top of page
Writer's pictureRich Washburn

Week in Tech: AI Chips, Voice Updates, and SpaceX Double Landing


Etched's Sohu Chip: A New Benchmark in AI Hardware


This week, Etched.ai, a startup founded by two 21-year-old Harvard dropouts, Gavin Uberti and Chris Zhu, has introduced Sohu, a groundbreaking AI chip set to revolutionize large language model (LLM) acceleration. Sohu boasts performance metrics ten times greater than Nvidia's H100 GPU and delivers 140 times the throughput per dollar. By embedding transformer architecture directly into the chip, Sohu enables real-time processing of thousands of words, making it highly efficient for LLM tasks.


The chip uses multicast speculative decoding and beam search, allowing it to handle models with up to 100 trillion parameters. This specialization is akin to the ASICs used in Bitcoin mining, which vastly improved efficiency by focusing on specific tasks. Etched.ai's approach reduces complexity and cost, potentially setting a new standard for AI hardware.


Delays in OpenAI's GPT-4 Voice Capabilities


OpenAI has announced a delay in the rollout of its GPT-4 voice capabilities, initially slated for a late June alpha release. The advanced voice mode, designed to facilitate natural, emotion-rich conversations with AI, will now be available to a small group of ChatGPT Plus users in late July. The delay is aimed at refining the model's safety and reliability, ensuring it meets the high standards expected by users. This feature aims to bring AI interactions closer to human-like conversations.


Claude 3.5: Dominating the AI Arena


Anthropic's Claude 3.5 has surged to the forefront of AI performance, now leading in coding tasks and hard prompts while securing the second overall position in open LLM leaderboards. This model has demonstrated superior performance in benchmarks, particularly in generating longer, detailed code snippets, making it a valuable tool for developers. Claude 3.5's efficiency in task success and project completion rates outpaces major competitors like GPT-4 and Gemini 1.5 Pro..



SpaceX's Double Rocket Landing: A Marvel of Engineering


In a remarkable feat, SpaceX successfully landed two rockets simultaneously for reuse, showcasing the pinnacle of engineering and space exploration. This achievement not only underscores SpaceX's technological prowess but also highlights the potential for more sustainable and cost-effective space missions. The precision required to accomplish this is a testament to the advancements in space technology and the future of reusable rockets.



AI-Generated Video Games: A Glimpse of the Future


AI-generated content is making significant inroads into the video game industry. Recent demonstrations of AI-rendered video games have shown stunning visuals and realistic soundscapes, pushing the boundaries of what's possible in game development. Although real-time AI generation of such high-quality content is still computationally intensive, the potential for AI to revolutionize video game creation is immense. This technology promises to take video games to new levels of immersion and realism.


Apple and Meta's AI Integration: Privacy Concerns Halt Progress


In a surprising move, Apple has decided not to proceed with integrating Meta's AI models into Siri, citing privacy concerns. Despite the potential benefits, Apple prioritized user privacy, a core value of the company. This decision reflects the ongoing tension between advancing AI capabilities and maintaining stringent privacy standards. The collaboration could have offered significant enhancements to Siri's functionality, but the concerns over data privacy led to a reconsideration of the partnership.



Comments


bottom of page