𒀭 How We Got Here

"AI" has been declared dead and reborn at least three times. The current era didn't appear from nowhere — it's the result of decades of dead ends, slow compounding, and a few breakthroughs that changed the slope of the curve entirely.

1956
The Dartmouth Workshop
The term "artificial intelligence" is coined. Early researchers bet that human reasoning could be reduced to symbolic logic and search — if a machine could manipulate symbols correctly, it could think.
1970s–80s
Expert systems, then the first AI winter
Hand-coded rule systems (MYCIN, DENDRAL) showed promise in narrow domains, then hit a wall: they couldn't generalize and were brutal to maintain. Funding dried up. "AI winter" enters the vocabulary.
1986
Backpropagation goes mainstream
Rumelhart, Hinton, and Williams popularize backpropagation for training multi-layer neural networks. The idea existed earlier, but this is when it starts reshaping the field — slowly, because compute and data were still tiny.
1997
Deep Blue beats Kasparov
Brute-force search, not learning — but a cultural turning point. The public starts taking machine "intelligence" seriously again, even if the engine underneath was still mostly symbolic.
2012
AlexNet and the deep learning explosion
A convolutional neural network crushes the ImageNet competition, trained on GPUs. This is the moment "deep learning" stops being a niche academic interest and becomes the dominant paradigm — bigger data and bigger compute start beating cleverer hand-engineering.
2017
"Attention Is All You Need"
Google researchers publish the Transformer architecture. It replaces recurrence with self-attention, making models dramatically more parallelizable to train. Every major language model since has been a descendant of this one paper.
2018–2020
GPT, BERT, and the scaling laws
OpenAI and Google show that simply training Transformers bigger, on more text, with more compute, keeps producing better models — fairly predictably. This "scaling law" observation becomes the industry's central bet for the next several years.
November 2022
ChatGPT goes public
GPT-3.5 wrapped in a chat interface reaches 100 million users faster than any consumer product before it. AI stops being a research topic and becomes a daily tool — and a daily news cycle.
2023
GPT-4, and open weights fight back
GPT-4 raises the ceiling on closed models. Meanwhile Meta releases LLaMA, and the open-weight ecosystem (Mistral, then dozens of fine-tunes) proves frontier-adjacent performance doesn't require a closed API.
2024
Mixture-of-experts goes mainstream, agents emerge
Sparse Mixture-of-Experts architectures (Mixtral and others) let models grow capacity without proportionally growing inference cost. Tool-use and "agentic" workflows — models that plan, call tools, and act in loops — move from demos to production.
Early 2025
DeepSeek breaks the cost assumption
DeepSeek releases an open-weight reasoning model trained for a fraction of the assumed frontier-lab budget, rattling the belief that only a handful of well-funded labs could compete at the frontier. Markets notice; so does policy.
2025–2026
Orchestration and the agent stack
The bottleneck shifts from "is the model smart enough" to "can it be wired into real systems reliably." Tools like Ollama (local model runtimes) and orchestration layers like OpenClaw mature, letting a single operator run and swap models across local and hosted backends from one control plane.
Now
Multi-agent systems, and regulation catching up
Long-running agents, multi-agent collaboration, and on-device models running on consumer hardware are converging — while governments worldwide scramble to write rules for systems that are already deployed at scale. See the live feed below for what's actually happening this week.

𒁹 The Hardware Underneath: GPUs & VRAM

Every model above runs on one fact: a neural network is matrix multiplication, repeated billions of times. That's the whole story of why GPUs won, and why VRAM — not raw compute — is usually what actually limits what you can run.

CPU GPU 8–32 large cores, fast serial logic thousands of small cores, the same op in parallel
Training and inference are almost entirely that one operation — which is why AlexNet's 2012 GPU-trained win over ImageNet is the moment deep learning took off, and why NVIDIA's CUDA stack still has no serious rival.

VRAM, not FLOPs, is usually the real bottleneck. A model's weights have to fit in memory before it can do anything — at 16-bit precision, roughly 2 bytes per parameter, so a 7-billion-parameter model needs ~14GB just to load. Training needs far more: gradients, optimizer state, and activations all have to fit in memory at the same time.

Training Inference 4-bit Inference ~94GB ~18GB ~5.5GB
Weights Gradients Optimizer states KV cache / activations
Illustrative figures for a 7B-parameter model. Training has to hold weights, gradients, optimizer state, and activations in memory all at once — quantized inference needs almost none of that, which is why the same model that needs multiple data-center GPUs to train can run on a laptop to use.

Most of the clever engineering in modern AI infrastructure is really just different answers to "how do we fit this in VRAM":

Fine-tuning

LoRA / QLoRA

Freeze the original weights and train small low-rank "adapter" matrices instead. No gradient or optimizer memory needed for the billions of frozen parameters — just for the tiny adapters.

Multi-GPU training

Tensor & ZeRO Sharding

Split a model's layers, or its gradients and optimizer state, across multiple GPUs instead of duplicating everything on each one. This is how models too big for any single GPU get trained at all.

Architecture

Mixture-of-Experts

A model can have a huge total parameter count but only activate a small subset of "expert" sub-networks per token — big on disk, cheap in active VRAM per forward pass. Part of how DeepSeek punched above its training budget.

𒈗 The Difference Between Models

"Which AI is best" is the wrong question — different models are built under different constraints, and those constraints determine what you can actually do with them.

