Learning – Stacking Trades

The Fall of the Giants

Stacking Trades — Mon, 08 Dec 2025 21:01:21 +0000

In late 2025, it has become quite common for the most impressive new AI features to come first from companies that many still call “startups.” These features are then released later, redesigned and repackaged, on platforms that serve billions of users. The shift is not that Big Tech has stopped developing technology. Google is still advancing systems like Gemini 3 in ways only Google can, including in Search.

The change is that the center of gravity for model iteration has shifted. The companies moving fastest have built their entire operating system around one loop: train, evaluate, deploy, learn, repeat. When the product is the model, and the customer is a developer, the distance from breakthrough to shipping is short.

The new AI-native cadence

Startups are not “winning” because they have a secret ingredient that giants cannot buy. They are winning because their feedback cycles are tighter, and their constraints are more legible.

When Anthropic releases a new frontier model like Claude Opus 4.5, it can quickly connect it with the distribution that matters most to its main audience: developers and teams who write code for their jobs, using the tools they already work with. That’s why adding Claude Code to Slack is not just a minor addition. It is central to the offering.

Mistral’s recent push tells a similar story from the open-weight angle: release models, let the ecosystem pressure test them, and win mindshare by making adoption easy for builders who want control.

“Speed is not a personality trait, it’s an organizational design choice.”

Why small teams ship faster than large empires

In Big Tech, delivering a capability is seldom just about delivering a capability. It involves passing privacy reviews, managing brand risk, facing litigation risks, undergoing safety audits, monitoring for abuse, meeting localization and accessibility requirements, negotiating with partners, and dealing with the sheer size of the product. Even when a model is ready, the distribution process has its own challenges.

By contrast, AI-first startups can choose narrow front doors. They can ship to a smaller set of customers, watch what breaks, patch, and ship again. They can decide that “developer happiness” is the main KPI for the quarter, because it often is.

Even OpenAI’s own internal “code red” to refocus on improving ChatGPT shows how competitive pressure now punishes diffusion. When the field moves week to week, a roadmap that made sense six months ago can become a liability.

The distribution tax of being huge

Big Tech still has distribution that startups cannot replicate, and that advantage is real. When Google puts Gemini 3 into Search experiences on day one, it can reshape how hundreds of millions of people encounter information without requiring a new habit or a new app.

That same scale also creates a tax. Every new capability must behave safely across edge cases that only appear at massive volume. Every change has reputational blast radius. A startup can afford to be wrong in a contained way. A giant often cannot.

This is why “the giants are falling” is the wrong mental model. The better model is that giants are becoming systems integrators, while AI-native companies act like high-frequency traders for capability. One discovers and packages speed. The other turns that speed into something stable enough to deploy everywhere.

Regulation turned speed into strategy

Another accelerant is governance. The EU’s AI Act era has begun to bite in practical ways, including obligations for general-purpose AI providers and the emergence of compliance mechanisms like the General-Purpose AI Code of Practice published in July 2025.

Large incumbents tend to internalize these requirements earlier and more broadly, because they already operate under intense regulatory scrutiny and because they have more to lose. Startups can still take on risk, but many do it selectively: they ship into developer channels, enterprise sandboxes, and opt-in workflows where safety controls are easier to enforce than on a default consumer surface.

The market consequence: capability is commoditizing, velocity is not

If you zoom out, a pattern emerges. Model capability keeps rising across the board, and the frontier is crowded. In that world, differentiation shifts toward iteration speed, tooling, and the practical ergonomics of using AI to do real work.

That is why the most meaningful “breakthroughs” are increasingly bundled as workflows: agentic systems that automate internal business processes, and model releases that arrive with practical migration paths, evaluation harnesses, and lifecycle planning instead of hype.

The giants are not out of the race. They are just running a different one. Startups are sprinting on tight loops to capture the next few weeks of developer attention. Big Tech is hauling those capabilities across the last mile of scale, compliance, and everyday usefulness.

In 2025, the interesting question is not who builds the smartest model. It is who can ship the future, repeatedly, without breaking the present.

AI Is Not One Technology

Stacking Trades — Mon, 01 Dec 2025 20:22:40 +0000

The Illusion of a Single System

Most people discuss artificial intelligence as if it were one thing. A model. A brain. A system that can understand and respond. But the reality within companies is very different. AI today is not a single technology. It is a layered structure of tools, pipelines, memory systems, evaluators, guardrails, and agents, with each part performing a different function.

