Course · 15 modules · 105 lessons · 849 min

LLM Evolution

The history and trajectory of large language models — from pre-transformer foundations through the 2025 frontier.

← All courses
Your progress0 / 105 lessons· 0%

The course at a glance

15 modules · click any tile to jump to its lessons.

All lessons

№ 01Pre Transformer Foundations
01Word Embeddings: Word2Vec and GloVeWord2Vec, GloVe, and FastText gave words numerical meaning by learning dense vector representations from massive text corpora, establishing the distributional foundation for all modern NLP.7 min→02Recurrent Neural Networks and LSTMsRNNs processed language one token at a time like reading left to right, and LSTMs solved their crippling memory problem with learned gates — dominating NLP from 2014 to 2017 before the Transformer made their sequential bottleneck obsolete.7 min→03Sequence-to-Sequence ModelsThe Seq2Seq framework (Sutskever et al., 2014) established the encoder-decoder paradigm for mapping variable-length inputs to variable-length outputs, achieving breakthrough machine translation results while revealing the fixed-length bottleneck that would drive the invention of attention.7 min→04Attention Mechanism OriginsBahdanau attention (2014) let decoders dynamically focus on different parts of the input sequence, solving the fixed-length bottleneck of Seq2Seq and laying the conceptual foundation for the Transformer's self-attention.7 min→05ELMo and Contextual EmbeddingsELMo (Peters et al., 2018) demonstrated that deep bidirectional LSTMs pre-trained on language modeling could generate context-dependent word representations, breaking the static embedding paradigm and pioneering the pre-train-then-fine-tune approach.7 min→06ULMFiT and Transfer Learning for NLPULMFiT (Howard & Ruder, 2018) demonstrated that a three-stage transfer learning recipe — pre-train a language model, fine-tune it on domain text, then fine-tune on the task — could match or beat state-of-the-art NLP systems trained from scratch, establishing the methodology that GPT and BERT would scale to transformative effect.8 min→07The Bottlenecks That Motivated TransformersThree fundamental limitations of RNN-based NLP — sequential computation preventing parallelism, vanishing gradients limiting memory, and fixed-length bottleneck vectors losing information — created an urgent need for a fully parallel architecture, setting the stage for the Transformer.8 min→
№ 02The Transformer Revolution
01Attention Is All You NeedVaswani et al. (2017) introduced the Transformer — a fully parallel architecture based entirely on self-attention that eliminated recurrence, achieved 28.4 BLEU on English-German translation in 3.5 days on 8 GPUs, and became the foundational architecture for every major language model that followed.8 min→02GPT-1: Generative Pre-TrainingGPT-1 (Radford et al., 2018) combined a decoder-only Transformer with unsupervised generative pre-training followed by supervised fine-tuning, establishing the paradigm that decoder-only models trained on next-token prediction could develop broad language understanding.7 min→03BERT: Bidirectional Encoder Representations from TransformersBERT (Devlin et al., 2018) introduced masked language modeling and bidirectional pre-training with an encoder-only Transformer, achieving state-of-the-art results on 11 NLP tasks and triggering the "BERT-ification" of the entire field — the most influential NLP paper since the Transformer itself.8 min→04GPT-2: Language Models Are Unsupervised Multitask LearnersGPT-2 (Radford et al., 2019) scaled the GPT-1 architecture to 1.5 billion parameters, demonstrated zero-shot task performance without any fine-tuning, sparked the first major AI safety debate with its "too dangerous to release" rollout, and established the scaling hypothesis that larger models develop qualitatively new capabilities.8 min→05T5: The Text-to-Text Transfer TransformerT5 (Raffel et al., 2019) unified every NLP task into a single text-to-text format, conducted the most systematic empirical study of transfer learning design choices, and introduced the C4 dataset — demonstrating that encoder-decoder models could match or exceed decoder-only approaches when all tasks are treated as text generation.8 min→06XLNet: Permutation Language ModelingXLNet (Yang et al., 2019) introduced permutation language modeling to capture bidirectional context without BERT's [MASK] token corruption, combining the strengths of autoregressive and autoencoding approaches while integrating Transformer-XL's recurrence mechanism for longer-range dependencies — outperforming BERT on 20 benchmarks before being eclipsed by simpler alternatives.8 min→07Encoder-Only vs Decoder-Only vs Encoder-Decoder: The Three Architecture ParadigmsThe Transformer spawned three architectural families — encoder-only (BERT), decoder-only (GPT), and encoder-decoder (T5) — each with distinct strengths, and the surprising dominance of the decoder-only paradigm in the scaling era is one of the most consequential developments in modern AI, though the story is more nuanced than "decoder-only won."8 min→