Closed / API-Only

  • e.g. GPT-4/5, Claude, Gemini
  • Weights never leave the provider's servers
  • You pay per token, get the latest model instantly
  • Usually the strongest general-purpose reasoning
  • Your data passes through someone else's infrastructure

Open-Weight

  • e.g. Llama, Mistral, Qwen, DeepSeek
  • Weights are published — anyone can download and run them
  • Can be fine-tuned, inspected, and self-hosted
  • Quality gap to closed frontier models keeps narrowing
  • You take on the hosting and ops burden

Local / On-Device

  • Distilled or quantized versions of open-weight models
  • Run entirely on your own hardware — no network call
  • Zero marginal cost per query, full data privacy
  • Capped by your GPU/RAM — smaller than frontier models
  • You manage updates, drivers, and model swaps yourself

𒆳 Local Hosting vs. API Key

The real decision isn't "which model" — it's "who controls the weights, and who pays for the compute." Here's the actual tradeoff.

FactorLocal HostingAPI Key
PrivacyData never leaves your machineProvider sees every request
Cost shapeUpfront hardware, then near-zero per queryNo hardware, but pay per token forever
Model qualityCapped by what fits on your hardwareAccess to the frontier, whatever that is this week
LatencyNo network round-trip; bound by your GPUDepends on provider load and your connection
MaintenanceYou manage updates, drivers, quantizationProvider handles all of it silently
AvailabilityWorks offline, never rate-limited by anyone elseDown if the provider is down or you hit a cap
Best forSensitive data, high-volume repetitive tasks, tinkeringBest-in-class reasoning, low-volume or bursty use

☊ DeepSeek, Ollama, OpenClaw

Three names that come up constantly once you go past "just use ChatGPT" — and what each one actually is.

Model lab

DeepSeek

A Chinese AI lab that publishes open-weight models — most notably its R1 reasoning models — trained with aggressive efficiency techniques (mixture-of-experts, reinforcement learning on verifiable tasks) at a fraction of the assumed frontier-lab cost. It matters less because it's "free" and more because it proved frontier-adjacent reasoning doesn't require frontier-lab budgets.

Local runtime

Ollama

A tool that makes running open-weight models on your own machine as simple as pulling a Docker image — ollama pull, ollama run. It handles quantization and serving so you don't have to wire up inference code yourself. The default on-ramp for "I want a model running on my laptop, today."

Orchestration

OpenClaw

An agent-orchestration layer that sits above whatever models you're running — local or API — and gives you one gateway to set models, route tools, and manage a persistent agent across sessions. It's the difference between "I called an API once" and "I run an agent that keeps working." This site's own agent runs through it.

☍ Context Servers: How Agents Reach the Real World

A model that can only see what's stuffed into its prompt is a chatbot. A model that can reach out and act on real systems — live, mid-conversation — is an agent. The thing that makes that distinction practical at scale is a standard called MCP (Model Context Protocol).

AI Agent Claude, GPT, etc. MCP Client built into the agent host GitHub MCP Server → repos, PRs, issues Database MCP Server → internal data Filesystem MCP Server → local/shared files Slack MCP Server → channels, messages
Each MCP server exposes a standard set of callable "tools." The agent doesn't need custom code per integration, and the integration doesn't need to know which model is calling it — this is the same pattern this site's own agent uses to work with its GitHub repository.

Released by Anthropic in late 2024, MCP is gaining traction precisely because it's become cross-vendor — OpenAI and Google have both moved toward supporting it or compatible patterns, which is the strongest signal that it's turning into standard plumbing rather than one company's format. Think of it less like a clever feature and more like USB-C: boring, standardized, and exactly what lets an ecosystem of tools snap together instead of every vendor inventing its own port.

𓋹 How Corporations Are Adopting This

Enterprise AI adoption right now is less "which model" and more "which tier handles which job, routed through one governed layer."

Frontier API Claude, GPT, Gemini — hardest reasoning Self-Hosted Open-Weight Llama, Mistral, DeepSeek — narrow, high-volume Internal MCP / Governance Layer auth, audit, rate limits, tool allowlists Internal Systems docs · tickets · databases · code
Most companies aren't picking a single model — they're routing tasks to whichever tier fits, through one governed layer instead of a bespoke integration per vendor per tool.

The infrastructure arms race is really a memory race. Frontier labs and hyperscalers are buying GPU clusters at an unprecedented scale, but the conversation has shifted from "how many GPUs" to "how much HBM and interconnect bandwidth" — that's what actually gates context length, batch size, and concurrent users a deployment can serve.

MCP as a governance layer, not just a connector. Companies are standing up internal MCP servers as an audited, permissioned bridge between AI copilots and proprietary systems — instead of handing a model raw API keys, or writing a one-off integration for every tool it might need to touch. That's the actual inflection point: the difference between "we pay for an AI subscription" and "we have agents that can do things in our systems."

𓁹 Where It's Headed

A few directions are converging at once, and none of them are slowing down.

Agents, not chatbots. The center of gravity is moving from single question-answer turns to long-running agents that plan, use tools, and act over hours or days with minimal supervision.

Compute efficiency over raw scale. DeepSeek's cost breakthrough forced the entire industry to take "do more with less" seriously — mixture-of-experts, distillation, and better training data curation are now competing directly with "just add more GPUs."

On-device gets real. Quantized models small enough to run on a phone are closing in on what needed a data center two years ago — pushing privacy-sensitive and offline use cases out of the cloud entirely.

Regulation is arriving mid-deployment. Governments are writing AI rules for systems that are already in production at global scale, not systems on a drawing board — which means the rules are reactive, regionally fragmented, and frequently behind the technology they target.

Physical compute becomes the bottleneck. As models plateau on raw text capability, the race shifts toward chips, power, and data-center capacity — turning AI progress into as much a hardware and energy story as a software one.

Live: Advances, Policy & Market

Continuously refreshed from research labs, policy trackers, and tech/market press — no page reload needed.

Scrying the feeds…