project rollup
Default
301
mentions
across 18 keywords
across 18 keywords
net sentiment
+25%
107 positive
161 neutral
33 negative
volume & sentiment over time
May 13 – Sep 28, 2026 · 3-day buckets
1–7 of 7 keywords
| keyword | mentions | net | trend | top theme | srcs | last seen | |
|---|---|---|---|---|---|---|---|
| Quill | 32 | -10% | performance | 4 | 6d | ||
| Google SynthID AI image watermark discussion | 1 | +100% | — | 1 | 2mo | ||
| Opus 5.5 versus Claude Opus models | 2 | +100% | — | 1 | 2mo | ||
| SynthID | 16 | +100% | gemini / google | 1 | 5d | ||
| Opus 5.5 official announcement | 50 | +68% | pricing / model | 1 | 4d | ||
| Opus 5.5 model documentation | 51 | +61% | claude / pricing | 2 | 4d | ||
| Cline | 149 | +72% | code | 3 | 6d |
mentions
Showing 1–25 of 301 mentions
-
Yesterday, I did something crazy, deleted 10,000+ lines from my project repo. ▶Majorly all documentation and otherwise redundant tests that models keep making over the time, a deep clean, you can say! I decided to do something about my bloated documentation (Including AGENTS. md) after reading a lot about how latest i… Yesterday, I did something crazy, deleted 10,000+ lines from my project repo. ▶Majorly all documentation and otherwise redundant tests that models keep making over the time, a deep clean, you can say! I decided to do something about my bloated documentation (Including AGENTS. md) after reading a lot about how latest intelligent capability improvement in models lile opus 5.5 and Astra→ that folks have been talking about taking down the 'plan mode' from their harnesses (CRAZYY!!) When I started this project in Google's Antigravity using maybe Gemini 3.5 Flash or something, I created a lot of documentation around the architecture and some nuances using the most intelligeny model of that time: Opus 4.6 at that time in Antigravity, which is still there (Hopeless team maintain or governing the decisions for model selection). For the past two to three months, I've had these documentation trying to gauge and help my models maneuver and understand the priorities I have. But I realized that maybe it is creating a lot of noise and a lot of unnecessary context for the really intelligent models like Astra that I have started using, so that they are not unintentionally restrainted to deliver the value that they possibly can. So I've deleted every documentation (because I am a solo founder, I know exactly they were and what I what in my head) and only kept an agents. md file that contains the most important rules, more like a philosophy of the project. From today, I will be using a freshly installed Codex with slight personal configuration change along with Astra-reformed Pstack. Which is a Codex version of 'PStack by Lauren Tan @poteto ' at SpaceX, and I triple verified my version with Astra adversarial judge to perform a comparison between using Lauren's skills and the new Codex plugin that I've created to see if the inspired plugin that has been created by Astra is performing with the philosophy that Lauren has made PStack with OR not. And so now it's a new time to seeview source ↗
-
Compare Opus 5.5 vs GPT 6 Sol for coding, writing, agents, and marketing. See which AI model delivers better quality, speed, and value. #OpenAI #Claude #GPT6 #Opus5.5 https://t.co/S3UFei4bKe Compare Opus 5.5 vs GPT 6 Sol for coding, writing, agents, and marketing. See which AI model delivers better quality, speed, and value. #OpenAI #Claude #GPT6 #Opus5.5 https://t.co/S3UFei4bKeview source ↗
-
🎙️ Claude Opus 5.5 Is Now Live on https://t.co/He4LVLveJn: Built for Long-Running AI Agents 🚀🤖 1️⃣📢 A New Claude Generation Arrives on https://t.co/He4LVLveJn https://t.co/He4LVLveJn has added Claude Opus 5.5, described as the first model in the new Claude 5.5 family. Its positioning is notably different from a gene… 🎙️ Claude Opus 5.5 Is Now Live on https://t.co/He4LVLveJn: Built for Long-Running AI Agents 🚀🤖 1️⃣📢 A New Claude Generation Arrives on https://t.co/He4LVLveJn https://t.co/He4LVLveJn has added Claude Opus 5.5, described as the first model in the new Claude 5.5 family. Its positioning is notably different from a general-purpose chatbot. The model is designed around long-running agentic coding and complex knowledge work, targeting workflows that require sustained reasoning and execution. 