№ 03The Bert Ecosystem
01RoBERTa: A Robustly Optimized BERT Pretraining ApproachRoBERTa (Liu et al., 2019) demonstrated that BERT was dramatically undertrained by removing the Next Sentence Prediction task, using dynamic masking, training with larger batches on 10x more data for longer — matching or exceeding XLNet on all benchmarks with zero architectural changes, and proving that training methodology matters as much as model design.8 min→02ALBERT: A Lite BERTALBERT (Lan et al., 2019) introduced factorized embedding parameterization and cross-layer parameter sharing to reduce BERT's parameter count by up to 18x while maintaining competitive performance, replacing Next Sentence Prediction with the harder Sentence Order Prediction task — an early and influential exploration of parameter efficiency that foreshadowed the model compression revolution.8 min→03DistilBERT: Knowledge Distillation Applied to BERTDistilBERT (Sanh et al., 2019) applied knowledge distillation to compress BERT into a model 40% smaller and 60% faster while retaining 97% of its language understanding capabilities — the first major "deployment-ready" BERT variant and Hugging Face's foundational research contribution that helped establish them as the central platform of the NLP ecosystem.8 min→04DeBERTa: Decoding-Enhanced BERT with Disentangled AttentionDeBERTa (He et al., 2020) introduced disentangled attention — separating content and position into independent representations with dedicated attention matrices — and an enhanced mask decoder that reintroduces absolute position for prediction, surpassing human performance on the SuperGLUE benchmark and representing the high-water mark of the encoder-only paradigm.8 min→05ELECTRA: Efficiently Learning an Encoder That Classifies Token Replacements AccuratelyELECTRA (Clark et al., 2020) replaced masked language modeling with a generator-discriminator framework where a small generator creates plausible token replacements and the main model learns to detect which tokens were replaced — training on all input tokens instead of just 15%, achieving 4x greater sample efficiency and matching RoBERTa-level performance with a fraction of the compute.8 min→06ModernBERT and the Encoder RevivalModernBERT (Warner et al., 2024) applied 2024-era techniques — RoPE positional encodings, Flash Attention 2, GeGLU activations, unpadding, and training on 2 trillion tokens — to the encoder-only architecture, outperforming all existing encoders and disproving the narrative that "encoders are dead" by showing they were not obsolete but simply under-invested.8 min→
№ 04The Scaling Era
01GPT-3OpenAI's 175-billion-parameter language model demonstrated that massive scale unlocks in-context learning, allowing a single model to perform diverse tasks from just a few examples in the prompt.7 min→02Kaplan Scaling LawsKaplan et al. discovered that language model loss follows smooth power-law relationships with model size, dataset size, and compute, providing a quantitative roadmap for building ever-larger models.7 min→03Chinchilla and Compute-Optimal TrainingDeepMind's Chinchilla paper overturned the prevailing wisdom on model scaling, proving that a 70B model trained on 1.4 trillion tokens could beat models 2-8x its size by simply using more training data.7 min→04PaLMGoogle's 540-billion-parameter Pathways Language Model demonstrated that a single dense Transformer, trained across 6,144 TPU v4 chips, could achieve breakthrough performance on reasoning, code, and multilingual tasks simultaneously.6 min→05Codex and Code GenerationOpenAI's Codex, a GPT-3 model fine-tuned on 54 million GitHub repositories, proved that language models could write functional code and launched the AI-assisted programming revolution through GitHub Copilot.7 min→06Emergent Abilities of Large Language ModelsCertain capabilities — like few-shot arithmetic, chain-of-thought reasoning, and