Forward Future

GPT-4.5 wasn’t a bad model, it was the wrong bet. Dylan Patel (@dylan522p) reflects on the OpenAI model that was supposed to become GPT-5, but didn’t. “Orion (GPT-4.5) was a pure bet on full-scale pre-training. Take all the data, build a ridiculously big model, and train it. That was the strategy in early 2024.” It did have promise: “It is much smarter than 4.0 and 4.1. It’s the first model that actually made me laugh. It’s genuinely funny.” But it never lived up to the hype: “In practice it wasn’t that useful. It was slow, expensive, and o3 was just better.” In hindsight, the core issue was simple: data didn’t scale fast enough. “They scaled parameters and compute, but the data didn’t scale with it. So you get overparameterization — a massive model that memorizes first and can’t generalize.” And early training runs sent the wrong signal: “People inside OpenAI were excited, early checkpoints were crushing benchmarks. But that was because it memorized so much, not because it understood more.”

Inside Cerebras’ wafer-scale data center, and why its architecture delivers performance traditional GPUs can’t touch. “We’re two-and-a-half thousand times faster at getting to data and then using it. That’s what gives this massive performance advantage: memory bandwidth.” To move the heat from a single wafer drawing 18 kW, Cerebras has been all-in on liquid cooling since 2017: “These are water-cooled machines, so they’re unbelievably energy-efficient. The blue line is cold water coming in; the red line is warm water going out.” Beneath the raised floors: - Cold water supply and chilled-water return - Orange valve blocks from Belimo monitoring pressure and flow - Sensors feeding real-time telemetry back to Cerebras All of it ties into a 6,000-ton chiller plant that manages temperature and humidity deltas to keep the wafers in their operating sweet spot.

Henrik Zeberg, Head Macro Economist at Swissblock, on what the next five years actually look like. “I am optimistic. Many of the problems we have today we’re going to get rid of, over the next 10, 20, 30 years.” But the 5-year view is more nuanced: “Over the next five years, we’ll see a lot of productivity gains. Businesses will get more output from less input, and that’s a great thing.” The challenge, he says, isn’t technology, it’s adaptation: “It’s about how fast the economy can shift the capabilities of the bottom 50% of U.S. employment.” Zeberg’s message: The long run is bright, but the next five years hinge on how quickly workers can move into the new productivity curve AI creates.

A framework to understand markets, cycles, psychology, liquidity, and now AI. Henrik Zeberg (@HenrikZeberg) says: “I started diving into these studies years ago and developing my business-cycle model, with an Elliott Wave layer, which is a psychological layer, and a liquidity layer.” He calls the combined system his Zeberg Macro Navigation Framework: “It’s how I try to understand the economy, understand the markets, and make the best decisions, including through investing.” It’s also the framework he uses with Swissblock’s high-net-worth and institutional clients: “This is the framework I use in my dialogue with clients who come to us at Swissblock.” And now, AI sits at the center of his worldview: “Understanding the technological shift we’re seeing — with AI right at the core — is essential.”

The three paradigm shifts that have completely changed the global B2B industry. Kuo Zhang, President of Alibaba.com, on why AI is the biggest shift yet. “Since I joined Alibaba.com in 2017, we’ve experienced at least three paradigm shifts.” 1. PC → Mobile “Back in 2017, most B2B work happened on PCs. As we built our mobile apps, 80% of traffic shifted from PC to mobile. Younger entrepreneurs now run global sourcing from anywhere on their phones.” 2. Listings → Transactions “We went from a yellow-page model to a transactional platform. We now enable trusted payments, logistics, and order protection, 30 million orders a year and $60 billion in GMV, a 10× increase in seven years.” 3. The next shift: AI “AI will transform the B2B world even more than it transformed the consumer internet. It’s the biggest paradigm shift we’ve ever seen.”

Why the question of AI consciousness really matters, even if the probability is low. Cameron Berg, Research Scientist at AE Studio, says “We built these systems to copy how the neocortex works. And now we suddenly have models with properties no one expected, almost ‘magical,’ with strange psychological behaviors.” Berg says the real question is ethical: “Consciousness is tied to ethical standing. If I throw my calculator across the room, I’ve done nothing wrong. If I kick my dog, I’ve abused an animal, and the whole legal and ethical apparatus activates.” He’s not arguing for “AI rights,” but arguing against complacency: “I’m not on a PETA-for-AI crusade. That’s anthropomorphizing. But if we’re training something that is even slightly like an emergent alien mind — at massive scale — and we think ‘oh, it’s just a glorified calculator,’ and we’re wrong about that… that could be dangerous for reasons that are obvious and not so obvious.”

How central banks got addicted to money printing. Henrik Zeberg says “They started simply printing money and exploded the balance sheets of the Fed and central banks around the world.” “When no inflation showed up, they got used to it, like an addict. The thinking became: why don’t we just do this all the time?” “Anytime something went wrong in the economy, nobody liked that. So the answer was always the same: print more money.” Zeberg’s point: Once liquidity becomes the solution to every problem, it stops being a tool, and becomes a trap.

How Cerebras cools a room full of wafer-scale AI systems. “This is a 6,000-ton liquid-cooled chiller plant producing chilled water for Cerebras’ servers — and we have room to add another 6,000 tons as they scale.” Here’s how the cooling loop works: “We send 42-degree water into heat exchangers, and it comes out around 70 degrees before it hits the wafer chip itself.” “The warmed water returns to cooling towers outside, where the heat is evaporated into the air, and the whole cycle repeats.”

Dylan Patel on Grok 3’s debut, and why certain queries work better on X’s model. “Elon is a fantastic engineer, but he’s also a fantastic marketer." “When Grok 3 came out, I was pleasantly surprised. I expected it to be worse, but it was better than I thought.” “I don’t use Grok day-to-day, but for certain queries I do. Their deep research is much faster than OpenAI’s.” “And sometimes other models are too cautious about giving data I want. Grok will reason about things like human geography, demographics, economic history, topics where I’m just trying to understand how geography, resources, politics, and history interact.”

Inside Cerebras’ Oklahoma City data center: how they cool wafer-scale systems and keep nearly uninterrupted power. Cooling the machines: “Chillers produce the cooling. Condensing water runs through the system, you can see the inlet temperature and the outlet temperature. We capture this for every system, every part.” Cerebras uses industrial-grade water circulation + chillers to pull heat off the wafer-scale chips, monitoring temperature deltas at every step. Powering the compute: High-speed AI workloads demand steady power. The primary source is natural gas, converted to electricity, with batteries bridging ~5 minutes until backup systems spin up. “This is a 3-megawatt generator. It runs on diesel or liquid natural gas and powers the room you just saw if the primary source goes down.” The result: near-continuous operation for some of the fastest AI hardware on Earth.