2️⃣🤖 Agentic Coding Takes Center Stage The key theme is autonomy. Rather than handling only isolated coding prompts, Claude Opus 5.5 is designed for workflows where an AI agent can work through multiple stages of a task. That makes it relevant to larger software projects, debugging cycles, codebase analysis, and other development processes that cannot be completed effectively through a single prompt. 3️⃣📚 1M-Token Context Window A 1-million-token context window gives the model substantial room to work with large amounts of information in one workflow. For developers, that can be useful when dealing with: 💻 Large codebases 📄 Extensive documentation 🧩 Multiple project files 🔍 Long-running research tasks 🤖 Agent memory and context Large context does not automatically guarantee better output, but it expands the amount of information an agent can potentially consider at once. 4️⃣📝 Up to 128K Output Tokens The model also supports up to 128K output tokens. Combined with the large context window, this is particularly relevant for tasks requiring lengthy reasoning or substantial generated output rather than short conversational responses. The architecture is therefore aimed at workloads where both input and output can become unusually large. 5️⃣💰 Lower Running Cost Than Opus 5 https://t.co/He4LVLveJn states that Claude Opus 5.5 can deliver comparable performance to Claude Fable 5.1 on most tasks while costing approximately 40% less to run than Opus 5. If that cost relationshipview source ↗
-
This is why LLMS will NEVER be conscious!!! We nailed the toughest function of all. Full character swap or just face swap. It was soooo hard to nail. This at low res. Scaling up is easy since it's already managed. But this was so tough; now it's unlimited. And I even added a new function through my main agent that can… This is why LLMS will NEVER be conscious!!! We nailed the toughest function of all. Full character swap or just face swap. It was soooo hard to nail. This at low res. Scaling up is easy since it's already managed. But this was so tough; now it's unlimited. And I even added a new function through my main agent that can sharpen even more I did not ask it to generate a sad face at all. This is a frame of a video. But look at the man's face. Fully consistent with the video frame itself. It was the toughest of all. The first above is the original. Build 7 is live and finished. The docs and test kit are installed on your laptop, and the employees' briefing is updated (attached above). 2 things I have proven. Continuous learning is only possible through client-side harnesses. Otherwise, all compute starts from a generalized training baseline. Every context compaction is every new session. You see it yourself. This was really tough; it looked so fake. In such a childish way. 2. I also proved that a native open-weight TRANSFORMER model is wich my clipping studio tool is built on. It does not come across as sophisticated at all. I swear it comes like the old Will Smith video that went viral. Because it is one of the last truly open weight model that you don't have to pay for a license and wich is not banned in US & EU like some are. Everything after was paid. But I believed I could build it to the next level. I even invented functions genuinely. Like patching things from other professional software tools. Tried many ways to arrive at my goal. But I will probably use this function like never. I do not know why I would do a face swap if I can just have consistent characters. And have nailed the motion physics already in an earlier phase. And emotional expression. But since I have it in my tool as a perfectionist, it did not feel complete. Everything else came together. And even after these images, what you see now. I have built a large range of FUNCTIONALITIES purely forview source ↗
-
Models, Machines and Rockets: The Expanding Competition Among ChatGPT, Claude and Grok. https://t.co/yGKb4BWou0 Models, Machines and Rockets: The Expanding Competition Among ChatGPT, Claude and Grok. https://t.co/yGKb4BWou0view source ↗
-
Models, Machines and Rockets: The Expanding Competition Among ChatGPT, Claude and Grok. https://t.co/yGKb4BWou0 Models, Machines and Rockets: The Expanding Competition Among ChatGPT, Claude and Grok. https://t.co/yGKb4BWou0view source ↗
-
