Blog

Evidence notes

Technical guides and data-led notes grounded in primary sources and source-linked evidence.

Latest notes

06

5 August 20268-minute read

Open models x inference hardware

How Much Hardware Do Open Models Actually Need?

A 16-model deployment map compares seven precision and memory-placement modes, from notebook-scale low-bit checkpoints to multi-node frontier inference.

Central finding

Four-bit remains the practical center. Native low-bit formats and SSD expert streaming change the hardware equation, but they do not guarantee interactive speed.

05

3 August 202616-minute read

Local LLMs x hardware

What is the most capable LLM you can run on a Mac mini?

A 64 GB M4 Pro Mac mini can keep a 70B model at Q4 fully resident, while model-specific Q2 quantization and SSD expert streaming can execute DeepSeek V4 Flash.

Central finding

48 GB is the best balanced configuration. Choose 64 GB for dense 70B models and advanced capacity experiments.

03

11 July 20267-minute read

Build vs buy x time

The AI market is maturing. So why is build winning?

73% of 2,685 classified enterprise AI deployments are custom-built. A time-series view shows how that lead has evolved as the product market has expanded.

Central finding

Custom engineering holds 73% of the classified corpus. The most autonomous workloads lean even further toward custom ownership.

02

11 July 20267-minute read

AI agents x industries

Which industries have entered the agent era?

Agents are 36% of the 1,263 deployments documented in the last 12 months — this note maps the agent share of every measurable industry against the 2023-era baseline.

Central finding

A handful of industries already publish more agentic deployments than classic AI; the fastest risers are physical-economy sectors.

01

10 July 20266-minute read

AI agents x reported outcomes

What changes when AI starts acting?

A controlled look at 2,753 AI deployments published since 2024, comparing the outcomes reported by other AI, single-agent, and multi-agent systems.

Central finding

Agent cases report automation about twice as often as other AI cases. Multi-agent cases show the strongest speed and throughput signals.