{"id":"o5hKqYjQSOE","title":"DeepSeek V3 25x Cheaper than GPT-5, GPU Breakthrough (66x faster), Fine Tuned Models Outperforming","description":"Join the NEAR Legion at https://nearlegion.com/\n\nFollow us on X:   / nearlegion  \n\n#NEAR #nearprotocol #NEARAI #ai #ainews","channel_name":"NEAR Protocol","published_at":"2025-12-02T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/o5hKqYjQSOE/hqdefault.jpg","transcript_text":"Good morning Legion, happy Tuesday. Tomorrow on Wednesday at 12 p.m. eastern time is our next ecosystem live stream will be answering the mailbag questions, talking about upcoming events in the real world, and outlining the grand vision around Legion buildout in 2026. Bring a beverage and join us! Our A.I. stories of the late have been breaking news and world changing. Most of you don't realize that, but each and every advancement increases us closer and closer to a transformed ASI first World. Gradually, then all at once will be there. Our first story today is a story you can't miss. DeepSeek version 3.2 just dropped in. It's extremely competitive as an open source model. It beats Gemini 3.0 Pro on multiple math coding benchmarks, and the base model lands around GPT 5 level High. But here's the real catch. It's roughly 25 times cheaper than GP-5 and 30 times cheaper than Gemini 3 Pro on output tokens. The East-West divide continues to expand, and China is proving time and again there is a way to do A.I. that's open, cheap and with good results. Our second story is more a reminder that constant innovation along the AI supply chain is happening, and the status quo can be disrupted at any time. Now, whether that will happen remains to be seen. For today's purposes, a company known as Mind Aptiv has a new way to generate machine instructions with extreme speed, precision and, most importantly, energy efficiency. While some are touting it as a breakthrough in GPU optimization, MindAptiv proposes to offer 20 to 60 times faster performance with up to 99% less energy. Sounds a little good to be true, but it supposedly also runs on standard Hyperscaler GPU instances. But other companies like this continue to scale, and how the data center compute paradigm evolves are some of the most important industry topics we'll be following. Because hint: NEAR compute is a real thing in the pipeline. Our third and final story for the day on the difference between general purpose LLMs and niche fine tuned models for a specific use case. Shanghai hosted an AI vs Doctor competition to diagnose a 73 year old woman with complex symptoms. Chinese AI and a group of Chinese doctors both predicted correctly that this woman had gastric cancer. The generic US model failed. This case study is just one instance, hinting at the value of fine tuning a niche small model development over the expensive and gas guzzling generic models. This is another major paradigm we'll be tracking as the AI arms race continues at record pace. Thanks for tuning in! Be sure to join the Legion if you enjoyed this video. We'll do this all again real soon.","transcript_synced_at":"2026-11-03T04:46:52.924Z","segments":[{"start_seconds":0,"text":"Good morning Legion, happy Tuesday. Tomorrow on Wednesday at 12 p.m.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=0s"},{"start_seconds":5,"text":"eastern time is our next ecosystem live stream will be answering the mailbag questions, talking about upcoming events in the real world, and outlining","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=5s"},{"start_seconds":12,"text":"the grand vision around Legion buildout in 2026. Bring a beverage and join us! Our A.I.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=12s"},{"start_seconds":19,"text":"stories of the late have been breaking news and world changing.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=19s"},{"start_seconds":22,"text":"Most of you don't realize that, but each and every advancement increases us closer and closer to a transformed ASI first World.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=22s"},{"start_seconds":29,"text":"Gradually, then all at once will be there.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=29s"},{"start_seconds":32,"text":"Our first story today is a story you can't miss. DeepSeek version 3.2 just dropped in.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=32s"},{"start_seconds":36,"text":"It's extremely competitive as an open source model.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=36s"},{"start_seconds":39,"text":"It beats Gemini 3.0 Pro on multiple math coding benchmarks, and the base model lands around GPT 5 level High.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=39s"},{"start_seconds":47,"text":"But here's the real catch.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=47s"},{"start_seconds":48,"text":"It's roughly 25 times cheaper than GP-5 and 30 times cheaper than Gemini 3 Pro on output tokens.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=48s"},{"start_seconds":55,"text":"The East-West divide continues to expand, and China is proving time and again there is a way to do A.I. that's open, cheap and with good results.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=55s"},{"start_seconds":65,"text":"Our second story is more a reminder that constant innovation along the AI supply chain is happening, and the status quo can be disrupted at any time.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=65s"},{"start_seconds":72,"text":"Now, whether that will happen remains to be seen.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=72s"},{"start_seconds":75,"text":"For today's purposes, a company known as Mind Aptiv has a new way to generate machine instructions with extreme speed, precision and, most importantly, energy efficiency.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=75s"},{"start_seconds":85,"text":"While some are touting it as a breakthrough in GPU optimization, MindAptiv proposes to offer 20 to 60 times faster performance with up to 99% less energy.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=85s"},{"start_seconds":96,"text":"Sounds a little good to be true, but it supposedly also runs on standard Hyperscaler GPU instances.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=96s"},{"start_seconds":103,"text":"But other companies like this continue to scale, and how the data center compute paradigm evolves are some of the most important industry topics we'll be following.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=103s"},{"start_seconds":111,"text":"Because hint: NEAR compute is a real thing in the pipeline.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=111s"},{"start_seconds":115,"text":"Our third and final story for the day on the difference between general purpose LLMs and niche fine tuned models for a specific use case.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=115s"},{"start_seconds":122,"text":"Shanghai hosted an AI vs Doctor competition to diagnose a 73 year old woman with complex symptoms.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=122s"},{"start_seconds":128,"text":"Chinese AI and a group of Chinese doctors both predicted correctly that this woman had gastric cancer. The generic US model failed.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=128s"},{"start_seconds":135,"text":"This case study is just one instance, hinting at the value of fine tuning a niche small model development over the expensive and gas guzzling generic models.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=135s"},{"start_seconds":144,"text":"This is another major paradigm we'll be tracking as the AI arms race continues at record pace. Thanks for tuning in!","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=144s"},{"start_seconds":151,"text":"Be sure to join the Legion if you enjoyed this video. We'll do this all again real soon.","url":"https://www.youtube.com/watch?v=o5hKqYjQSOE&t=151s"}],"youtube_url":"https://youtube.com/watch?v=o5hKqYjQSOE","url":"https://near-pedia.xyz/videos/o5hKqYjQSOE"}