word unscrambling — appear to emerge unpredictably at specific model scales, sparking a fierce debate about whether these phase transitions are real or artifacts of how we measure.7 min→07LaMDA and Conversational AIGoogle's 137-billion-parameter dialogue model, trained on 1.56 trillion words of conversation data and optimized for safety, factual grounding, and conversational quality, became unexpectedly famous when a Google engineer claimed it was sentient.7 min→08The Scaling Hypothesis DebateThe contested idea that intelligence is an emergent property of sufficient scale — that making models bigger and training them on more data will eventually produce general intelligence — became the defining intellectual debate of the LLM era.9 min→
№ 05Alignment And The Chatgpt Moment
01InstructGPT and RLHFOpenAI's InstructGPT demonstrated that a 1.3B parameter model aligned with human preferences via reinforcement learning from human feedback could be preferred over the 175B GPT-3, proving that alignment technique matters as much as raw scale.8 min→02ChatGPTReleased on November 30, 2022, ChatGPT was a conversationally fine-tuned GPT-3.5 model that reached 100 million users in two months, transforming large language models from research curiosities into the fastest-growing consumer product in history.7 min→03Constitutional AIAnthropic's Constitutional AI replaced the need for extensive human labeling of harmful content by having the model critique and revise its own outputs according to a written set of principles, then training a preference model using AI-generated judgments (RLAIF).8 min→04Direct Preference Optimization (DPO)Rafailov et al. showed that the RLHF objective could be mathematically reformulated as a simple classification loss on preference pairs, eliminating the need for a separate reward model and the instability of RL training while matching or exceeding PPO's quality.8 min→05Instruction Tuning and FLANGoogle's FLAN showed that fine-tuning language models on diverse NLP tasks phrased as natural-language instructions dramatically improves zero-shot generalization, and scaling to 1,800 tasks produced some of the largest gains in model capability per dollar ever observed.8 min→06Synthetic Data for TrainingThe practice of using language models to generate training data for other (or the same) models became a defining technique of the LLM era, enabling everything from Stanford Alpaca's $600 chatbot to DeepSeek-R1's reasoning breakthroughs.8 min→07GPT-4OpenAI's GPT-4, released in March 2023, was the first multimodal frontier model to accept both text and image inputs, and it achieved a massive leap in reasoning, coding, and factuality that set a new ceiling for AI capabilities — while revealing almost nothing about how it was built.7 min→08The AI Arms Race BeginsChatGPT's explosive success in late 2022 triggered a global technology arms race, with Google declaring "code red," Microsoft investing $10B+ in OpenAI, and annual AI investment surpassing $100 billion as every major tech company scrambled to compete.9 min→
№ 06The 2023 Model Boom
01LLaMA 1Meta AI's LLaMA proved that smaller models trained on more data could outperform much larger ones, and its leaked weights ignited the open-source AI movement.7 min→02The Alpaca EffectStanford's $600 fine-tuning of LLaMA triggered a Cambrian explosion of open-source instruction-tuned models, proving that capable AI assistants could be built on a graduate student budget.7 min→03LLaMA 2Meta's LLaMA 2 was the first truly commercially licensed open-weight language model, combining 2 trillion tokens of training with extensive RLHF alignment to narrow the gap between open and closed AI.7 min→04Mistral 7BA Paris-based startup released a 7.3-billion-parameter model via a torrent magnet link with no paper and no marketing, and it outperformed every open model twice its size.7 min→05Mixtral 8x7BMistral AI's sparse Mixture of Experts model used 46.7 billion total parameters but only 12.9 billion per forward pass, matching LLaMA 2 70B quality at a fraction of the inference cost and proving MoE was practical for the open-source community.8 min→06FalconThe Technology Innovation Institute's Falcon models proved that exceptional data curation alone — without novel architectures or proprietary text — could produce world-class language models, briefly topping the Hugging Face Open LLM Leaderboard.7 min→07Claude 1 and 2Anthropic's Claude models brought Constitutional AI from theory to product, establishing the "safety-first" brand in commercial AI and pioneering the long-context paradigm with 100K and eventually 200K token windows.8 min→08Gemini 1Google DeepMind's Gemini was the first natively multimodal large model — trained from the ground up on text, images, audio, and video — and represented Google's consolidated answer to GPT-4 after a year of playing catch-up.8 min→
