About this episode AI Unfiltered delivers a concise daily briefing, cutting through the noise to highlight the AI developments that genuinely shift the landscape. We analyze new models, product launches, research breakthroughs, and funding rounds, identifying true signals from an experienced researcher's perspective. Tune in to understand what truly matters, gaining clarity and foresight in the rapidly evolving world of artificial intelligence.
0:00 A new model from a four-person team in Paris just achieved ninety-four percent on the MMLU benchmark.0:06 That’s a score that puts it on par with models built by teams of thousands with billion-dollar budgets.0:12 In episode ninety-five, we covered Alphabet's massive eighty-four-point-seven-five-billion-dollar capital raise and the DeepMind exodus that followed.0:21 Today’s news from Paris makes you wonder if that giant war chest is aimed at the wrong war.0:27 Okay, let's break this down.0:28 The company is called Cognition Libre.0:31 The model is CL-1.0:32 And it did NOT get this score by being bigger.0:35 It’s a one-hundred-seventy-billion parameter model, which is large, but not frontier-scale anymore.0:41 The breakthrough is the architecture.0:43 They’re calling it a Sparse Activation Matrix, or SAM.0:46 Instead of lighting up the entire network for every single token, SAM intelligently routes the query to only the most relevant one to two percent of the model’s neurons.0:57 Think of it like this: current models are like turning on every light in a skyscraper to find a file in one office.1:04 It works, but the energy waste is astronomical.1:07 Cognition Libre found a way to build a smart elevator that takes you directly to the right floor, the right office, and turns on ONLY that desk lamp.1:16 The result?1:16 A ninety percent reduction in inference cost.1:19 Ninety.1:20 Percent.1:20 This isn't a minor efficiency gain.1:22 This is a phase change.1:24 It means you can run a top-tier reasoning engine on a local server cluster, not a state-sized GPU farm.1:30 This is the kind of leap that enables entirely new products.1:34 It’s what the engineers who left DeepMind were probably hoping to build.1:38 And it just showed up from a tiny, unknown team in France.1:42 The signal here is that the brute-force scaling era might have just peaked.1:46 The game is shifting from who has the most GPUs to who has the smartest architecture.1:51 Second, Adobe just pushed an update to Premiere Pro.1:55 This isn't some beta feature in a side menu.1:57 This is a new core tool called SceneGen, and it's live for all Creative Cloud subscribers as of this morning.2:04 You type a description of a scene—say, "a detective walking through a rainy neon-lit alley at night"—and it generates a high-fidelity, eight-second video clip.2:14 Here's what's different.2:15 One, the quality is there.2:17 No weird artifacts, no six-fingered hands.2:19 Two, and this is the BIG one, it understands continuity.2:23 You can generate a second clip, "close up on the detective's face, rain dripping from his hat," and it renders the SAME character in the SAME environment.2:32 It maintains character and scene consistency across multiple shots.2:36 This is the moment generative video stops being a toy for making viral memes and starts becoming a tool for professional workflows.2:44 B-roll, establishing shots, even simple cutaways can now be generated on the fly, inside the editor you’re already using.2:52 For corporate video, for social media content, for independent filmmakers… the economics of video production just changed.2:59 Radically.3:00 The barrier to creating professional-looking video didn't just get lower—it’s been vaporized.3:06 Finally, let's talk about the hardware.3:08 While everyone is watching Nvidia's stock and fighting for H100s, a startup called Lighthouse Photonics just closed a five hundred million dollar Series B.3:17 That's a huge number for a company with no shipping product.3:21 So what are they building?3:23 Optical processors.3:24 They're using light—photons—instead of electricity to perform matrix multiplications, which is the core mathematical operation of all modern AI.3:33 The advantage is speed and power efficiency.3:35 Light moves faster than electrons and generates almost no heat.3:39 Lighthouse is promising a one-hundred-X improvement in performance-per-watt over the best available GPUs.3:46 This isn't just another chip company.3:48 This is a bet on a post-silicon future for AI.3:51 Five hundred million dollars says that very serious investors believe the path forward isn't just more and more complex silicon, but a fundamental change in the physics of computation.4:02 It’s a long-term play, but it’s a direct challenge to the current hardware oligopoly.4:07 It’s the sound of the next foundation being laid while everyone else is still decorating the penthouse of the current building.4:15 So today, we saw the software get smarter, the tools get integrated, and the fundamental hardware get challenged.4:22 The biggest companies are spending fortunes to build bigger walls and deeper moats around their AI castles.4:29 But today's news—a hyper-efficient model, a professional tool for the masses, and a bet on light-speed computing—shows that the innovation isn't happening inside the castle.4:39 It's happening everywhere else.4:41 The game isn't about building the biggest thing anymore.4:44 It's about building the smartest.
