Latency isn't a flaw — it's a feature of intelligence — and possibly the key to reclaiming our power in the markets.

Originally published March 7, 2025 on Medium

"A particularly intriguing phenomenon observed during the training of DeepSeek-R1-Zero is the occurrence of an 'aha moment'… a captivating example of how reinforcement learning can lead to unexpected and sophisticated outcomes." — DeepSeek AI

Thomas Edison was obsessed with the hidden intelligence of the subconscious. He believed that breakthroughs happen in the in-between spaces — just beyond conscious grasp. To capture these fleeting moments of insight, he devised an unusual technique: he would sit in a chair holding a heavy metal ball in his hand. As he drifted off, his grip would loosen, the ball would drop, and the sudden noise would wake him — right at the threshold between wakefulness and the subconscious.

He wasn't alone. Salvador Dalí used a key on a plate instead, calling this practice "slumber with a key." Einstein strategically embraced micro-naps, harnessing subconscious processing for profound insights. They all understood something fundamental: intelligence isn't just about direct problem-solving — it's about what happens in the shadows of the mind, where ideas incubate before surfacing. In other words, real intuition isn't just about what you know, but how you process what you don't yet know.

And now, AI might be learning this too.

Is this the birth of AI self-awareness?

I was deep in thought about the relationship between AI, attention, and human cognition. I realized that the Transformer's self-attention mechanism — pioneering the leap toward GPT-3 — mirrored the stages of the creative process.

I called it active incubation — the idea that breakthroughs aren't always about pushing forward, but about holding space for ideas to form. Creativity isn't just about immediate inspiration but about letting thought processes evolve in the background before something finally clicks.

Two years later, DeepSeek's AI breakthroughs led me back to this idea — but on an entirely new scale. AI, too, is shifting away from instant pattern recognition toward something more human: latent reasoning.

"A particularly intriguing phenomenon observed during the training of DeepSeek-R1-Zero is the occurrence of an 'aha moment.' This moment occurs in an intermediate version of the model. During this phase, DeepSeek-R1-Zero learns to allocate more thinking time to a problem by reevaluating its initial approach. This behavior is not only a testament to the model's growing reasoning abilities but also a captivating example of how reinforcement learning can lead to unexpected and sophisticated outcomes." — DeepSeek AI

But something even more profound is happening. DeepSeek-R1-Zero isn't just solving problems — it's recognizing when it's learning in real-time. Is this the birth of AI self-awareness?

DeepSeek-R1-Zero has shown a phenomenon resembling metacognition — or, at the very least, the early sparks of self-monitoring intelligence. In its intermediate training phase, the model demonstrated an 'aha moment' that wasn't just an improvement in reasoning — it was the recognition that it needed to rethink its approach.

"Wait, wait. Wait. That's an aha moment I can flag here." — DeepSeek-R1-Zero

This is a major shift in AI cognition. Instead of blindly following programmed heuristics, DeepSeek's model is developing an awareness of when its own assumptions are flawed.

"What's really impressive is DeepSeek models' ability to reason," says Kush Varshney, an IBM Fellow. Reasoning models essentially verify or check themselves, representing a type of "metacognition," or "thinking about thinking," Varshney says. "We are now starting to put wisdom into these models, and that's a huge step." (IBM)

A narrative in finance that's been picking up steam lately — particularly in quantitative and algorithmic trading which is heavily dominated by institutions — is that speed is everything, and that only those with the fastest data feeds, lowest latency, and highest-frequency algorithms can compete.

Institutions use co-located servers next to exchanges to execute trades in microseconds. Quant firms rely on cutting-edge Nvidia chips to process vast amounts of market data before anyone else.

Retail traders are often left behind, unable to compete on execution speed alone.

But what if speed isn't the ultimate edge? What if the future of trading lies in how well you can extract meaning from latency, rather than simply reducing it?

The greatest traders aren't just fast — they intuit shifts in market behavior before they happen.

A regime shift is when the fundamental structure of a market changes — bull to bear, volatility spikes, liquidity shifts, or macroeconomic pivots.

These shifts often defy past correlations — what worked yesterday no longer applies. Traditional AI and quantitative algorithms trained on historical statistical patterns struggle here.

An AI that can detect its own learning threshold — like DeepSeek-R1-Zero — could do something markets have never seen before: recognize when it's in a new regime before human traders do — not through speed, but through latent reasoning.

DeepSeek's metacognition — or ability to 'pause' and recognize when it needs to rethink its approach — is the foundation for this.