What the outside world perceives as one answer is often the result of many components collaborating behind the scenes. The model is just one part of this complex system.

The Model Is the Interface, Not the Machine

The visible part of AI is the model that produces text, analyzes images, or suggests actions. It feels like the core engine because it is the part that interacts with us. But the engine depends on the layers beneath it.

A retrieval system can provide context from documents. A data pipeline ensures that context stays updated. A memory layer keeps long-term patterns. A tool invoking layer determines when to call external systems. An evaluator checks if the model followed the rules. A monitoring system tracks drift and failure cases. A security layer filters out harmful or non-compliant requests.

Remove any one of these layers and the model behaves unpredictably. This is why enterprises deploying AI at scale talk less about models and more about architecture.

“The intelligence we see from AI does not come from a single model. It comes from the architecture surrounding it.”

How Companies Actually Build AI Workflows

Inside organizations, a single AI task often starts a small network of cooperating parts. A planning part breaks down the request. A reasoning part drafts an approach. A retrieval part gathers context. A synthesis part produces an answer. A verification part checks the constraints. A scoring part measures reliability.

Companies like Microsoft, Google, and Anthropic have published research showing that multi component systems consistently outperform single model setups. Stanford’s 2024 AI Index documented the same trend across enterprise deployments. Coordination beats scale.

The intelligence we perceive is not coming from a lone model. It is emerging from the way these systems interact.

Why This Matters for Understanding AI’s Limits

Treating AI as a single technology creates unrealistic expectations. People imagine the model should know everything, remember everything, and decide everything. But the strengths of AI do not come from omniscience. They come from orchestration.

When a model hallucinates, it is often because it lacks a retrieval layer. When it forgets context, it is due to the absence of a memory module. When it struggles with long tasks, it is because there was no planning scaffold. When it contradicts itself, it is because it is missing evaluation.

Most weaknesses in AI systems come from absent pieces of the stack, not from the model itself.

The Coming Shift in How People Build and Evaluate AI

As AI becomes embedded in critical operations, companies will evaluate systems the way they evaluate software infrastructure, not user facing apps. They will talk about reliability, latency, tooling, routing, and guardrails before they talk about model size.

The real breakthroughs will come not from bigger models but from better architectures. These include multi-agent systems, smarter memory, structured reasoning, automated evaluation, domain-specific tool use, and secure data retrieval.

In the same way that operating systems defined the PC era and cloud infrastructure defined the software era, the emerging AI stack will define the intelligence era.

The future belongs to the builders who understand that AI is not a brain. It is an ecosystem.

The New Billion-Dollar Job

Stacking Trades — Wed, 26 Nov 2025 20:49:39 +0000

From Labels to Logic

For years, the main jobs related to AI seemed unexciting. Data labeling, annotation, and tagging were the tasks at hand. Workers drew boxes around objects in images, marked sentiment in text, or checked answers for accuracy. It was repetitive work, but it was important. Models required examples, and people provided them.

That era is still with us, but something new is emerging on top of it. As models gain raw power, the bottleneck is shifting. The hardest problem is no longer teaching systems what a cat looks like. It is teaching them how to decide, how to reason, and how to follow norms that resemble judgment rather than pattern matching.

A new class of work is forming around that challenge. Not labeling. Not simple prompting. Something closer to instructing machine minds in how to think.

The Limits of Raw Scale

Over the last few years, large language models have improved thanks to more parameters, more data, and more computing power. However, even the most advanced systems still show familiar weaknesses. They make up information. They contradict themselves. They have a hard time with multi-step reasoning even when they seem confident. Benchmarks from research groups at Stanford, Berkeley, and other institutions repeatedly highlight gaps in logical consistency, planning, and reliable tool use, despite fast improvements in basic performance.

Scaling has brought us to a new plateau. More data and more GPUs move the ceiling, but they do not change the fact that the models are learning correlations, not principles. Organizations can no longer assume that throwing more tokens at the problem will yield better judgment.

This is where the new job appears.