Claude Opus 5.5 on xhigh reasoning. Ran all day, 10 hours or so. --- Tools --- I didn't tell Claude to do any of this, it just used the API keys I put in its environment: - Generated images using gpt-image-2.5-flare via API - Generated 3D models using Hunyuan via API, and some from Poly Haven (CC0) - Got the iPhone D… Claude Opus 5.5 on xhigh reasoning. Ran all day, 10 hours or so. --- Tools --- I didn't tell Claude to do any of this, it just used the API keys I put in its environment: - Generated images using gpt-image-2.5-flare via API - Generated 3D models using Hunyuan via API, and some from Poly Haven (CC0) - Got the iPhone Duo model out of Xcode somehow and rigged it in Blender - Did all the motion graphics in Python - Prompted and generated the background music using Lyria 3 Pro via API - Generated the voice clips and sound effects using ElevenLabs via API --- Cost --- Raw API cost would've been ~$265. Actual cost was 10% of the weekly quota (20x Max plan), plus $16 in external API costs. I believe the Max plan gets something like 6x the Pro plan's overall usage, so this should fit in the $20 Pro plan, but you'd hit the 5h limit a few times. Annoying, but you can make it work. Claude usage - Model calls: 1,824 (including 8 review subagents) - Total tokens: 781M - Cache reads: 772M - Cache writes: 7M - Output: 1.5M (0.6M of it thinking) - Cost at Opus 5.5 API prices: ~$233 External APIs - 3D food models (Hunyuan 3D + bake-off, fal): ~$15 - Food photos (OpenAI gpt-image-2.5, ~115 images): free via Codex CLI but ~$14 at API prices - Music, voice, sound effects (Lyria 3, ElevenLabs, fal): ~$3.50 - Image model bake-off (FLUX.2 Pro, Gemini): ~$1 - External subtotal: ~$33 --- Prompt 1 (one-shot app, first draft of video) --- I want you to build an iOS calorie tracker app. It should have a super clean, minimalist design: very design-forward and aesthetic. Sweat all the tiny details and make everything feel super premium. Make every interaction extremely delightful. You can go so far as to write custom Metal shaders, custom UI components, etc., to make everything fluid, unique and delightful. The idea I have here is to make it one unified conversational experience. I should be able to page through different days, each of which is a conversation thread with an agent. Theview source ↗
-
Googleアカウントを持つすべてのユーザーが、Google Vids上でGemini Omni 1.1 Flashモデルを用いた1080pのAI動画を無料で生成できるようになったと発表 個人アカウントは月6本まで。 シーン延長、正確なシーン長設定、旧クリップのアップスケーリングなど新たな編集機能も追加された。すべての動画にはAI検出用のSynthID透かしが埋め込まれる。 https://t.co/LUAA3aobCa Googleアカウントを持つすべてのユーザーが、Google Vids上でGemini Omni 1.1 Flashモデルを用いた1080pのAI動画を無料で生成できるようになったと発表 個人アカウントは月6本まで。 シーン延長、正確なシーン長設定、旧クリップのアップスケーリングなど新たな編集機能も追加された。すべての動画にはAI検出用のSynthID透かしが埋め込まれる。 https://t.co/LUAA3aobCaview source ↗
-
Your next manuscript and grant will carry a hidden AI label. And your reviewers will read it wrong. All Claude outputs from its latest models now carry an invisible watermark. OpenAI and Gemini will soon follow. Here's what almost everyone is getting wrong: → An AI watermark is not an AI detector. Confuse them, and… Your next manuscript and grant will carry a hidden AI label. And your reviewers will read it wrong. All Claude outputs from its latest models now carry an invisible watermark. OpenAI and Gemini will soon follow. Here's what almost everyone is getting wrong: → An AI watermark is not an AI detector. Confuse them, and researchers get hurt. 1️⃣ A detector guesses. A watermark counts. Detectors like Turnitin and GPTZero read your writing style and make a guess. A watermark never reads your style. It runs a keyed statistical test on the word pattern (SynthID-Text, Nature 2024). No hidden characters. The signal lives in the word choices. When "overcast" and "grey" both work, a secret key settles the pick. 2️⃣ Anthropic tells you what the mark can't say. "It cannot distinguish 'Claude wrote this' from 'Claude heavily edited this.'" And it "doesn't say anything about ownership or authorship." What your reviewer will hear anyway: Claude wrote this paper. 3️⃣ It can be forged for under $50. ICML 2024: attackers can scrub watermarks or stamp one onto human-written text. Picture that in a hostile authorship dispute. 4️⃣ The rules never asked for word-level provenance. NIH targets ideas "substantially developed by AI." ICMJE targets accountability. A watermark says nothing about whose hypothesis it was or whether the analysis is sound. What to do: ↳ Write your disclosure now: tool, task, what stayed human. ↳ Keep receipts. Version history beats any vendor's scan. ↳ Never paraphrase just to strip a mark. That's concealment with extra steps. Your institution will write its first watermark policy soon. Somebody in that room needs to know a watermark from a detector. Make sure it's you. 💬 If a scan flagged your paper tomorrow, what evidence of your own work could you show? — Here’s the link to join 15,000+ enthusiastic researchers on AI in academic research: https://t.co/XTq6dj6kz3view source ↗
