Nvidia, king of the data center server, was the surprise star of this year's Consumer Electronics Show (CES), the tech industry's version of the Golden Globes.
Late Monday, the chipmaker announced a slew of consumer-focused hardware, including a $3,000 "personal AI supercomputer" called Digits. Not to be outdone, Amazon-backed Anthropic is closing in on a new fundraising round at a $60 billion valuation, according to a Wall Street Journal report on Tuesday.
Call My Agent
Digits brings the power of Nvidia's AI-enabling cloud servers to home offices, and comes stocked with the company's AI software library, such as development kits and some pre-trained models. In other words, Digits will likely be most useful for hardcore hobbyists and at-home developers when it's released in May. But Nvidia says the platform will also allow for the integration of so-called computer-using AI agents — a.k.a., systems that employ large-language models to autonomously perform tasks on behalf of users, like sifting through email archives or even booking events or reservations online.
According to AI evangelists, agents are the next step in the robo-revolution, a level-up from our current chatbot status quo. Google, Microsoft, and Anthropic have already debuted agents of their own, in various capacities. With Digits, Nvidia wants to offer users a way to run agents directly at home, rather than relying on remote cloud servers. But just how powerful is the hardware?
- Digits, which is about the size of a standard hardcover book, comes packed with 128GB of unified memory (a normal laptop has maybe 32GB of RAM) and a GB10 chip capable of up to 1 petaflop of AI performance. Which means it can run AI models with up to 200 billion parameters — an amount of power that is hard to obtain today without renting cloud space.
- Ok, now the laymen's translation: That's just enough to capably handle industry-standard LLMs; the last-gen ChatGPT-3.5 requires about 175 billion parameters while Meta's top model, Llama 3.1, requires around 405 billion parameters. Nvidia says two Digits systems can be linked together to reach that threshold.
Mission Control: Anthropic began rolling out a developer beta for its AI agent system, the aptly named Computer Use, in October; Anthropic said then it wants to expand access to it early this year, just in time for its latest reported fundraising round. OpenAI, meanwhile, is preparing to launch its computer-using agents as soon as this month, The Information reported Tuesday. The roll-out has been deliberate for good reason: Via what's been dubbed "injection attacks," malicious actors have flooded the web with dangerous websites designed to lure in and trick AI agents into exposing sensitive user data, like credit card information. Building an agent capable of dodging such digital booby traps requires tons of safety testing. Just last month, OpenAI co-founder Wojciech Zaremba accused Anthropic of releasing an agent that came up short on safety mitigations, saying "If OpenAI would do the same, we would get tons of hate." Just make sure the good robots beat the bad robots, and we're good, guys, really.
Written by Brian Boyle
0 comentários:
Postar um comentário