Author: Evolving AI - Bewertung: 0x - Views:7
Sam Altman just dropped one of the biggest signals in AI this year: he told Stanford students he’s willing to bet that Transformers, the architecture behind ChatGPT, Gemini, and Claude, will be replaced by something as disruptive as Transformers replacing LSTMs. In this video, we break down what Altman actually meant, why Transformers have a brutal quadratic scaling problem (longer context gets exponentially more expensive), and why researchers are hunting for the “next architecture” using today’s models as the research tool. We also dive into the most promising alternative direction mentioned right now, state space models like Mamba, and why “linear scaling” could change everything about long-context AI, inference cost, and what agents can do. Then we zoom out to the other major releases that quietly prove we’re entering a new phase: Apple Research’s LiTo, which reconstructs realistic 3D + correct lighting from a single photo (a huge step for AR, product visualization, and content pipelines), Manus launching My Computer to run an AI agent directly on your desktop with permission-based command execution, Z.ai’s GLM-5-Turbo built specifically for long-chain agent workflows, and Mistral’s Leanstral, a specialized code agent for formal verification that tries to “prove code is correct” instead of just testing it. We end with InSpatio’s WorldFM, a real-time multi-view spatial AI model that runs on a single RTX 4090, hinting that robotics, AR, and world-modeling are about to get cheaper and way more accessible. If the Transformer era is ending, this is what the next era looks like: agents, verification, spatial AI, and a new architecture race.
SOCIAL SHARE CARD GENERATOR