Quick personal note: I skipped last weekās digest while I was deep in finalising the book launch - UNHYPED: From Hype to Hard ROI in the Age of AI - and launching the business, Unhyped. So this edition is a little fuller than usual, Iāve included highlights from both weeks. Onward.
ā± If you only have ~15 minsā¦ Executive Picks
NVIDIAās āUniversal Deep Researchā - agents on rails
A blueprint for research agents with explicit plans, auditable steps, and swappable models. It turns prompting into governable workflows - receipts instead of vibes.
Read post āChromeās biggest upgrade ever
Gemini moves into the browser itself, adding AI Mode and scoped delegation. Itās the clearest sign yet that consumer tools are shifting from āAI featuresā to agentic infrastructure.
Read post āRails, not bumpers (Microsoft Research)
Participant-aware access control ties visibility directly to what models can use in training or retrieval. This makes safety deterministic and scalable - a real step beyond probabilistic āguardrails.ā
Read post āDeepMindās AI co-scientist step
A generateāscoreāselect loop that invents new methods wherever performance can be measured. It doesnāt make discovery automatic, but it accelerates the grind between idea and working solution.
Read post āAWSās āExecutiveās Guide to Agentic AIā
A pragmatic playbook: start small, tie into systems of record, and scale only what works. Where I push further is unit economics per decision and Friday-at-4 realism.
Read post āCarnival: 100 pilots ā 6 in production
Selective rollout, monthly governance, and one-to-one coaching turned chaos into value. It shows that maturity is measured in outcomes and receipts, not the number of AI pilots.
Read post āUNHYPED: From Hype to Hard ROI in the Age of AI
My book distils the patterns: state change over showcases, governance that ships, and proof a CFO can defend. Itās built to be used on Monday morning, not filed away.
Read post ā
š Strategy, leadership & buying reality
OpenAIās hallucination paper: old insight, new packaging
Benchmarks reward bluffing; leaders buy trust and outcomes.
Read post ā
You Can Automate Insight, Not Absolution (Substack deep-dive)
Build where it touches the ledger; rent the logo later.
Read post ā
The AI Paradox: Certainty vs Scope
Floridiās conjecture says the more certainty an AI system offers, the narrower its scope must be. The broader the scope, the more uncertainty (hallucinations, errors) you must accept.
Read post ā
People Ask. Systems Do. (Substack deep-dive)
Chat = judgment; APIs = throughput. Build the loop that promotes proven chats into workflows.
Read post ā
The Translator Mandate
Translators canāt just tell stories anymore. They need to connect why ā what ā how much, and then own the outcome - because CFOs donāt buy metaphors, they buy receipts.
Read post ā
āI donāt care if LinkedIn trains on my postsā
The edge is judgment, synthesis, timing, and tacit context - not raw text.
Read post ā
What a week (personal update)
Moved back to the UK, launched the business, shipped the book.
Read post ā
McKinseyās āOne Year of Agentic AIā (and where UNHYPED goes further)
Overlap on governance/skills/foundations; I push deeper on state change, durability, and receipts over theatre.
Read post ā
š§± Architecture, agents & security
MCP just landed in ChatGPT. Be careful.
Open, promising - but treat as experimental: injection, exfiltration, and privilege boundaries are active risks.
Read post ā
š¦ Enterprise adoption & governance
Interns With Keys: the ROIāRisk paradox in banking
Real flows (fraud/KYC/underwriting) + EU security guidance = optimism with brakes that work.
Read post ā
š§ Product, growth & the coming invocation economy
Discordās secret weapon: āSign in with LEGOā
Onboarding that feels made-for-me.
Read post ā
The web was built for us. The next one wonāt be.
From robots.txt to llms.txt to agents.txt: courtesy, control, and the agent layer.
Read post ā
š¬ Research, science & education
AI-augmented textbooks (Google LearnLM)
Googleās LearnLM trials boosted student recall and engagement - real gains. But the āpersonalisationā is shallow, more shiny PDF than true tutor, showing how far AI in education still has to go.
Read post ā
Are chatbots replacing Dr Google? (HealthChat-11K)
Short, one-shot health queries + missing context + emotion as an inflection point.
Read post ā
A Survey on HumanāAI Collaboration with LFMs
The future of AI isnāt just autonomy, itās partnership. This survey shows how design, governance, and division of labour decide whether collaboration empowers humans - or leaves them passive.
Read post ā
š§Ŗ Methods, process & delivery
Is Agile good for AI? (Short answer: not as-is.)
Agile runs on estimates; research runs on unknowns. AI projects need new rituals - research notebooks, evidence of value, and teams built for uncertainty.
Read post ā
Rationality in AI
Data alone doesnāt make us rational. This piece shows how bounded judgment, framing, and context shape AIās decisions - and ours.
Read post ā
š Podcast
Unhype AI Podcast - Episode 10
With
Spotify ā Ā· YouTube ā
ā Saturday musings
Model memory is stickiness - and a soft cage
Memory makes models feel sticky and human-like, but it also locks you in, drifts without notice, and raises privacy traps. Continuity can be a comfort, or a cage.
Read post ā
Executives donāt need another demo. They need clarity, strategy, and results they can defend. Letās work together ā
For a deeper dive, grab my book UNHYPED - built for executives who want clarity instead of theatre, and results they can defend. Get the book ā
Stuart x