Skip to content

AI Technologies

A long read on how modern AI became possible: from history and scientific breakthroughs to large language models, agents and what comes next.

019 min read

What artificial intelligence is and why it works

Modern AI makes the most sense when described without mysticism. It is not a digital person and not a magical black box that suddenly woke up. It became practically useful because compute, data, architecture, and engineering practice finally aligned at scale. To understand current AI clearly, it helps to separate the philosophical question of intelligence from the engineering question of how large representational systems become broadly useful.

What intelligence means here

In AI, intelligence is best treated as the ability to build internal models, extract regularities, transfer structure across tasks, and generate useful actions or outputs in new situations. That is drier than popular mythology, but far more useful for serious understanding.

When a model writes code or explains an idea, it is not becoming human. It is showing that a learned representational system can compress large amounts of structure and reuse that structure across downstream tasks.

Why the breakthrough happened now

Neural-network ideas existed long before the current wave, but GPU infrastructure, large corpora, stable optimization, and mature engineering were not yet aligned. Once they did align, the same general family of methods began producing much broader practical capability.

This matters because major breakthroughs often come from several curves meeting at once rather than from one isolated theoretical insight.

Why next-token prediction becomes so powerful

To predict the next token well over a massive corpus, the model has to compress grammar, style, recurring concepts, and fragments of reasoning structure into its parameters. That is why a seemingly narrow objective can still produce broad practical behavior.

The usefulness does not emerge in spite of the simple objective. It emerges partly because a simple objective, applied at enormous scale, forces the system to internalize a rich structure of written knowledge and language.

Where the brain analogy stops helping

Brain comparisons can be intuitively helpful, but they are easy to overextend. Current LLM systems are not digital brains and should not be treated as one-to-one replicas of human cognition.

Their strength is still better described as large-scale statistical representation learning than as biological equivalence. That distinction keeps the discussion grounded.

Primary sources:LeCun, Bengio, Hinton · Deep LearningNIH · More than 100 trillion synapses

0212 min read

What an LLM is and how transformer architecture works

An LLM is best understood not as a magical autocomplete engine, but as a transformer-based computational core trained at massive scale. To really understand an LLM, you need to understand both the language-model side and the architectural side: tokenization, embeddings, self-attention, stacked transformer blocks, pretraining, alignment, and the external systems that turn the model into a practical tool.

Tokens, embeddings, and sequence structure

LLMs operate on tokens rather than directly on words or human ideas. A token can be part of a word, punctuation, a number, or a short text fragment. That matters conceptually because the model builds competence through statistical structure in token sequences rather than through hand-coded meanings.

Those tokens are turned into embeddings, dense vector representations, and combined with positional information so the model can distinguish what came earlier from what came later. From the beginning, then, the model is already a system for transforming discrete sequences into layered internal representations.

Why the transformer architecture matters

The decisive transformer idea is self-attention. Each token can dynamically attend to other tokens in the context and weigh which ones matter more for the current computation. That makes long-range dependence much easier to model than in earlier recurrent architectures.

In practice, a transformer is a repeated block of multi-head attention, feed-forward computation, residual connections, and normalization. Multi-head attention lets the model track different forms of structure at once, while the rest of the block stabilizes deep optimization. That combination is what made present large-scale language modeling practical.

Why Attention Is All You Need changed the field

Attention Is All You Need was not important only because it proposed a new architecture, but because it aligned model design with the realities of large-scale computation. By removing recurrence from the center of sequence modeling, it offered a cleaner path to parallelization, scale, and longer-context behavior.

That is why the paper changed the field so deeply. It did not merely improve one benchmark; it reset the default architectural language of modern AI. Much of the current LLM wave is downstream of that turn.

How transformer training becomes an LLM

During pretraining, the model learns to predict the next token over a huge corpus of text and code. That objective looks simple, but it forces the network to absorb grammar, recurring concepts, stylistic forms, and many regularities about how information is written and connected.

What emerges is not a literal database of facts, but a compressed representational structure that can be activated at inference time. This is why LLMs are often strong at synthesis, reformulation, and contextual reasoning, while still remaining imperfect at factual precision without verification.

Why post-training and system design still matter

A pretrained transformer is not yet a polished user-facing system. Post-training methods such as instruction tuning, preference optimization, and RLHF help shape the model into something that follows instructions, keeps useful formats, and behaves more reliably in practical interaction.

Even then, context limits, data quality, long error chains, and the need for external verification remain central constraints. That is why the strongest applications are usually broader systems around the model, not the model in isolation.

Primary sources:Attention Is All You NeedGPT-3 paper

0310 min read

From AI winters to AlexNet and Geoffrey Hinton

The current AI wave only makes sense when placed inside a longer history of symbolic systems, hardware limits, scientific persistence, and repeated cycles of disappointment. Geoffrey Hinton matters in that story not only because of a few famous ideas, but because he helped carry the deep-learning line through the years when it still looked too early and too costly to dominate.

