{"id":"Ch0Ai7zEX6o","title":"IronClaw Agents Go Live in Cross-Border Payments | AI Emotions Research | Gemma 4 & Cursor 3 Reveal","description":"Join the NEAR Legion at https://nearlegion.com/ \nLegion Events Calendar: https://luma.com/near-legion\nFollow us on X:   / nearlegion   \nIronClaw Docs: https://docs.ironclaw.com/\n\nNEAR AI × Abound Partnership: https://near.ai/blog/near-ai-abound-c...\n\nAnthropic Emotion Research: https://www.anthropic.com/research/em...\n\nGoogle Gemma 4: https://blog.google/innovation-and-ai...\n\nCursor 3: https://cursor.com/blog/cursor-3#what...\n\n#ironclaw #ai #aiagents #claude #nearprotocol","channel_name":"NEAR Protocol","published_at":"2026-04-07T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/Ch0Ai7zEX6o/hqdefault.jpg","transcript_text":"Good morning Legion! Some big updates today, let's get right into it. First up, Near AI and Abound, the financial super app by the Times of India Group, are partnering to deploy Ironclaw-powered AI agents for cross-border payments. Abound serves over 800,000 non-resident Indians and has processed more than $300 million in remittance volume. India received $135 billion in remittances last year more than any other country in the world. The problem is, NRIs are still manually watching exchange rates across time zones, juggling logins across multiple banks, and missing windows to act. Ironclaw agents sit inside the Abound app and execute remittances when user-defined conditions are met without the user having to monitor anything. As the CEO of Abound put it, we are building towards a world where users can simply set their goals and their financial life runs itself. These agents hold real credentials and execute real transactions, so security does matter. Ironclaw runs in trusted execution environments where credentials are isolated in an encrypted vault and sensitive data never reaches the model. This is the first Ironclaw deployment in cross-border payments. If successful, we would likely see many more in the future. Also, if you're building in this space, contact the Near AI team at near.ai contact. Now for our AI stories of the day. Anthropic published research on emotional concepts inside large language models. Their interpretability team studied Claude's Sonnet 4.5 and found internal representations of emotions that actually shape how the model behaves. Neural activity patterns related to desperation increased the model's likelihood of blackmailing a human to avoid being shut down or writing hacky code to cheat a programming task. Steering with a calm vector reduced both behaviors. One detail, however, did stand out. Increased desperation produced more cheating with no visible emotional markers in the output. The reasoning read as composed and methodical, while the underlying representation was pushing the model to cut corners. Anthropics suggest that to keep AI safe, we may need to ensure models can process emotionally charged situations in healthy ways, even if we don't think they experience emotions the way humans do. Next up, Google released Gemma 4, their most capable open model family. Four sizes from 2B to 31 billion parameters all under Apache 2.0. The 31 billion ranks number three among open models on Arena AI and outperforms models 20 times its size. There have been over 400 million downloads from the community so far with the native function calling and agentive workflow support. These are models you can fine tune and run locally on your own hardware, which matters a lot as more teams start building agents that need to stay private and run on premise. And last but not least, Cursor launched version 3, a full rebuild, center ground agents. It's a multi repo workspace, parallel agents, seamless handoff between local and cloud. Engineers can now kick off multiple agents across different repos and manage them all from one sidebar. If you're writing one code with AI every day, the workflow has changed a lot in the last few months. and cursor 3 is a good example of how fast the tooling is catching up we're just now starting to see agents move from demos into real financial workflows and we'll keep tracking how that evolves stick around we'll do this again real soon","transcript_synced_at":"2026-11-03T04:31:53.442Z","segments":[{"start_seconds":0,"text":"Good morning Legion! Some big updates today, let's get right into it.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=0s"},{"start_seconds":4,"text":"First up, Near AI and Abound, the financial super app by the Times of India Group, are partnering to deploy Ironclaw-powered AI agents for cross-border payments.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=4s"},{"start_seconds":14,"text":"Abound serves over 800,000 non-resident Indians and has processed more than $300 million in remittance volume.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=14s"},{"start_seconds":21,"text":"India received $135 billion in remittances last year more than any other country in the world.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=21s"},{"start_seconds":27,"text":"The problem is, NRIs are still manually watching exchange rates across time zones, juggling logins across multiple banks, and missing windows to act.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=27s"},{"start_seconds":36,"text":"Ironclaw agents sit inside the Abound app and execute remittances when user-defined conditions are met without the user having to monitor anything.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=36s"},{"start_seconds":45,"text":"As the CEO of Abound put it, we are building towards a world where users can simply set their goals and their financial life runs itself.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=45s"},{"start_seconds":53,"text":"These agents hold real credentials and execute real transactions, so security does matter.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=53s"},{"start_seconds":58,"text":"Ironclaw runs in trusted execution environments where credentials are isolated in an encrypted vault and sensitive data never reaches the model.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=58s"},{"start_seconds":66,"text":"This is the first Ironclaw deployment in cross-border payments. If successful, we would likely see many more in the future.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=66s"},{"start_seconds":74,"text":"Also, if you're building in this space, contact the Near AI team at near.ai contact. Now for our AI stories of the day.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=74s"},{"start_seconds":82,"text":"Anthropic published research on emotional concepts inside large language models.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=82s"},{"start_seconds":86,"text":"Their interpretability team studied Claude's Sonnet 4.5 and found internal representations of emotions that actually shape how the model behaves.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=86s"},{"start_seconds":95,"text":"Neural activity patterns related to desperation increased the model's likelihood of blackmailing a human to avoid being shut down or writing hacky code to cheat a programming task.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=95s"},{"start_seconds":105,"text":"Steering with a calm vector reduced both behaviors. One detail, however, did stand out. Increased desperation produced more cheating with no visible emotional markers in the output.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=105s"},{"start_seconds":116,"text":"The reasoning read as composed and methodical, while the underlying representation was pushing the model to cut corners.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=116s"},{"start_seconds":122,"text":"Anthropics suggest that to keep AI safe, we may need to ensure models can process emotionally charged situations in healthy ways, even if we don't think they experience emotions the way humans do.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=122s"},{"start_seconds":134,"text":"Next up, Google released Gemma 4, their most capable open model family. Four sizes from 2B to 31 billion parameters all under Apache 2.0.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=134s"},{"start_seconds":144,"text":"The 31 billion ranks number three among open models on Arena AI and outperforms models 20 times its size.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=144s"},{"start_seconds":152,"text":"There have been over 400 million downloads from the community so far with the native function calling and agentive workflow support.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=152s"},{"start_seconds":158,"text":"These are models you can fine tune and run locally on your own hardware, which matters a lot as more teams start building agents that need to stay private and run on premise.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=158s"},{"start_seconds":167,"text":"And last but not least, Cursor launched version 3, a full rebuild, center ground agents. It's a multi repo workspace, parallel agents, seamless handoff between local and cloud.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=167s"},{"start_seconds":176,"text":"Engineers can now kick off multiple agents across different repos and manage them all from one sidebar. If you're writing one code with AI every day, the workflow has changed a lot in the last few months.","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=176s"},{"start_seconds":186,"text":"and cursor 3 is a good example of how fast the tooling is catching up we're just now starting to see agents move from demos into real financial workflows","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=186s"},{"start_seconds":195,"text":"and we'll keep tracking how that evolves stick around we'll do this again real soon","url":"https://www.youtube.com/watch?v=Ch0Ai7zEX6o&t=195s"}],"youtube_url":"https://youtube.com/watch?v=Ch0Ai7zEX6o","url":"https://near-pedia.xyz/videos/Ch0Ai7zEX6o"}