№ 07The 2024 Frontier Race
01Claude 3 FamilyAnthropic's March 2024 release introduced a three-tier model system — Haiku, Sonnet, and Opus — all with 200K context windows, with Opus becoming the first model to credibly challenge GPT-4's supremacy across major benchmarks.6 min→02Gemini 1.5Google DeepMind's Gemini 1.5, released in February 2024, introduced a Mixture of Experts architecture with an unprecedented 1 million token context window — later extended to 2 million — fundamentally redefining what it means to give a model "enough context."7 min→03GPT-4oOpenAI's GPT-4o ("Omni"), released in May 2024, was the first truly unified multimodal model — trained end-to-end to accept and generate text, audio, images, and video through a single neural network, at 2x the speed and half the cost of GPT-4 Turbo.6 min→04Claude 3.5 SonnetReleased on June 20, 2024, Claude 3.5 Sonnet shattered the assumption that mid-tier models must be inferior — it outperformed Claude 3 Opus on nearly every benchmark at 2x the speed and lower cost, becoming the most influential single model release of 2024.7 min→05LLaMA 3 and LLaMA 3.1Meta's LLaMA 3 (April 2024) and LLaMA 3.1 (July 2024) proved that open-weight models could compete at the absolute frontier, with the 405B parameter model rivaling GPT-4o and Claude 3.5 Sonnet while being freely available for download.7 min→06LLaMA 3.2: Multimodal and Edge ModelsMeta's LLaMA 3.2 (September 2024) brought vision capabilities to the open-weight LLaMA family for the first time with 11B and 90B multimodal models, while also releasing tiny 1B and 3B text models for on-device deployment — and LLaMA 3.3 later showed a 70B model could match the 405B.7 min→07Grok and xAIElon Musk's xAI built Grok from zero to frontier-competitive in under two years, open-sourcing the 314B parameter Grok-1, scaling on the massive Colossus GPU cluster, and reaching the top of LMArena rankings by late 2025 — embodying the "move fast, scale hard" philosophy.7 min→08PaLM 2 and the Gemini EvolutionGoogle's journey from PaLM (540B dense, 2022) through PaLM 2 (Chinchilla-optimal, 2023) to Gemini 1.0 (2023) and Gemini 1.5 (MoE, 2024) traces the company's strategic pivot from "scale the biggest model" to "scale efficiently with MoE and long context."7 min→09Mistral Large and Enterprise ExpansionMistral AI expanded from its scrappy open-source origins into a full enterprise AI platform through 2024, releasing Mistral Large 2 (123B dense), Codestral (22B code specialist), Pixtral (12B multimodal), and Mistral Nemo (12B) — establishing Europe's first credible frontier AI lab.8 min→
№ 08Reasoning And Inference Scaling
01OpenAI o1: Trained ReasoningOpenAI o1 was the first model explicitly trained to reason through reinforcement learning on chain-of-thought, proving that thinking longer at inference time could dramatically improve performance on hard problems.9 min→02The o-Series Evolution: o1 to o4-mini (and Beyond)OpenAI's o-series evolved from o1's proof-of-concept in reasoning through o3 and o4-mini, achieving dramatic improvements in capability and cost efficiency across five models in eight months, before its reasoning advances were fully absorbed into the GPT-5 line.11 min→03DeepSeek-R1: Open Reasoning from Pure RLDeepSeek-R1 demonstrated that sophisticated reasoning capabilities could emerge from pure reinforcement learning without supervised fine-tuning, matching OpenAI o1 at a fraction of the cost and releasing everything under an open license.9 min→04Test-Time Compute Scaling: Thinking Longer Beats Training BiggerTest-time compute scaling is the paradigm that allocating more computation during inference (letting a model think longer) can be more cost-effective than training a larger model, opening a second axis for improving AI capabilities.9 min→05The Reasoning Paradigm ShiftAI reasoning evolved in three phases, from chain-of-thought prompting tricks in 2022, through search-based improvements in 2023, to fully trained reasoning via reinforcement learning in 2024, transforming reasoning from a fragile prompt hack into a robust learned capability.9 min→06Hybrid Thinking Models: On-Demand ReasoningHybrid thinking models give users the ability to toggle reasoning on and off and set thinking budgets, combining the speed of traditional LLMs with the depth of reasoning models in a single system.11 min→
