AI NEWS
AI news, read by someone who has built it
Most AI news tells you what was announced. Almost none of it tells you what to do about it.
Latest
Showing 17–24 of 42
Analysis | 2024–2025 source review
Perplexity Comet: assess the browser as a work environment
This analysis is published on 25 October 2026 and examines the original 2025-07-09 announcement or report. It is not a claim of a launch this week or a current availability guide. Evaluate the whole browser session, including contextual access and user correction time, rather than isolated assistant answers.
Read the note →Analysis | 2024–2025 source review
Agentic Commerce Protocol: reconcile the order, not just the chat
This analysis is published on 25 October 2026 and examines the original 2025-09-29 announcement or report. It is not a claim of a launch this week or a current availability guide. Approve a checkout integration only when payment authorization, order creation and recovery can be reconciled independently.
Read the note →Analysis | 2024–2025 source review
Stargate and AI data centers: separate investment from usable capacity
This analysis is published on 25 October 2026 and examines the original 2025-01-21 announcement or report. It is not a claim of a launch this week or a current availability guide. Evaluate infrastructure through contracted service, delivery milestones and workload performance rather than headline investment totals.
Read the note →Analysis | 2024–2025 source review
AI energy consumption: turn the IEA report into a workload budget
This analysis is published on 25 October 2026 and examines the original 2025-04-10 announcement or report. It is not a claim of a launch this week or a current availability guide. Budget energy against useful completed work and local electricity constraints, not a universal per-prompt estimate.
Read the note →ANALYSIS
Small language models: decide what stays on device
Apple’s 2024 architecture illustrates a practical choice: place each task where its data, latency and quality requirements can actually be met.
Read the note →ANALYSIS
NVIDIA Blackwell: measure cost per useful answer
The 18 March 2024 Blackwell announcement made a strong inference-economics claim. Turning that claim into a budget requires workload-specific evidence.
Read the note →ANALYSIS
Phi-4 reasoning: a small model can still think expensively
Microsoft’s April 2025 report shifted attention from parameter count to inference-time computation. Buyers need to test that trade-off in their own tasks.
Read the note →ANALYSIS
DeepSeek R1: open weights do not remove operating responsibility
The January 2025 release widened deployment choices, but enterprise adoption still requires version, license and data-path decisions.
Read the note →
Want this as a keynote for your leadership team?
I turn the same material into a 45–60 minute keynote, customised for your industry and the decisions your room is facing.
BOOK AMER →