Why early AI stalled

Rule-based systems provided structure and control, but they scaled poorly to messy real-world data filled with ambiguity, noise, and exceptions. Symbolic approaches were elegant, but they struggled to absorb language, vision, and behavior at scale.

That helps explain the AI winters: the field repeatedly moved through cycles of overpromising and retrenchment, which shaped both funding and public trust.

Why Geoffrey Hinton mattered

Hinton mattered because he kept working on neural-network learning when much of the field treated it as too early, too expensive, or too unstable. His role was not only to produce ideas, but to hold a direction long enough for the rest of the world to catch up to it.

That persistence matters historically. Scientific shifts often depend not only on correctness, but on whether someone carries a line of thought through the years when it still looks strategically inconvenient.

Why deep learning looked premature for so long

Backpropagation and deep learning were not ignored simply because people failed to understand them. For a long time they really were difficult to exploit at useful scale. The ecosystem lacked the hardware, datasets, and engineering maturity needed to unlock them.

That is a broader lesson: methods can look weak not because they are wrong, but because the surrounding infrastructure is not ready to reveal their strength.

What AlexNet made undeniable

AlexNet did more than win a benchmark. It showed that a deep network trained on large data with GPU acceleration could shift the quality frontier abruptly enough to change the center of gravity of the field.

After AlexNet, the debate moved from whether deep learning worked at all to how far and how fast it could be scaled. That is why Geoffrey Hinton and AlexNet belong in the same frame: one represents the long intellectual line, the other the moment of unmistakable public proof.

Primary sources:Deep Learning · NatureImageNet Classification with Deep CNNs · Abstract

048 min read

DeepMind and the scientific turn

DeepMind’s line matters because it turned AI from game performance into a broader story about planning, learning in complex environments, and eventually scientific discovery.

Games as laboratories

AlphaGo and AlphaStar showed that AI could plan in settings where brute force alone was not enough. Games provided a compact world for strategy, reward, and long-horizon behavior.

From games to protein folding

AlphaFold changed the public meaning of AI by turning it into a scientific instrument. That shift helped move AI from a technology story into the core of knowledge production.

Primary sources:About Google DeepMindAlphaGoAlphaStarAlphaFoldDemis Hassabis & John Jumper awarded Nobel Prize in Chemistry

058 min read

OpenAI and the scaling era

Scaling turned from a brute-force instinct into a methodology. Once models, data, and compute were expanded in the right way, capability shifts became large enough to reorganize product design.

GPT-3 and few-shot behavior

GPT-3 showed that large models could perform many tasks from instructions and examples alone. That changed the interface model from task-specific tuning toward prompt- and context-based programming.

Chinchilla and compute-optimal scaling

Chinchilla made the economics of scaling sharper by showing that model size and training tokens have to grow in a more balanced way. That was a major correction to naïve bigger-is-better thinking.

Primary sources:GPT-3 · Language Models are Few-Shot LearnersChinchilla scaling laws

068 min read

Agents, tools, and MCP

The major shift ahead is not only stronger models, but models becoming agents: systems that can reason, act, use tools, and continue work across multiple steps.

Reasoning plus action

ReAct demonstrated that reasoning and acting reinforce each other. Plans help guide actions, while external actions reduce hallucination by bringing in new evidence and feedback.

Why protocols matter

As tool ecosystems grow, structured protocols like MCP become essential. They let models discover capabilities, invoke tools, and receive predictable structured responses at scale.

Primary sources:ReActIntroducing the Model Context Protocol

077 min read

AGI and superintelligence without mythology

AGI is more useful as a framework for comparing capability, breadth, and autonomy than as a mystical threshold. Superintelligence becomes meaningful when the discussion shifts into deployment, control, and governance.

Why AGI is hard to define

Intelligence does not collapse into one benchmark. Systems vary across language, planning, transfer, and autonomy, which is why levels and ontologies are more useful than single slogans.

Why governance becomes central

As capability and autonomy grow, deployment questions become more important: access, control, military use, concentration of compute, and international coordination all enter the picture.

Primary sources:Levels of AGISuperintelligence · Oxford Martin School

088 min read

The near future: multimodality, robotics, and AI factories

The next AI phase is likely to be defined not by one best model, but by a restructuring of the full stack: multimodal systems, robots, simulation, scientific tools, and increasingly industrial compute infrastructure.

Multimodality and action

As models work across text, vision, audio, interfaces, and tools, they stop being purely linguistic systems. That pushes AI from answer generation toward a more general task-execution layer.

Infrastructure and factories

The most strategic competition is shifting toward chips, racks, networking, energy, and AI factory design. That infrastructure determines the real cost and scale of deployment.

Primary sources:Orbital data centers framework