№ 09The Cost Revolution And Global Competition
01DeepSeek V2: Multi-head Latent AttentionDeepSeek V2 introduced Multi-head Latent Attention (MLA), a novel attention mechanism that compressed the KV cache by 93.3%, making frontier-quality inference dramatically cheaper and signaling that architectural innovation could substitute for brute-force compute.9 min→02DeepSeek V3: Frontier Quality at Startup CostDeepSeek V3 matched Claude 3.5 Sonnet and GPT-4o across most benchmarks while training for just $5.576 million, combining innovations in FP8 training, multi-token prediction, and efficient MoE routing to shatter assumptions about the cost of frontier AI.9 min→03The DeepSeek Cost RevolutionDeepSeek demonstrated through V2, V3, and R1 that frontier AI could be built for a fraction of Western lab budgets, triggering a trillion-dollar market shock and forcing the entire industry to rethink the relationship between compute spending and AI capability.9 min→04Qwen 1 and 2: Alibaba's AscentAlibaba's Qwen model family evolved from a competent bilingual system in 2023 to a leading open-weight family by late 2024, demonstrating that consistent iteration on data quality and architecture could close the gap with Western frontier models.8 min→05Qwen 3: The Open Frontier ChallengerQwen 3 brought hybrid thinking, MoE scaling, and 119-language support to the open-weight ecosystem, challenging the notion that frontier reasoning required closed, proprietary models.10 min→06Chinese AI Labs: The Global Competition LandscapeBeyond DeepSeek and Qwen, a diverse ecosystem of Chinese AI labs emerged between 2023 and 2025, collectively challenging Western dominance through architectural innovation, massive domestic deployment, and creative adaptation to chip export restrictions.12 min→
№ 10The Small Model Revolution
01Phi SeriesMicrosoft Research's Phi models proved that training data quality matters more than model size, achieving frontier-class performance with models as small as 1.3 billion parameters.7 min→02GemmaGoogle DeepMind's Gemma series brought Gemini-class technology to the open-weight ecosystem, evolving from simple text models to multimodal, multilingual systems designed for edge deployment.7 min→03Knowledge Distillation for LLMsKnowledge distillation evolved from compressing BERT-era models by mimicking output probabilities to a modern paradigm where large "teacher" models generate entire synthetic training datasets -- including reasoning traces -- that transfer intelligence through data rather than architecture mimicry.8 min→04Quantization and CompressionQuantization techniques evolved from a niche optimization into the critical bridge that brought frontier-class language models from data center clusters to consumer laptops, shrinking memory requirements by 4x with less than 1% quality loss.8 min→05LoRA and Fine-Tuning DemocratizationLow-Rank Adaptation (LoRA) transformed LLM fine-tuning from a privilege of well-funded labs into something any developer with a single GPU could do, by training only 0.1-1% of a model's parameters through injected low-rank matrices.8 min→06llama.cpp and Local InferenceGeorgi Gerganov's llama.cpp project, started in March 2023 as a C/C++ port of LLaMA inference, sparked a revolution in local AI by proving that large language models could run on ordinary laptops and even phones without a GPU.8 min→07The SLM RevolutionThe Small Language Model revolution proved that for the majority of real-world tasks, right-sized models -- optimized for quality data, efficient architecture, and targeted deployment -- outperform the brute-force scaling approach on every practical metric.8 min→
№ 11The 2025 Frontier
