Three weeks in owned AI: what changed in October

October was a dense month for anyone who wants a company to own its AI rather than rent it. Here is what moved, what it means and where we stand. Each line links to a longer piece or to its source.

Models

• Open weights are a few points from the top. Meta published Muse Spark Open, Reflection AI published Beam, and Kimi K3.1 Pro sits three points below the best closed model on one public index.1 Read the licence first: why it matters.
• Public scores flatter them. Epoch AI puts open models six to seven months behind on cyber tasks, and finds the gap wider on private tests than on public ones.2 Our answer is a sealed test from your own work: more here.

Machines

• On-site got real. NVIDIA’s DGX Station, with 748 GB of memory, ships now, and NVIDIA released a 74B model built to fit in 24 GB.3 What fits where.

Money

• Rented GPUs got about a fifth dearer in a month, while blended token prices fell a quarter in four weeks. Both are set by someone else.4 The numbers.
• A fine-tune lives as long as its base. OpenAI retired a batch of models on 23 October, and its fine-tunes go with them.5 What that means.

Rules and records

• Regulators asked for evidence, not new bans. The UK FCA named self-hosting, notice of model changes and exit options as good practice; the FSB finalised twelve sound practices; the EU moved its high-risk dates to December 2027 and August 2028.6 The detail.
• Courts keep count. A public database lists 2,217 decisions on AI hallucinations, 879 of them involving lawyers.7 What it says to a firm.

And at Eigna

A third candidate of our first client’s forecasting model met our release gate on 23 October. It tied, and a tie keeps the current version, so it was held. The record.

The thread through all of it: the evidence that decides is the evidence on your own work, and control sits with whoever holds the weights and the machine. That is the bet Eigna is built on, and we will keep testing it in the open.