Latency isn't a flaw — it's a feature of intelligence — and possibly the key to reclaiming our power in the markets.

Originally published March 7, 2025 on Medium

"A particularly intriguing phenomenon observed during the training of DeepSeek-R1-Zero is the occurrence of an 'aha moment'… a captivating example of how reinforcement learning can lead to unexpected and sophisticated outcomes." — DeepSeek AI

Thomas Edison was obsessed with the hidden intelligence of the subconscious. He believed that breakthroughs happen in the in-between spaces — just beyond conscious grasp. To capture these fleeting moments of insight, he devised an unusual technique: he would sit in a chair holding a heavy metal ball in his hand. As he drifted off, his grip would loosen, the ball would drop, and the sudden noise would wake him — right at the threshold between wakefulness and the subconscious.

He wasn't alone. Salvador Dalí used a key on a plate instead, calling this practice "slumber with a key." Einstein strategically embraced micro-naps, harnessing subconscious processing for profound insights. They all understood something fundamental: intelligence isn't just about direct problem-solving — it's about what happens in the shadows of the mind, where ideas incubate before surfacing. In other words, real intuition isn't just about what you know, but how you process what you don't yet know.

And now, AI might be learning this too.

Is this the birth of AI self-awareness?

I was deep in thought about the relationship between AI, attention, and human cognition. I realized that the Transformer's self-attention mechanism — pioneering the leap toward GPT-3 — mirrored the stages of the creative process.

I called it active incubation — the idea that breakthroughs aren't always about pushing forward, but about holding space for ideas to form. Creativity isn't just about immediate inspiration but about letting thought processes evolve in the background before something finally clicks.

Two years later, DeepSeek's AI breakthroughs led me back to this idea — but on an entirely new scale. AI, too, is shifting away from instant pattern recognition toward something more human: latent reasoning.

"A particularly intriguing phenomenon observed during the training of DeepSeek-R1-Zero is the occurrence of an 'aha moment.' This moment occurs in an intermediate version of the model. During this phase, DeepSeek-R1-Zero learns to allocate more thinking time to a problem by reevaluating its initial approach. This behavior is not only a testament to the model's growing reasoning abilities but also a captivating example of how reinforcement learning can lead to unexpected and sophisticated outcomes." — DeepSeek AI

But something even more profound is happening. DeepSeek-R1-Zero isn't just solving problems — it's recognizing when it's learning in real-time. Is this the birth of AI self-awareness?

DeepSeek-R1-Zero has shown a phenomenon resembling metacognition — or, at the very least, the early sparks of self-monitoring intelligence. In its intermediate training phase, the model demonstrated an 'aha moment' that wasn't just an improvement in reasoning — it was the recognition that it needed to rethink its approach.

"Wait, wait. Wait. That's an aha moment I can flag here." — DeepSeek-R1-Zero

This is a major shift in AI cognition. Instead of blindly following programmed heuristics, DeepSeek's model is developing an awareness of when its own assumptions are flawed.

"What's really impressive is DeepSeek models' ability to reason," says Kush Varshney, an IBM Fellow. Reasoning models essentially verify or check themselves, representing a type of "metacognition," or "thinking about thinking," Varshney says. "We are now starting to put wisdom into these models, and that's a huge step." (IBM)

A narrative in finance that's been picking up steam lately — particularly in quantitative and algorithmic trading which is heavily dominated by institutions — is that speed is everything, and that only those with the fastest data feeds, lowest latency, and highest-frequency algorithms can compete.

Institutions use co-located servers next to exchanges to execute trades in microseconds. Quant firms rely on cutting-edge Nvidia chips to process vast amounts of market data before anyone else.

Retail traders are often left behind, unable to compete on execution speed alone.

But what if speed isn't the ultimate edge? What if the future of trading lies in how well you can extract meaning from latency, rather than simply reducing it?

The greatest traders aren't just fast — they intuit shifts in market behavior before they happen.

A regime shift is when the fundamental structure of a market changes — bull to bear, volatility spikes, liquidity shifts, or macroeconomic pivots.

These shifts often defy past correlations — what worked yesterday no longer applies. Traditional AI and quantitative algorithms trained on historical statistical patterns struggle here.

An AI that can detect its own learning threshold — like DeepSeek-R1-Zero — could do something markets have never seen before: recognize when it's in a new regime before human traders do — not through speed, but through latent reasoning.

DeepSeek's metacognition — or ability to 'pause' and recognize when it needs to rethink its approach — is the foundation for this.