01Claude 4 SeriesAnthropic's Claude 4 series (2025-2026) pushed the frontier of coding, agentic capability, and alignment -- from Opus 4's autonomous task dominance through Sonnet 4.5's 30-hour sustained focus and Opus 4.5's benchmark-leading efficiency, to the 4.6 generation's agent teams, with Opus 4.6 Thinking reaching #1 on LMArena at 1506 Elo.9 min→02GPT-5OpenAI's GPT-5 (August 2025) unified traditional language modeling, chain-of-thought reasoning, and native tool use into a single architecture, converging the separate GPT and o-series product lines into one model — then GPT-5.2 (December 2025) pushed the frontier further with three model variants and near-saturating benchmark scores, followed by GPT-5.2-Codex (January 2026) for agentic coding.13 min→03Gemini 2.x and 3: Google's Agent EraGoogle's Gemini series from 2.0 through 3.1 (2024-2026) evolved from a fast multimodal model into the industry's most aggressive push toward agent-native AI, combining native tool use, visible reasoning traces, million-token context, and deep integration with Google's ecosystem — culminating in Gemini 3 Flash outperforming its own flagship on agentic coding, and Gemini 3.1 Pro achieving 94.3% GPQA Diamond and #1 rankings on 12 of 18 tracked benchmarks.13 min→04Llama 4Meta's Llama 4 (April 2025) brought native Mixture of Experts and early-fusion multimodality to the open-weight frontier, with Scout's 10 million-token context window setting a new record for open models.9 min→05Qwen 3 Coder: Domain-Specialized Open ModelsAlibaba's Qwen3-Coder (July 2025) demonstrated that domain-specialized open-weight models could approach frontier closed models on targeted tasks, representing a broader trend of specialization as a path to competitive performance.9 min→06Agent-Native Models: Built for AutonomyAgent-native models (2024-2026) represent a paradigm shift from language models designed to generate text toward models trained from the ground up for autonomous action — using tools, navigating interfaces, recovering from errors, and completing multi-step tasks in the real world.10 min→07Open vs Closed: The Narrowing GapThe capability gap between open-weight and closed frontier models collapsed from ~17.5 MMLU points in 2023 to near-parity by 2025, and by early 2026 the best open model trailed the best closed model by less than 1% on SWE-bench coding — driven by better training data, MoE architectures, and reasoning distillation, with remaining edges narrowing to multimodal, safety, and ecosystem differentiation.10 min→
№ 12Architectural Innovation Threads
01Attention Mechanism EvolutionThe journey from every attention head having its own memory to groups sharing compressed memory — a relentless drive to make attention cheaper without making it dumber.7 min→02Positional Encoding EvolutionHow Transformers went from rigid, pre-set notions of word order to flexible, rotatable representations that let models generalize to sequences far longer than anything seen during training.7 min→03Flash Attention and Hardware-Aware ComputingThe realization that attention's bottleneck was not arithmetic but memory bandwidth, and the tiling algorithm that turned that insight into a 2-4x speedup with zero approximation.7 min→04Mixture of Experts EvolutionThe three-decade journey from a theoretical gating idea to the dominant architecture for frontier models — getting more parameters without paying for all of them at inference time.8 min→05State Space Models and MambaThe bet that linear-time sequence models can challenge the Transformer's quadratic attention — and the selective state space mechanism that made that bet credible.8 min→06KV Cache and Serving OptimizationHow the field borrowed operating system concepts — virtual memory, paging, demand allocation — to solve the memory crisis of storing every token's past for every concurrent request.8 min→07Long-Context TechniquesThe twenty-fold expansion of context windows from 512 tokens to 10 million — achieved through positional encoding tricks, memory-efficient attention, and the hard-won realization that nominal context length and effective context length are not the same thing.8 min→08Normalization and Activation EvolutionThe quiet evolution of normalization (LayerNorm to RMSNorm) and activation functions (ReLU to SwiGLU) in transformers represents the kind of incremental architectural refinement that individually yields small gains but collectively defines the "modern LLM recipe."10 min→09Speculative Decoding and Inference SpeedupsSpeculative decoding and related inference optimization techniques overcome the autoregressive bottleneck — generating tokens one at a time — to achieve 2-10x speedups in production LLM serving without sacrificing output quality.8 min→
№ 13Training Innovation Threads
