sayzardInflectiv AI ⧉ (@inflectivAI)
Anthropic이 Fractile 같은 칩 스타트업과 함께 추론(inference) 역량을 강화하려는 것은 합리적이라는 의견이다. 이를 통해 AI 응답을 더 빠르고 저렴하게 대규모로 제공할 수 있을 것으로 보인다.
gulifare2A Fractile apresenta chips de processamento de IA com desempenho superior:
🚀 100x mais veloz que modelos atuais
💰 Redução de custos em até 10x
🤝 Parceria estratégica com a Anthropic para otimização de LLMs
O avanço promete transformar a infraestrutura de dados global. 💻⚡️
Al SuttonNow all the paperwork is complete, I'm happy to announce that I've been offered, and accepted, a position at #Fractile, who are working on some of the things that concern me the most about #AI and #LLM inference.
Fractile - Radically Accelerate Frontier Model Inference | Al Sutton
Now all the paperwork is complete, I'm happy to announce that I've been offered, and accepted, a position at Fractile, who are working on some of the things that concern me the most about #AI. After a few months of exploring #LLM technology, I've reached a point where I don't think it's ever just going away. It's shown itself to be useful enough, to enough people, to justify its existence, and so I'm confident it's not a "3D TV" style technology, but, instead, is more of a "smart phone" technology where it'll become a thing that's used, as part of daily life, without anyone making a big fuss about it. One of the glaring problems I currently see around AI is power consumption. Meta is expected to spend hundreds of billions of dollars on AI datacenters which will need multi-gigawatt power feeds. Google/Alphabet Inc. in their 2025 environmental report, showed they're continuing to increase their power consumption, and alluded that is due to their AI usage ramp-up. Similarly Apple, Amazon Web Services (AWS), etc., etc., etc., are all increasing their power needs to support the use of AI. Fractile is designing hardware to change that. They're designing hardware from the ground up to achieve a 10x improvement in tokens per second per Megawatt, with the capability to generate tokens significantly faster than today's systems. That means that if current inference systems using Nvidia's hardware were replaced by Fractile's systems, in the next hardware refresh (so 3 years time), we'd see AI inference power consumption drop by 90%, with users getting faster results as well. And that's a vision I believe needs to become reality. I'm joining to initially help with tooling. Hardware is great, but tooling can make a difference between a 50% power reduction, and a 90% one, by highlighting and helping implement optimisations. During my time I'm hoping to grow my knowledge of AI systems and hardware, and become even more useful to the company as it grows. I start in just over one week's time, and it's the most excited I've been about joining a company in a very long time. So if you've got some AI tooling tips, recommendations, or ideas, feel free to send them my way. https://www.fractile.ai/
Hillphantom