they're cutting off Trey? What do these companies have in common? What is it that Anthropic is so sensitive about? And in particular, why is Cursor different? Why is it that Cursor is not getting this weird treatment that the others are? I think the reason is very specific and it's a weird thing that these two companies have in common. Enthropic is scared of these two more than the others. I think the reason that they hate OpenAI so much is pretty **[04:04]** obvious and apparent. Enthropic came as a result of employees at OpenAI quitting and building their own alternative like model company that has went way better than anyone expected. Enthropic is killing it to be frank. And Trey is a very different reason. There is overlap to that reason though. Trey is happening because of this report that they put out September 4th of this year updating restrictions of sales to unsupported regions. Anthropic's terms of service prohibit use of our services in certain regions due to legal, regulatory, and security risks. However, companies from **[04:35]** these restricted regions, including adversarial nations like China, continue accessing our services in various ways, such as through subsidiaries incorporated in other countries. is my understanding that there are parts of Bite Dance and Tik Tok that are incorporated here. That's why they have an office for Trey in my city, San Francisco. This is also why I can hang out with people who work on Trey. I've hung out with the whole team at this point. They're great devs. They care a lot. They feel as American as any other team that I work with, but the company that owns them is Chinese and the **[05:06]** subsidiary in the US has allowed them to have access to Claude and it seems like Anthropic is now not allowing such a thing to happen. Companies subject to control from authoritarian regimes like China face legal requirements that can compel them to share data, cooperate with intelligence services, or take other actions that create national security risks. These requirements make it difficult for companies to resist the pressures regardless of where they operate from or the personal preferences of the individuals at the company. When these entities access our services through subsidiaries, they could use our capabilities to develop applications and **[05:37]** services that ultimately serve adversarial military and intelligence services in broader authoritarian objectives. They could also potentially use our models to advance their own AI development through techniques like distillation and to compete globally with trusted technology companies headquartered in the US and allied countries. This is a lot of words. The thing they actually care about is right here. use our models to advance their own AI development through techniques like distillation. Yeah, Anthropic doesn't want openweight models to come **[06:08]** back and compete with them, especially from China. This fear seems a little overplayed until you realize something important. Hear me out on this one. I was playing with Windsurf a bunch recently because I wanted to try out their new model sw.5 which is their super fast up to like 2,000 TPS model that they trained on data from users using Windsurf. It's hosted on Cerebrus because they have really fast infra for this type of inference and Cerebrris loves doing these types of weird partnerships. The reason I'm here is a bit different. I **[06:39]** was using this to try and upgrade my SnitchBench benchmark to use the latest version of the AI SDK. It's one of my favorite tests to run on new models, new applications, new background agent tools, all these types of things. So, I tried it and it runs really fast. I was impressed with how fast and surprisingly decent the model was. But I noticed something particular here. Outdoor went through all of this. It had screwed some things up and I told it that I don't think you are persisting tool calls correctly. This is what they should look like in logs. You're right. Huh. Later **[07:12]** on, I said it seems like the problem is all of the types we created. Can we do this without so many custom types? The types in the AI SDK V5 are much better and should be usable directly. Read the docs if you need to. You're absolutely right. That sound familiar? This model has clawed brain. It has clawed brain after being hit with a few hammers and squeezed into a much faster shape. But this model has very bad case of claw brain. It's hard to describe. It has a specific vibe with it **[07:44]** and how it does things that is so claudy like let me simplify this by using or perfect. Now I can see the exact types like this just this is claude. Other models do this some amount as well. Like apparently Grock 4 fast will match these patterns. I would suspect the majority that are doing this are doing it because they're trained on data that was created by Claude because Claude 3.5 sonnet was the first model good enough to do these types of things. Yeah, I don't have **[08:15]** official confirmation on this, but it seems relatively obvious that the SWE 1.5 model by Windsurf was trained on what anthropic models have been doing in their IDE. My honest guess is that they've collected a certain amount of data from people who are using clawed models in windsurf and used that to do a refined version of what's probably an openweight model to make it this fast and good for codegen for this use case. I suspect that cursor has done the same thing with their recent composer one model feels a bit more different than **[08:46]** the others do, but it's very clear that the thing that windsurf and cursor now have that allows them to do this is insane amounts of data of a user making a request in a giant codebase and a result coming out and a thumbs up or down on whether or not the user liked the result and what they then went and changed. That is super valuable data and it's not surprising to me at all that they're using this data to train new models and the result is these fast models that are not quite as