Teaching Frameworks Instead of Answers

Inside AI labs and companies that deploy models at scale, people are starting to work less as annotators and more as curriculum designers. They do not just identify the correct answer. They define what a clear chain of thought looks like. They specify which tools a model should use and the order in which to use them. They write policies that describe acceptable reasoning paths and those that are unacceptable. They build scaffolds.

Some of this shows up publicly in research on tool using models and reasoning agents. System prompts now include detailed instructions about steps, constraints, and evaluation criteria. Teams design synthetic tasks where models practice decomposing problems rather than jumping to a guess. In more advanced settings, models are trained or fine tuned on traces of their own reasoning, corrected and curated by humans who act less like graders and more like tutors.

The work is not about discrete labels. It is about teaching structure.

The People Behind the Structure

This role does not fit neatly into old titles. It blends parts of machine learning, product thinking, behavioral design, and even a little philosophy.

Some people doing this today have titles like AI researcher, alignment engineer, or reasoning specialist. Others work in product teams but spend much of their time designing evaluation frameworks, system instructions, and feedback loops for agents instead of users. They select which examples to show models. They determine how to phrase objectives. They establish what counts as a solid solution.

They are not writing traditional software, but they are programming behavior. The medium is not code. It is thought.

As more companies rely on AI for complex decisions, this role starts to matter as much as traditional engineering. A model can draft a hundred options. It still takes a human, working at the framework level, to decide what the model should value.

“The most valuable AI work is shifting from giving models answers to teaching them how to think.”

Why This Becomes a Billion-Dollar Job

This is not about salaries alone. It is about leverage.

If one person improves the reasoning framework of a widely used model, that improvement affects every user, every workflow, and every integration. A single insight on how to shape a decision can spread through a system that serves millions. The economic effect of that work far surpasses the effect of adding another app or feature.

Companies are already signaling this. Job postings for roles focused on model behavior, evaluation, and policy design have grown significantly in the last two years. Investors increasingly ask frontier model companies about alignment, reliability, and governance, not only benchmark scores. Enterprises deploying AI in finance, healthcare, and logistics want to know who is responsible for how the models think, not just how fast they run.

The people who can shape that thinking are quietly becoming some of the most leveraged individuals in the stack.

Beyond Prompting

It is tempting to confuse this with prompt engineering. At the beginning of the generative wave, prompt engineering looked like a cheat code. Clever phrasing, special tokens, and long detailed instructions could produce surprisingly good results. But as models and tooling have matured, the focus has shifted.

The new work focuses less on clever lines and more on repeatable systems. It includes test suites, scenario libraries, failure catalogs, and structured rubrics that define good reasoning in a domain. It involves working together with legal, risk, and domain experts. It treats model behavior as something that can be directed using frameworks rather than relying on one-off hacks.

Prompting is to this work what a single lesson is to an entire curriculum.

The Next Layer of Responsibility

There is also a deeper responsibility here. Training models how to think is not neutral. Choices about what counts as valid reasoning, what risks are acceptable, and which trade offs matter are all value laden. They reflect the priorities of the organizations building the systems.

That is why this emerging role is not only technical; it involves governance. The people who create the reasoning frameworks of powerful models will impact decision-making in areas that involve money, health, safety, and information. They are not just building tools; they are shaping the standards of machine judgment.

As AI systems become more autonomous and more embedded in critical processes, this influence will only grow.

A Job That Did Not Exist Ten Years Ago

Ten years ago, there was no such job. There were machine learning engineers, data scientists, and research scientists. There were product managers and architects. There were compliance officers and domain experts.

Today, we are beginning to see something new at their intersection. People whose primary work is to teach machines not what to think, but how.

It is difficult to measure this role using traditional categories. It will not appear as a separate item in standard org charts for long. However, as AI progresses from pattern matching to something resembling reasoning, the people who influence that reasoning will become some of the most crucial builders in the stack.

The new billion-dollar job is not labeled as such yet. It lives under different titles and in different departments. But its shape is already clear. Somewhere between engineering and instruction, between governance and design, people have started training AIs how to think.

PromptOps

Stacking Trades — Mon, 17 Nov 2025 22:18:24 +0000

The End of Improvised Prompting

There was a time when prompting felt spontaneous. People shared clever inputs in chats, exchanged tricks on social media, and treated AI models like tools that responded to instinct instead of a set format. The idea was straightforward. Anyone could write a prompt. You simply typed your way to a result.

