GPT-4.5 vs Claude 3.7 vs Gemini 2.0: a cloud architect's take
Three frontier models on the table at the end of February, three different bets about what "frontier" means. The interesting comparison isn't who wins, it's who fits which workload.
AI
What "thinking models" do to your context-window math
Reasoning tokens have to live somewhere. The somewhere is your context window. Worth working through what the budget actually looks like once you stop pretending output is the only thing that costs.
Claude 3.7 Sonnet: the new defaults of "extended thinking"
Anthropic shipped a model that knows when to think. The interesting question isn't the benchmark numbers, it's what happens when reasoning becomes a default behavior instead of a separate product tier.
What "Knowledge as a Service" actually meant: and why DeepSeek made it cheap
An idea sketched two years ago assumed expertise-as-licensable-artifact would be a premium-tier product. The economic floor just dropped through that assumption. Worth re-examining what the idea was actually about now that the substrate has changed.
Self-hosting open weights at home: what 2025 actually requires
The open-weights frontier is genuinely usable from a home rig now. The bottleneck has moved, it isn't capability anymore, it's memory bandwidth and patience.
Why a $5M training run rattled three trillion in market cap
The DeepSeek-R1 number is small. The market reaction wasn't. The interesting bit isn't the dollar figure, it's which assumption it broke.
The DeepSeek-R1 reality check, one week later
A week ago, an open-source reasoning model erased about a trillion dollars of market cap. Most of the takes from week one are already wrong. Here's what actually changed.
Llama 2 changes the math on owning your model
Meta released Llama 2 last week with a permissive commercial license. Five months after the leak made local inference possible, the legal layer just caught up to the technical one.
Claude 2 and why context windows are the new RAM
Anthropic shipped Claude 2 yesterday with a 100K-token context window. The capability is genuinely interesting; the architectural shift it forces is the bigger story.
After ChatGPT plugins: agents are coming, and most of us aren't ready
ChatGPT plugins shipped a month ago. Function calling shipped last week. The substrate for agents that take action on your behalf just landed, and the second-order problems are going to be ugly.
Encoding a person: what would it actually take?
If the goal is a working representation of how a specific person reasons in their domain, what's the actual stack? Walking through corpus, model, retrieval, eval, and where each piece breaks today.
Could there be a marketplace for AI training data?
The model providers scraped the open internet and called it fair game. The next phase needs an actual market for the data, and the structural pieces of that market don't exist yet.