-
@brucebatman007 Según Gemini sí. Cita hasta que tiene patrones claros hechos con SynthID. @brucebatman007 Según Gemini sí. Cita hasta que tiene patrones claros hechos con SynthID.view source ↗
-
@Yeagerista1595 Son imágenes falsas, Chris, generadas por IA con SynthID. Universal ni siquiera ha confirmado haber comenzado el rodaje. @Yeagerista1595 Son imágenes falsas, Chris, generadas por IA con SynthID. Universal ni siquiera ha confirmado haber comenzado el rodaje.view source ↗
-
Google launches Gemini 3.8 TTS for voice cloning 8 TTS (Text-to-Speech). This isn’t just another update; it’s a significant move in the world of voice synthesis and automation. What Is Gemini 3.8 TTS? Gemini 3.8 TTS comes in two flavors: Gemini 7.8 Flash TTS for creative work and character voices, and Gemini 7.8 Flas… Google launches Gemini 3.8 TTS for voice cloning 8 TTS (Text-to-Speech). This isn’t just another update; it’s a significant move in the world of voice synthesis and automation. What Is Gemini 3.8 TTS? Gemini 3.8 TTS comes in two flavors: Gemini 7.8 Flash TTS for creative work and character voices, and Gemini 7.8 Flash-Lite TTS for large-scale projects where cost efficiency is crucial. The model boasts an impressive array of features designed to enhance voice generation across multiple languages and dialects. Voice Generation Across Languages One of the standout features of Gemini 3.8 TTS is its ability to generate voices from text descriptions in over 100 languages and dialects, including regional variations like Mexican Spanish, Quebecois French, and Scottish English. This versatility opens up new possibilities for content creators who need localized voiceovers. Cloning Voices with SynthID Another exciting feature is the capability to clone a voice from just a 30-second sample. The model even includes SynthID watermarks for verification purposes, ensuring that cloned voices are authentic and traceable. This could be a game-changer for those looking to replicate specific vocal styles or personalities. Handling Direction with Ease Gemini 3.8 TTS excels in handling direction through plain language stage directions embedded directly into scripts. The model can accurately interpret these cues, maintaining voice stability over long audio sessions and supporting natural dialogues between multiple speakers. It even nails non-verbal sounds like laughs, sighs, and interruptions. Benchmarks and Performance According to benchmarks from Hume AI Voice Design and Voice Arena, Gemini 3.8 Flash TTS leads in several categories: • Hume AI Voice Design: Score of 71.4 • Accent Accuracy: Score of 60.8 • Voice Arena: Top rankings for Japanese, Brazilian Portuguese, Vietnamese, Arabic, Mexican Spanish, and Hindi These scores reflect the model’s ability to replicate real human voiceview source ↗
-
Google's Gemini 3.8 Live now has "Live Avatar": a real-time animated persona that lip-syncs across 97 languages without losing sync, and can run background tool calls (e.g. hotel check-in) mid-conversation. Enterprise-only for now, SynthID watermarked. https://t.co/XWKTB940OE Google's Gemini 3.8 Live now has "Live Avatar": a real-time animated persona that lip-syncs across 97 languages without losing sync, and can run background tool calls (e.g. hotel check-in) mid-conversation. Enterprise-only for now, SynthID watermarked. https://t.co/XWKTB940OEview source ↗
-
https://t.co/0s0ZBntV9n https://t.co/0s0ZBntV9nview source ↗
-
