Insights for running a business that doesn’t need you.
Practical operations and AI insights to help your business become easier to run, and less dependent on you.
Page 5 of 7
Harness Engineering: How Meta Cut Agent Failures 87.5 Per Cent Without Changing the Model
Meta's Auto-RecSys preprint cut agent operational failures from 4.0 to 0.5 per experiment without changing the model. What harness engineering is and what transfers to smaller operations.
6 min read
Jev in Practice: Where a Decision Model Beats a Text LLM, and Where It Never Will
Jev is a decision engine, not a chatbot: it picks from answers you defined and scores its own confidence. Where that fits, what it costs, and what it cannot do.
5 min read
The Obligation You Inherited at Completion: TUPE and the Employee Records You Must Have by Week One
UK business transfers move employment terms, history and liability to you at completion, and TUPE sets a four-week records duty before day one. Here is the week-one audit that establishes what you actually inherited.
7 min read
The Recurrent Looped Transformer: When a Model's Reasoning Depth Grows With the Conversation
A new architecture lets a model's reasoning path grow with the conversation. What the Recurrent Looped Transformer does, the measured evidence behind recurrent depth, and the limits its author flags.
6 min read
GLM-5.3-Flash vs DeepSeek V4.1 Flash: Memory Decides the Two-Box Local AI Cluster
DeepSeek V4.1 Flash scores higher on agentic benchmarks, but its 510GB checkpoint cannot fit a 256GB two-box cluster. GLM-5.3-Flash is the frontier-class model that does, and fit decides.
5 min read
The Passwords and Permissions You Inherited: Build the Access Baseline in Week One
Undocumented access is the fastest operational risk you inherit at close. Count every system, account and admin right in week one, and file the register.
7 min read
The Best AI Model for a Small Cluster Is the One That Fits
DeepSeek's V4.1 Flash leads on speed and price, but GLM 5.3 Flash is the model with a working recipe on a 256GB desk cluster. Fit beats benchmark rank.
7 min read
TypeSafe's Jev Sells Decisions, Not Words: What a Decision-Only AI Model Changes
TypeSafe AI launched Jev, a decision-only model that returns typed, confidence-scored answers with output tokens billed at zero. What the launch proves, what it does not, and the inventory an owner can build now.
8 min read
The Weekly Meeting Most Founders Skip Is the Evidence Buyers Want
A fixed weekly meeting with a one-page pack turns founder-run operations into documented process, and leaves the dated evidence buyer diligence asks for.
7 min read