$100B AI Models

RX1, an open-source humanoid robot developed by Red Rabbit Robotics, offers enthusiasts the chance to assemble their very own for lower than $1,000. This human-sized robot features two arms able to grasping and manipulating objects, and it will possibly be operated remotely via computer connection, utilizing either machine learning algorithms or a VR headset interface.

Utilizing a mixture of 3D-printed parts and available components, RX1’s design emphasizes affordability and accessibility. Initial plans for the humanoid servo component are already accessible on GitHub, with comprehensive construct instructions, additional software, and further details slated for release within the near future. Read more.

The AI industry faces mounting pressure to justify its massive investments in infrastructure, with concerns over a possible bubble intensifying following a report by Sequoia Capital. Despite substantial expenditures on AI data centers by tech giants like Nvidia, AWS, Google, Meta, and Microsoft in 2023, revenue growth from AI applications has yet to match expectations.

In keeping with analyst David Cahn, these corporations must collectively earn around $600 billion annually to offset their infrastructure costs effectively, including data centers primarily powered by compute GPUs. Nevertheless, achieving such revenue stays a frightening challenge. Speculative investments and increasing competition from latest players like AMD and Intel within the AI hardware market are putting pressure on GPU prices, further muddying profitability forecasts for AI ventures. Read more.

The price of coaching AI models is experiencing a steep rise, with projections suggesting future models could require training budgets as high as $100 billion, in comparison with current costs of around $100 million.

Dario Amodei, CEO of Anthropic, discussed these estimates on the In Good Company podcast. He anticipates that by 2025 or 2026, advancements in algorithms and hardware could enable training costs to scale dramatically. This evolution reflects the shift towards more powerful AI systems able to outperforming humans across various tasks, albeit requiring substantial computational resources.

Corporations like Nvidia are already responding with innovations just like the B200 AI chip, priced between $30,000 to $40,000, aiming to fulfill the escalating demands for AI infrastructure. Read more.

Kuaishou’s AI video generator, KLING, is now accessible online. This updated version allows users to create videos as much as ten seconds long, with features like camera movements (panning, tilting, zooming) and support for negative prompts.

Free users could make three 10-second videos day by day using this improved model, but registration still requires a Chinese phone number, although demo videos could be viewed without signing up. KLING competes directly with OpenAI’s Sora by producing videos as much as two minutes long at 1080p resolution and 30 frames per second. It uses a 3D space-time attention system and a diffusion transformer to realistically simulate complex motions and interactions. Currently, KLING is out there as a public demo in China. Read more.