On social media and in portfolios you only get platform-contaminated files, not original PSDs, so even official content checks are no longer credible. A SynthID or C2PA flag on an image or video is not proof it was created by generative AI. It simply means the file contacted generative AI at some point, which can happe… On social media and in portfolios you only get platform-contaminated files, not original PSDs, so even official content checks are no longer credible. A SynthID or C2PA flag on an image or video is not proof it was created by generative AI. It simply means the file contacted generative AI at some point, which can happen anywhere, including after it left the pipeline. I can add that to a frame grab from a Maya render by having OpenAI or Gemini enhance it or run a super-resolution pass on traditional video. Kindly stop talking out of your ass like you know how AI and content ID work, because you are clearly still in a soggy nappy.view source ↗
-
Ännu en dåligt gjord bluffpropaganda. (25 september 2026) påstår att ukrainska kvinnor massivt får inkallelser från rekryteringscentra (TCC) och att landet köpt 150 000 militärjackor för kvinnor, med slutsatsen ”ända till den sista ukrainskan”. Nej, inlägget stämmer inte. Det blandar verkliga fakta om frivillig kvinn… Ännu en dåligt gjord bluffpropaganda. (25 september 2026) påstår att ukrainska kvinnor massivt får inkallelser från rekryteringscentra (TCC) och att landet köpt 150 000 militärjackor för kvinnor, med slutsatsen ”ända till den sista ukrainskan”. Nej, inlägget stämmer inte. Det blandar verkliga fakta om frivillig kvinnlig tjänstgöring med överdrifter, förfalskade dokument och ett återkommande ryskt desinformationsnarrativ om att Ukraina skulle tvinga kvinnor i strid.intellinews.comInlägget från @benoitm_mtl De specifika påståendena i inlägget: Inkallelsebilden (Saltivskyi TCC i Charkiv till en kvinna född 1993, för ”dataverifiering” 30 september 2026): Ukrainska CPD och flera medier beskriver de cirkulerande fotona av kvinnliga inkallelser från Kyiv och Charkiv som förfalskningar som sprids via ryska Telegram-kanaler. Enstaka äkta kallelser för registrerade kvinnor kan förekomma, men det är inte en https://t.co/IWEB4J5vKO 150 000 kvinnliga militärjackor: Påståendet kommer från ryska källor som visade en screenshot av en påstådd Prozorro-upphandling. CPD och ukrainska medier har kontrollerat: ingen sådan upphandling finns i det offentliga systemet. Screenshoten är fabricerad. (Däremot har Ukraina gjort mindre, verkliga inköp av kvinnospecifik utrustning, t.ex. några tusen kvinnliga skyddsvästar till redan tjänstgörande personal.) https://t.co/0Jno383uq0 Uppföljningsbilderna med reklam om att ”om du klarar tvättmaskinen/strykjärnet/dammsugaren klarar du FPV-drönare”: Dessa är AI-genererade. StopFake, CPD och andra faktakollar visar att ukrainska försvarets kommunikationsavdelning förnekar kampanjen, att zsu. team leder till ett kasino (inte en officiell rekryteringssida) och att bilderna bär OpenAI:s SynthID-vattenstämpel. https://t.co/MMNsZL4Rsfview source ↗
-
Just a fuckin fake, again. No, the post is incorrect. It mixes actual facts about voluntary female service with exaggerations, forged documents, and a recurring Russian disinformation narrative claiming that Ukraine is forcing women into combat. https://t.co/0Jno383uq0 The post by @benoitm_mtl (September 25, 2026) cl… Just a fuckin fake, again. No, the post is incorrect. It mixes actual facts about voluntary female service with exaggerations, forged documents, and a recurring Russian disinformation narrative claiming that Ukraine is forcing women into combat. https://t.co/0Jno383uq0 The post by @benoitm_mtl (September 25, 2026) claims that Ukrainian women are receiving draft notices from recruitment centers (TCCs) en masse and that the country has purchased 150,000 military jackets for women, concluding with the phrase "down to the last Ukrainian woman." Regarding the specific claims in the post featuring the draft notice image (issued by the Saltivskyi TCC in Kharkiv to a woman born in 1993 for "data verification" on September 30, 2026): The Ukrainian CPD and several media outlets describe the circulating photos of draft notices for women from Kyiv and Kharkiv as fakes spread via Russian Telegram channels. While isolated, genuine notices for registered women may occur, there is no mass campaign. https://t.co/GknU9yyOeE 150,000 women's military jackets: This claim originates from Russian sources displaying a screenshot of an alleged Prozorro procurement. The CPD and Ukrainian media have verified that no such procurement exists in the public system; the screenshot is fabricated. (However, Ukraine has made smaller, genuine purchases of women-specific equipment, such as a few thousand female-fit protective vests for personnel already in service.) https://t.co/0Jno383uq0 The follow-up images featuring advertisements stating "if you can handle the washing machine/iron/vacuum cleaner, you can handle FPV drones": These are AI-generated. StopFake, the CPD, and other fact-checkers report that the Ukrainian Armed Forces' communications department denies the existence of the campaign, that the "https://t.co/0w7xJYRNJl" link leads to a casino site (not an official recruitment page), and that the images bear OpenAI’s SynthID watermark. https://t.co/MMNsZL4Rsfview source ↗
