Google Stacking GPUs, Algorithms of Thought Reasoning for LLMs, and A21 Gets Fresh Funding
AI Infrastructure Alliance
We’re dedicated to bringing together the essential building blocks for the AI/ML applications of today and tomorrow.
A controversial analysis of the industry dropped last week from a team that analyzes the semiconductor industry.? It divided the world into the GPU rich and the GPU poor.? It got a lot of play and back and forth on social media, with OpenAI ’s Sam Altman quipping that Google managed to get their internal marketing published.?
But the question is whether there really is a GPU divide and it sure seems like there is right now?? Some companies are stacking up massive GPU farms like Inflection and Meta and some struggling to get their hands on the chips and put together the teams that can harness their power efficiently.?
谷歌 's MEENA once held the title of the best, until OpenAI's GPT-3 came along with its superior parameters and token count. But Google isn't backing down - it's predicted to leapfrog GPT-4's pre-training FLOPS by 5x by the end of this year, and a whopping 20x by the end of next.
The GPU-poor, like AlephAlpha and Together, face an uphill battle, even as they gain industry recognition. They are playing catch-up to giants like Nvidia, whose DGX Cloud service boasts multiple times more GPUs and has larger enterprises as customers. The race isn't over yet though - Google's Gemini and efficient infrastructure make it a strong contender, with its AI infrastructure superiority stemming from a focus on systems over microarchitecture.
A divide is emerging between GPU-rich firms like Google and GPU-poor ones, with the latter struggling to keep pace. Multiple Chinese firms are gearing up, boasting over 100k GPUs by next year's end. Meanwhile, some Bay Area ML researchers are flaunting their GPU access like peacocks. This situation is causing a stir among startups and open-source researchers who are GPU-poor, leading them to waste time on counterproductive tasks. Europe is lagging behind due to lack of GPU investment, while the Middle East is upping the ante in large-scale AI infrastructure.
Despite the industry recognition, GPU-poor firms like HuggingFace and Together will struggle to compete with Nvidia's cloud service. Databricks might have a fighting chance, but it needs to seriously up its infrastructure spending game.
The GPU divide is real and it's reshaping the industry, influencing career decisions of top ML researchers and the success of start-ups.
In the world of AI, it's often a tale of haves and have-nots.
Artificial intelligence (AI) is not just a buzzword; it's a tool with the potential to make a significant impact on our world. Central to its evolution is the growing open source AI community, though many projects are hamstrung by a lack of resources for sustained development. Enter the a16z Open Source AI Grant program - a lifeline for developers, offering funding for projects that are pushing the boundaries of AI. Their inaugural batch of grant recipients are making strides in areas like instruction-tuning LLMs and developing tools for LLM inference, all while fostering a spirit of open collaboration. Remember though, while this exciting news is informative, it's not intended as investment advice. Always consult your own advisors for specific matters.
领英推荐
Google's Duet AI has sashayed its way into all Workspace apps, ready to lend a digital hand in tasks such as whipping up slide decks, crafting email responses, and even summarizing those lengthy documents. It doesn't just stop there - Duet also brings app-specific treats to the table, including AI-driven lighting and sound adjustments in Google Meet and automatic summarizing in Chat. But remember, it's not all smooth sailing. While Duet's services come with a $30 price tag per user, it's wise to double-check its results as the AI, like all of us, isn't perfect. As Duet makes itself at home in emails and documents, it's clear that Google is ready to lock horns with Microsoft in the race to harness the potential of AI in transforming our workspaces.
AI21 Labs , a Generative AI startup, has secured a cool $155 million in its latest Series C funding round. The cash injection, led by an impressive roster of investors including Walden Catalyst and Samsung Next, has catapulted the company's valuation to a hefty $1.4 billion. AI21 Labs is making waves with its flagship product, AI21 Studio, a developer platform for crafting text-based business apps, and Wordtune, a multilingual AI assistant that's already charmed over 10 million users. Despite stiff competition from tech giants like Google and Microsoft, AI21 Labs is holding its own, thanks to its large language models and refined control. With this new funding, the company is set to supercharge its R&D efforts and take its AI to the next level.
Google DeepMind just pulled the wraps off SynthID, a watermarking tool designed to indentify AI-generated images. The tool, initially available to users of Google's AI image generator, Imagen, uses a pair of neural networks to create an elusive watermark, which remains detectable even if the image undergoes editing or screenshotting. While this move marks DeepMind as the first Big Tech to openly launch such a tool, it's not without its skeptics. Critics question the tool's long-term durability, citing how easily text watermarks can be disrupted. Despite the tool being in its experimental phase, this development is a promising stride towards safeguarding against deepfake imagery and copyright infringements.
Microsoft is pushing the envelope in ML once again, this time with its "Algorithm of Thoughts" (AoT). The paper looks to amplify the reasoning prowess of language models, marrying human-like intuition with algorithmic exploration for an optimized search process. Not only does AoT sidestep the hurdles of current in-context learning techniques, but it also offers more dependable results while efficiently juggling costs and computations. The real kicker? Microsoft is perfectly poised to weave this technology into advanced systems like GPT-4, potentially revolutionizing the way language models "think".
Also this week: