Sakana Fugu 不訓練模型改當指揮官，效能直逼 Anthropic Fable 5？ | S2E63

📖 This episode is sponsored by AiPPT 👉 Easy Talk in Silicon Valley AiPPT Exclusive Link: https://tinyurl.com/y47e9z4k 👉 Enter promo code JKtech for 25% off Have you ever had this experience: you've already planned out the content of your presentation, but you spend an entire afternoon on layout, choosing templates, and getting the visuals right, leaving little time for the actual content itself? AiPPT.com aims to solve this problem. It supports a variety of input methods. You can directly type in a topic, paste some random notes or markdown, upload a Word/PDF file, or even just send a URL for it to read the entire page. It generates a formatted presentation in seconds, complete with a cover, outline, body, and conclusion. This year, three new modes have been added: Classic, Flow, and Visual, corresponding to formal work reports, complex topics requiring step-by-step explanation, and more story-based content, respectively. It also features built-in AI-generated images, so you don't need to open other tools to upload pictures. If you happen to need to create presentations—whether it's a work report, a school assignment, or simply want to quickly organize an article into slides—you can use the exclusive link above or enter the discount code JKtech to enjoy a 25% discount. Just throw something in and see how it transforms in a few seconds; you'll get a feel for it. If you like my content, welcome to become a member to support me, so I can make the content more in-depth and better, and together we can make this channel the way we all want it to be! 👉    / @justkiddingtech   Sakana AI, a Japanese startup founded in Tokyo in 2023, recently posted a message on social media that essentially said: "Stop competing on computing power." Their Fugu Ultra, touted as comparable to Anthropic's Fable, doesn't actually train large models. Instead, it uses a mere 7-byte "commander" to manage the Opus 4.8, GPT-5.5, and Gemini 3.1 Pro processors. Can three heads are better than one? This episode will show you how it actually works. This company has quite a background. One of its co-founders was one of the eight authors of the 2017 Transformer paper; coincidentally, none of those eight people are still at Google, and even the one who won the 2024 Nobel Prize in Chemistry recently left. Is this a signal? I discussed my thoughts in the video. What I found most interesting was the training method for the 7B commander. It generates a complete "workflow," but doesn't directly provide answers. Its underlying scoring mechanism is so simple it's almost rudimentary, yet it exposes one of the most critical limitations of current AI progress: why some capabilities advance rapidly while others lag behind, even making "taste" the rarest thing. Of course, no matter how well it's explained, I still tested it myself. I topped up $20 and ran a test with a prompt; the results were somewhat unexpected. Whether it's worth using, and how it compares to Opus and Fable, I'll show you in the latter half of the video. What do you think of this approach of "combining existing models"? Feel free to leave a comment below after watching. 🔗 Link to "Silicon Valley Talk" 👉 https://linktr.ee/jktech (00:00) Introduction (01:28) AiPPT (03:09) A Japanese Company Breaks Through: Sakana AI and Fugu Ultra (04:42) Why Not Train Your Own Model Instead of Scheduling Others'? (06:07) Two Co-founders: David Ha and Transformer Author Llion Jones (06:55) None of the Eight Transformer Authors Remained at Google: A Major Talent Reshuffle (08:55) How Does Fugu Ultra Actually Work? (11:15) Commander Model: Trained with RL, Producing Workflows, Not Answers (13:22) Why Do Only "Verifiable" Capabilities Improve So Quickly? (15:08) This is Actually Driving Engineering: Will It Be Replaced by a New Model in Six Months? (16:19) Starting the Real-World Test: Can Benchmarks Still Be Trusted? (17:08) Community Feedback: Strong in Code Review, But Slow and Expensive (18:27) Pikachu Flappy Bird Real-World Test (19:32) My Thoughts on Sakana AI