-
Google just made AI video creation much more accessible. Anyone with a Google account can now generate AI videos for free in Google Vids using Gemini Omni 1.1 Flash. You can: • Generate new scenes in 1080p • Extend scenes while keeping characters, lighting and environments consistent • Choose exact clip durations • … Google just made AI video creation much more accessible. Anyone with a Google account can now generate AI videos for free in Google Vids using Gemini Omni 1.1 Flash. You can: • Generate new scenes in 1080p • Extend scenes while keeping characters, lighting and environments consistent • Choose exact clip durations • Upscale existing AI-generated clips • Build videos from ready-made templates Every AI-generated clip also gets an invisible SynthID watermark. Google says Gemini 3.8 Flash-Lite voiceovers in 100+ languages are coming next. You can try it now at https://t.co/kg3cxLLJl6 AI video creation is quickly moving from specialized tools into everyday productivity apps.view source ↗
-
https://t.co/tHGElE3N0s https://t.co/tHGElE3N0sview source ↗
-
Production systems (Google SynthID-Text, and variants used by Gemini / Claude) use a different sampler: tournament / g-function scoring instead of a simple green-list bonus. Same idea, different machinery: use secret randomness to bias which plausible token wins, then score the sequence against that randomness. Th… Production systems (Google SynthID-Text, and variants used by Gemini / Claude) use a different sampler: tournament / g-function scoring instead of a simple green-list bonus. Same idea, different machinery: use secret randomness to bias which plausible token wins, then score the sequence against that randomness. The goal is the same trade-off — detectability without obvious prose damage.view source ↗
-
@engcool إتاحة Gemini Omni Flash مجاناً في Google Vids فرصة حقيقية للمطورين والمبتكرين في المنطقة. والأهم اعتماد SynthID لترميز المحتوى المولَّد، خطوة إيجابية نحو الاستخدام المسؤول. أتطلع لبناء مشروع تجريبي بها قريباً. @engcool إتاحة Gemini Omni Flash مجاناً في Google Vids فرصة حقيقية للمطورين والمبتكرين في المنطقة. والأهم اعتماد SynthID لترميز المحتوى المولَّد، خطوة إيجابية نحو الاستخدام المسؤول. أتطلع لبناء مشروع تجريبي بها قريباً.view source ↗
-
@SuzunamiRei 实际上这个synthid也是很好破坏的 拿个ai放大图片的处理一下 然后再把图片分辨率降回去就查不出来的 https://t.co/g9z4z60PjB @SuzunamiRei 实际上这个synthid也是很好破坏的 拿个ai放大图片的处理一下 然后再把图片分辨率降回去就查不出来的 https://t.co/g9z4z60PjBview source ↗
-
8/ AI image authenticity Apple is also preparing support for SynthID later this year. The goal is to help identify images that have been generated or edited using AI. Apple Reference Image + image metadata + SynthID are designed to provide multiple signals around image authenticity. This could be particularly relev… 8/ AI image authenticity Apple is also preparing support for SynthID later this year. The goal is to help identify images that have been generated or edited using AI. Apple Reference Image + image metadata + SynthID are designed to provide multiple signals around image authenticity. This could be particularly relevant for photographers, journalists and anyone working with visual content.view source ↗
-
Latest in Web3 : Google's Free 1080p AI Video with SynthID Could Reshape NFT Provenance Full story: https://t.co/68qJ87cbUK Powered by @Chain_GPT #ChainGPTAI Latest in Web3 : Google's Free 1080p AI Video with SynthID Could Reshape NFT Provenance Full story: https://t.co/68qJ87cbUK Powered by @Chain_GPT #ChainGPTAIview source ↗
-
来源税：LLM 水印如何影响 AI agent 的行为 Anthropic 在 Claude 中部署 SynthID-Text 水印，研究发现其会改变模型拒绝行为和 Agent 工具调用，且效应随密钥而异。 https://t.co/nxslYN6rOo 来源税：LLM 水印如何影响 AI agent 的行为 Anthropic 在 Claude 中部署 SynthID-Text 水印，研究发现其会改变模型拒绝行为和 Agent 工具调用，且效应随密钥而异。 https://t.co/nxslYN6rOoview source ↗