01Pre-Training Objectives EvolutionThe training objectives used to pre-train language models have evolved from simple next-token prediction into a diverse ecosystem of techniques, each making different tradeoffs between efficiency, bidirectionality, and downstream performance.8 min→02The Data Quality RevolutionThe field's understanding of training data shifted from "more is better" to "quality, curation, and diversity matter more than raw volume," fundamentally changing how LLMs are trained.8 min→03Alignment Method EvolutionAlignment methods — techniques for making LLMs follow human intent and values — have evolved from complex multi-stage pipelines (RLHF) to simpler single-stage approaches (DPO) to pure reinforcement learning from verifiable outcomes.8 min→04Instruction Tuning EvolutionInstruction tuning — fine-tuning models on task instructions and desired responses — evolved from small hand-crafted datasets to massive LLM-generated corpora, becoming the critical bridge between raw pre-training and useful assistant behavior.8 min→05Distributed Training InfrastructureTraining modern LLMs requires distributing computation across thousands to hundreds of thousands of GPUs using sophisticated parallelism strategies, making distributed training infrastructure as critical as model architecture itself.7 min→06The Synthetic Data RevolutionSynthetic data — training data generated by LLMs themselves — has become the primary fuel for post-training, enabling cheaper instruction tuning, reasoning distillation, and alignment at a fraction of the cost of human-annotated data.8 min→07Training Efficiency BreakthroughsA series of compounding innovations in numerical precision, attention computation, communication scheduling, and architectural design have reduced LLM training costs by 10-50x, making frontier-quality models achievable without frontier-scale budgets.7 min→
№ 14Multimodal Evolution
01Vision-Language Models: Connecting Sight and LanguageVision-language models learn to connect visual perception with language understanding, evolving from contrastive image-text matching (CLIP) to full visual reasoning capabilities integrated into large language models.8 min→02Native Multimodal TrainingNative multimodal training jointly trains a single model on text, images, audio, and video from the ground up, producing cross-modal understanding that adapter-based approaches cannot achieve.8 min→03Audio and Speech ModelsAudio and speech capabilities in LLMs evolved from specialized speech recognition systems to native audio understanding and generation, culminating in models that can hold real-time spoken conversations with emotional nuance.7 min→04Video UnderstandingVideo understanding in LLMs extends visual reasoning from static images to temporal sequences, enabling models to comprehend narratives, track objects, and answer questions about events unfolding over minutes to hours.8 min→05The Convergence Toward Omni-ModelsThe AI field is converging from separate specialized models for each modality toward unified "omni-models" that perceive, reason about, and generate text, images, audio, video, and code within a single architecture.8 min→
№ 15The Llm Landscape
01The Benchmark and Evaluation LandscapeThe evolution of LLM benchmarks from MMLU through SWE-bench and Chatbot Arena reflects a recurring cycle — new benchmark, rapid progress, saturation, replacement — exposing fundamental tensions between measurability and meaningful evaluation.11 min→02The API Economy: How LLMs Are CommercializedThe LLM API economy — pioneered by OpenAI in 2020 and transformed by DeepSeek's cost revolution in 2025 — created a multi-billion-dollar industry where the fundamental business dynamics are shaped by relentless price deflation, tiered model strategies, and the competitive pressure of free open-weight alternatives.9 min→03AI Safety and GovernanceThe rapid scaling of LLM capabilities from 2023 to 2025 outpaced governance frameworks, producing a patchwork of legislation (EU AI Act), voluntary commitments (Responsible Scaling Policies), and technical safety measures (red-teaming, model evaluations) that reflect deep disagreements about how to balance innovation with risk.10 min→04The Open-Source EcosystemThe open-source AI ecosystem — from Hugging Face's model hub to llama.cpp's local inference to vLLM's production serving — created the infrastructure that turned open model weights into a global innovation engine, enabling anyone to run, modify, and build on frontier AI.8 min→05Where LLMs Are HeadingThe trajectory of LLMs points toward a convergence of agentic autonomy, efficient reasoning, multimodal integration, and open-weight parity — raising fundamental questions about the nature of understanding, the economics of knowledge work, and the alignment of increasingly capable systems.10 min→