smart but are surprisingly good. Another good point is that there's a lot of companies that sell data to these labs and most of **[09:18]** these data labeling companies are also going to use anthropic models for labeling their data which results in similar behavior. Makes sense. I would be very very surprised if any of these labs were training on data that didn't have claude involved in it somewhere, which definitely results in the the flavoring that we're seeing. Good friend of the channel, AI Code King, just showed up. If you're not already watching his videos, you should be. His shit's really good. He dropped this very fun bit of information I was not aware of, which is that Bite Dance's seed **[09:48]** model through Trey is currently crushing not on TurtleBench, but SWEBench. This is the highest current score on SWE verified. Trey and Duba seed code got a 78% as of about two months ago, a month and a halfish, around the same time that this updated restrictions page was published earlier in the same month. Seems like there's some salts going on here that Anthropic is scared of what's going on with these models that they almost certainly suspect at least are **[10:18]** being trained using the data that was collected from users using claude. The seed coder model is a titled let the code model curate data for itself. It's a family of lightweight yet powerful open source code LLMs comprising base instruct and reasoning models all of eight billion size. Seed coder demonstrates that with minimal human effort LMS can effectively curate code training data by themselves to drastically enhance coding capabilities. So this is a model trained on code data that was curated by another model and the results are that it's crushing in benchmarks and making anthropic very **[10:49]** sensitive. This makes so much sense. The mini SWE agent plus 4.5 sonnet was able to get a 70 versus the 78.8 that was just scored by Trey plus Duba. Maybe I shouldn't be so hard on them for having too many tools if they're able to score that high on things like this. Very interesting. So around the same time that this post came out by anthropic saying that they're going to be more restrictive on how companies that have Chinese ownership access their models. I'm guessing that was around the same **[11:19]** time that Windsurf and Cursor had started really diving into training their own models on the data that they had. And that made Enthropic scared. Enthropic doesn't want the insane amount of money, effort, research, time, and every other measurable thing that they've done that they've put into these models to be distilled out and thrown at other much cheaper things. And they're going to start fighting about this significantly harder. Some things that are worth knowing about Anthropic is that they're the only major lab that has **[11:51]** zero openweight models. None. Google has Gemma, OpenAI has Whisper, and now the GP OSS models. Llama's always been open source. Anthropic has not contributed meaning to openweight models. They're also the only major CLI with source not available. There's a lot of people who seem to think that cloud code is open source because there's a cloud code repo on GitHub and this repo on GitHub even has a license which is anthropic PBZ all rights reserved. Huh. What about the **[12:22]** code? Oh, I don't see any real code here. I see plugins and scripts. That's because cloud code doesn't actually have the cloud code source in this repo. This is for random community stuff like tracking issues which they now have 5,000 of. They did not open source cloud code. The code for cloud code is not in this repo. And I feel like they get away with this and a lot of people think cloud code is open source because this repo is really popular. This is probably the biggest repo on GitHub. 41.5K stars. That isn't **[12:56]** an actual repo with actual [ __ ] code in it. It's kind of insane. Anthropic just gets away with being loved by developers because they were the first to serve developers at all and now they just get away [clears throat] with it. Yeah. So the only one that's has zero open weight models, the only one that has a CLI with the source not available and they're also the only one that actively publishes research and benchmarks without releasing the majority of those benchmarks. Like it's they have a couple times like I'm really thankful they put out a misalignment bench in particular. I'm thankful for **[13:27]** that because they lied when they were talking about how OpenAI models performed on it by saying OpenAI's reasoning models don't understand the benchmark. And then when 4.5 sonnet had the same problem, they said that the model was now smarter than the benchmark and realized it was being tested. When in reality, OpenAI models noticed that earlier and even when they didn't, they scored a 0% murder rate on the murder bench that Opus scored over 60% on. Yeah, Enthropic has tried positioning themselves as the safety focused open AI **[13:58]** company that OpenAI pretends to be. And I wish they would hold themselves to that. I genuinely do because right now it feels like their excuses are being written by Claude. [snorts] Very real. Very, very real. And now as they look at companies like Trey, which have way more funding than either Cursor or Windsurf do, they're a real threat. Windsurf was not a threat when they were separate in their own company with a little bit of funding and a lot of weird UI problems to solve. Opening eye was a real threat because **[14:29]** they have more money than Anthropic. Trey is a real threat because they have real money and they're in China. Curser wasn't seen as much of as a threat. And at this point, if Anthropic was to pull Claude models from Cursor, I think that would be the the point where we all start to riot. It seems as though Anthropic is being really strategic here about how much can they take away without the average developer being mad at them because they're really relying on the average dev thinking of them as a good