- AI For All
- Posts
- Never Enough
In partnership with
Hello readers,
Welcome to the AI For All newsletter! NVIDIA is feeling the heat from investors with revenue growth forecasted to slow, the AI industry finally admitted that scaling isn’t the answer, China is racing towards AGI, and Coca-Cola assaulted everyone’s senses.
Never Enough
NVIDIA posted its Q3 financial results and despite demand for its chips still sky-high, investors were perturbed by slowing sales growth. NVIDIA’s fourth-quarter revenue growth will slow to 69.5% from 94%. This was one of the rare quarters where sales did not at least double. Following these results, shares initially fell 5%. With expectations for growth this high, not even the world’s most valuable company could measure up. Furthermore, supply chain snags and hardware issues have complicated delivery of NVIDIA’s Blackwell chips, which reportedly experienced overheating.
With the news of LLM performance plateauing, analysts asked CEO Jensen Huang on an earnings call whether or not NVIDIA was prepared to accommodate new approaches to AI like the “test-time scaling” underpinning OpenAI’s o1, which gives a model more time and compute during the inference phase to “think” through prompts, hopefully leading to better answers. Huang referred to test-time scaling as “a new scaling law.” Interestingly, Satya Nadella said the exact same at this year’s Microsoft Ignite event. Writer is also directly addressing the limits of scaling with “self-evolving” models.
So, it seems the industry is finally admitting that scaling LLMs is not the answer. Instead, scaling compute for the inference phase of LLMs is the answer! 🤦 For the record, I saw this coming. First, it’s not ideal to make LLMs respond slower just to squeeze a few measly drops of performance out of them. Second, this does nothing to improve the economics of LLMs. Even with businesses spending $13.8 billion on AI this year for some reason, that doesn’t even begin to cover the $200 billion in CapEx alone.
The problem is LLMs themselves, no matter how you scale them. In last week’s newsletter, I wrote about how a cyber animist perspective could prove helpful in building true synthetic intelligence. The prospect of conscious AI is relevant to this. It is often imagined that consciousness is something that would emerge in sufficiently complex AI systems. But what if consciousness is the prerequisite for an intelligence that is energy efficient and requires very little data? What if consciousness comes first? If this is the case, as it is in humans, then it would flip the entire AI paradigm.
In other news, the U.S.-China Economic and Security Review Commission proposed an initiative to fund the development of advanced AI systems. Why? Because China is “racing towards AGI” apparently (they’re not). China doesn’t have or know anything that America doesn’t. I guess no one told the U.S. government that LLMs have plateaued. Unless this funding is going towards radically new approaches, simply buying more GPUs and building more data centers won’t move the needle. The fact that a chatbot released in 2022 has caused the whole world to lose it remains mystifying.
Lastly, Coca-Cola released an AI-generated commerical, and it’s hideous. The appalling sludge was soundly rejected in the comments section. There are numerous problems with the commercial that are obvious to creatives and people with good taste, so I will instead describe the bigger picture implications of this. While I don’t expect high art from a Coca-Cola commercial, I must ask: has our society just given up? The point of society is to facilitate the flourishing of its inhabitants. Injecting technological efficiency at the expense of quality misses this point. Generative video is a perfect example of technology for technology’s sake. It doesn’t enrich or improve culture, it lays ruin to it.
🔥 Rapid Fire
DeepMind introduces AlphaQubit to identify errors in quantum computers
Mistral AI releases Pixtral Large, a 124B open-weights multimodal model
Research shows AI is insensitive to underlying meaning in language tasks
Leaked Amazon documents reveal critical flaws in delayed AI reboot of Alexa
Microsoft introduces Copilot Actions and agents plus Azure AI Foundry SDK
Samsung unveils improved multimodal generative AI model Gauss2
IBM expands AI accelerator offerings in new collaboration with AMD
Dell Technologies boosts enterprise AI with Dell AI Factory advancements
Capgemini, Mistral AI, and Microsoft collaborate on generative AI adoption
Accenture expands AI-powered cybersecurity services and capabilities
Biden and Xi Jinping agree to avoid giving AI control of nuclear weapons
Anthropic pursues new benchmark with statistical approach to model evals
Google announces $20M in funding for AI-driven scientific breakthroughs
Start learning AI in 2025
Everyone talks about AI, but no one has the time to learn it. So, we found the easiest way to learn AI in as little time as possible: The Rundown AI.
It's a free AI newsletter that keeps you up-to-date on the latest AI news, and teaches you how to apply it in just 5 minutes a day.
Plus, complete the quiz after signing up and they’ll recommend the best AI tools, guides, and courses – tailored to your needs.
📖 What We’re Reading
“One of the most pressing concerns today is the potential impact AI-generated voice impersonations will have on financial institutions. In recent years, high-profile incidents (such as when an unnamed CEO was scammed out of $243,000) have highlighted the devastating impact of these sophisticated attacks, where hackers use AI-driven tools to create convincing voice deepfakes, leading to severe financial and reputational damage.”