Inside companies that now rely on AI at scale, that world is gone. Prompting is no longer a creative stunt. It is becoming part of the machinery that runs the system. Prompts now live in registries, carry version histories, and undergo reviews the way software components do. What once felt playful is becoming procedural.

When Cost Forced Structure

The shift did not happen for style. It happened because prompting impacts money. The 2024 Stanford AI Index Report highlighted the rising cost of inference as models expanded and usage increased. A poorly written prompt uses more tokens, generates longer outputs, and leads to more retries. Across millions of calls, these small inefficiencies turn into significant expenses.

Companies began treating prompt structure as a lever for managing compute. A precise prompt reduces load. A clear prompt improves accuracy. A tested prompt lowers the chance of unexpected behavior. The input became as important as the output.

Compliance Makes the Input Matter

Regulation accelerated this trend. The EU AI Act, finalized in 2024, requires organizations to show how their AI systems reach decisions. That means logging not only the results but the instructions that shaped them. Prompts stored in managed libraries became part of compliance records. Legal teams began reviewing prompts for risk. Inputs shifted from casual requests to documented artifacts.

A New Workflow Around Prompts

Inside organizations, prompts became shared tools instead of personal projects. A workflow engineer builds the structure. A product owner defines the purpose. A legal reviewer examines the language. A data scientist evaluates performance using real benchmarks. What used to be a single line of text is now part of a team development process.

Prompts behave like specifications. They define how the system should think, respond, or reason within a specific task. They carry accountability.

“Prompting is becoming an operational language inside AI systems, not a creative trick.”

The Quiet Infrastructure Layer

Something subtle is happening underneath all of this. Prompting is becoming an operational language for directing machines. It shapes reasoning paths, influences compute load, and governs how AI systems behave as they scale.

PromptOps is not loud. It does not come with launches or marketing campaigns. It spreads through tools, policies, and the need for systems that can be trusted. But it marks the beginning of a new phase in AI. The frontier is not only in what models can do. It is in how precisely we can tell them what to do.

Marketing in the Age of Infinite Tests

Stacking Trades — Thu, 06 Nov 2025 20:00:11 +0000

For most of marketing’s modern history, optimization was an exercise in patience. Two versions of an ad — A and B — would battle it out across an audience sample. A winner would emerge, a report would be written, and the campaign would adjust. That flow of rhythm defined digital advertising for twenty years.

Now the rhythm is unrecognizable. AI doesn’t test campaigns sequentially; it tests them continuously. Every impression becomes an experiment, every click a data point in an evolving system that refines itself faster than humans can brief it.

We’ve entered an age where the experiment and the campaign has become one.

The Death of the Static Test

A/B testing was built for human limits — two variables, producing a clear result.

AI operates beyond that boundary. Using reinforcement learning, generative targeting, and contextual optimization, systems can run millions of micro-variations at once: color shades, phrasing swaps, emotional tone, even pacing in video ads.

Google’s Performance Max and Meta’s Advantage+ already do this at scale. They don’t compare A vs B — they orchestrate entire universes of versions, allocating spend dynamically as results emerge.

Creativity as Code

This shift is forcing marketers to think differently about creativity.

Instead of launching a single polished campaign, brands now deploy creative genomes — modular assets designed to evolve under algorithmic control.

At Wieden + Kennedy’s experimental lab, an AI copy system generated 12,000 variants of a tagline for a sports client, continuously rewriting itself based on regional engagement sentiment.

By week’s end, the line most consumers remembered had never been written by a human.

AI doesn’t brainstorm; it iterates. And iteration at machine scale changes what “good” even means.

From Optimization to Autonomy

The new question isn’t which version wins — it’s who decides.

As models learn to self-optimize, marketers risk losing sight as to why certain patterns succeed.

Amazon’s retail media algorithms, for example, already generate and retire ad variations within hours, sometimes before human teams even see the creatives. The result: higher ROI, but vanishing rationale.

Some agencies now employ “AI auditors,” data scientists who reverse-engineer why an algorithm favored one audience-message pair over another — a new discipline that sits between strategy and surveillance.

“Our job used to be taste. Now it’s traceability.”

A Market of One, Billions of Times Over

Infinite testing also means infinite personalization.

Where A/B testing sought a single best message, AI seeks a unique message for each person at each moment.