positive company trying to push the state-of-the-art forward. From the little I've heard, it seems like **[15:00]** anthropic employees don't even want to have an API because the API needs GPUs that they could be using for the things they want to do internally. Yeah, interesting. For better or worse, I still think Anthropic has the most pleasant to use models for doing these big agentic things, especially background tasks, because their models are way faster and good at handling large amounts of tool calls constantly going on. And a lot of these tools are built to handle that. Like when I was playing with Trey's solo mode, I asked it to list all of the tools it has **[15:31]** access to, and there's a lot of them. It has a run agent solo document, run agent solo coding, solo dev builder, search code, search by reax, view files, view folder, it has a ton of tools. And then when it runs the agent solo coding tool, that one has even more of them. Yeah, it's a lot. They built these tools around the quirks, behaviors, and things they observed claude doing. And now Claude isn't even an option, which sucks. If you want to bring your own API key, you still can. That's cool. And I'm **[16:02]** assuming that those requests are hopefully, fingers crossed, routed from your machine directly to OpenAI or Anthropics APIs. Not positive about that, but uh yeah, hard to know for sure. This is scary. I don't like the idea of Anthropic not letting specific businesses use their models because they don't like them. I hated when Google got banned from Amazon for selling chcasts because Amazon saw them as a competitor and didn't want easy access to chcast when they could be selling Fire TV. I **[16:33]** don't like that Apple was restricting other payment things from being available on iOS when you download apps depending on what industry category you're in. If Apple intends to compete someday, I don't like when businesses arbitrarily restrict access to their services based on the potential fear of competitors having a win from it. In this case, there's the angle of China bad and scary. And I understand that angle. I absolutely do. I'm more sympathetic here than I am most of the time. But this is a pattern from anthropic. They wait till something is **[17:03]** far enough along that it's now a concern, but not quite far enough that cutting it will hurt their image. They wait until they can cut and hurt the competitor and not themselves. Like when they did the open AI cut, that was right around when GBT5 was coming out, which meant it got destroyed in the news cycle. No one even cared. In fact, the writer who wrote the article we had pulled up here for the Wired, Kylie Robinson got fired from working at the Wired. That's how little Wired cared about this story and how little traffic this got. People don't care right now. **[17:35]** And I know a lot of the subtle reasons why. In fact, I'm sure if you scroll down and read this comment section, it's a bunch of people saying, "Why does Theo have such a hate boner for Anthropic? Why is Theo so mean to Anthropic? must be OpenAI paying him. I'll tell you from firsthand experience, OpenAI is kinder to Anthropic than I am. I am the hater, not them. Don't assume that I'm being paid to hate on this company when it could just be organic hatred. I just do not like how well they've positioned themselves where they're seen as the good guy when they're not doing good things. And you even see this on Twitter **[18:06]** when you look at the people who are complaining saying that this is Trey's fault. Switching from Trey after their drop of anthropic models. Every coder knows Claude is miles ahead. Just cancel my subscription. Damn it. I bought a year's Trey membership and now they're telling me cloud can't be used. Bytes move is really too scammy. No, Anthropic. Why Trey? Why? 80% of my reason to buy a year subscription was Cloud models. Will you give me an 80% refund then? At Trey, why did you remove Claude? GBT5 sucks. Just saw a poor anthropic employee trying to do damage control in my replies when I talked about this **[18:37]** claiming that this is just the supported region thing. Every other major ID has access to Claude. Windsurf has access because the OpenAI deal didn't go through and they got bought by Devon instead. And Enthropic has a big partnership with Devon. They don't want to lose. Curser, I wouldn't be surprised if they're cut off. Also notice he put Windser first here even though it's the least popular of these options. He knows what he's doing here. He knows that they were cut off and they reversed the cutoff when they realized that OpenAI wouldn't be making the acquisition. And **[19:08]** also by technicality, OpenAI doesn't have an IDE. And if they did, you know, Anthropic wouldn't let them because they've revoked the access already. Yeah, it kind of sucks that they keep getting away with this. And I will continue taking the PR hit whenever I [ __ ] about Anthropic models and anthropics business practices because I don't want this to keep happening. If Anthropic is going to be seen by developers as the good AI company, I really hope they start acting like it. For every good thing they do, like MCP, **[19:40]** which we could even argue about the goodness of, they're the only company that's doing some other weird, sketchy [ __ ] And we need to hold them accountable for this. We can't let Anthropic keep getting away with [ __ ] over the progress of the rest of the industry because we think that their models are really good at code sometimes. And I do honestly suspect they're going to start pushing their relationships and the limits of them with companies like Curser and now Windur and Devon. The Leopards will never eat my face. Let's hope this one doesn't get us cut **[20:10]** off with T3 chat. But uh if it does, no, I'm sorry and that I did it for the right reasons. Let me know what you guys think. Am I overreacting or is Anthropic not deserving of the positivity they currently have? Let me know. And until next time, peace nerds.