Netflix trailers, Spotify promos, and TikTok ads are already morphing based on micro-behaviors: pause length, skip velocity, or scroll rhythm. These signals feed neural networks that learn individual persuasion tempos — the precise moment when curiosity converts to intent.

Marketing no longer targets demographics; it synchronizes with cognition.

When Everything Works, Nothing Learns

But there’s a paradox.

When every message is optimized instantly, learning can flatten. Brands risk losing the narrative coherence that used to emerge from slower cycles of testing and reflection.

Several large CPG advertisers quietly report that hyper-optimization increases short-term metrics while long-term brand recall fades in the background. The signal becomes so fine-tuned it disappears into noise.

AI can win every micro-moment and still lose the story.

When One Brain Isn’t Enough

Stacking Trades — Wed, 05 Nov 2025 21:19:26 +0000

AI-Enhanced Humans and the Rise of Cognitive Companions

There has always been a quiet hierarchy in knowledge work — those who could think faster, absorb more, connect dots invisible to others. For decades, the edge came from memory, intuition, or the rare ability to hold messy ideas steady until clarity surfaced.

Now, that advantage is changing hands.

Not because humans suddenly forgot how to reason — but because the tools beside them have evolved and learned how to help. There is a new class of professionals emerging, not defined by job title or pedigree, but by how seamlessly they work alongside intelligent systems.

They aren’t “using AI” the way early computer users “used software.”

They co-think with it.

The machine drafts, they shape.

The machine surfaces patterns, they interpret.

The machine remembers everything, they decide what matters.

And the gap between those who work alone and those who work with a cognitive companion is widening faster than industries are prepared to admit.

The New Advantage: Judgment + Inference

Early automation replaced repetitive labor.

This wave does something different — it amplifies mental capacity.

Not by answering questions alone, but by increasing the quality of the questions asked.

In meetings, these augmented workers don’t say “Let me research that.”

They ask richer questions in real time because they’re never thinking with one brain.

They have:

• Memory that never decays

• Research that runs while they sleep

• Feedback without embarrassment

• Perspective that compounds instead of erodes

Their work isn’t faster.

It’s deeper.

And depth, in an age of noise, is leverage.

The First Cognitive Super-Teams

These enhanced workers aren’t defined by tools, but by habits:

• They think with their models, not after them

• They build personal knowledge systems that learn alongside them

• They treat intuition as input, not gospel

• They interrogate assumptions with machine feedback

Some keep their companion in the background.

Others speak openly about the model “sitting beside them.”

Neither is right or wrong.

It is simply a new mode of work — one where intelligence is no longer solitary.

You don’t win by knowing more.

You win by processing more efficiently.

What Changes Now

Industries built on memory, recall, and pattern recognition will shift first:

• Law shifts from precedent recall → interpretation at scale

• Strategy shifts from instinct → scenario synthesis

• Medicine shifts from manual diagnosis → augmented reasoning

• Finance shifts from “feel” → behavioral inference assisted by machines

• Education shifts from teaching facts → training judgment

This isn’t replacement.

It’s reinforcement — the cognitive equivalent of an exoskeleton.

The jobs that disappear won’t vanish because AI is better.

They’ll vanish because augmented humans outperform unaugmented ones.

“Intelligence is no longer solitary. It’s shared, amplified, and scaffolding our best thinking.”

A Cultural Shift, Not a Technical One

For many years, intelligence was viewed as an inherent trait — unchanging, much like one’s height.

This era breaks that assumption.

Intelligence becomes interactive.

Elastic. Scaffolded. Shareable.

When combined with a disciplined AI workflow, a person with average talent will do better than a person with exceptional talent working alone.

This changes more than productivity.

It changes power, confidence, and identity.

Some will see this as competition.

Others will see it as companionship.

Only one group gets left behind.

The Human Part Still Matters

There are things models still cannot do:

• Care

• Choose values

• Hold moral weight

• Understand stakes deeply enough to lose sleep

• Feel responsibility in the bones instead of the logs

Machines surface possibilities.

Humans choose the ones that are worthwhile.

In that selection, the essence of work — and humanity — remains.

The augmented don’t surrender intuition.

They refine it.

The future does not belong to machines or to humans alone.

It belongs to those who learn to think in chorus.