Frameworks and honest lessons from enterprise AI.
Longer arguments and thesis pieces.
AI is becoming more expensive to create and cheaper to use. The next advantage will come from turning the same machine intelligence into more verified value.
Why the next software war will be won at the improvement layer.
Top SLMs are already capable enough for a growing share of enterprise workflows. The next moat is the system that gives those models the right goal, memory, workflow state, constraints, and validation at each step.
Why renting AI will cost more than owning it. The case for enterprise-owned intelligence — from sovereignty to compounding advantage.
Shorter pieces that started on LinkedIn.
A new paper reframes hallucination and overconfidence as symptoms of unbounded autonomy — and proposes a four-state framework (SMARt) for governing when agents can and cannot act.
WorkstreamBench tested agents on real financial modeling tasks. Claude's general web interface scored 69.1. Its specialized Excel add-in scored 60.4. Same model, different harness, 3.8× gap.
SKILL0 trains small models with curated skills then progressively withdraws them. The result: a 3B model scores 87.9% where GPT-4o scores 48.0%, at 5× lower inference cost.
The OpenAI Deployment Company launched with $4B, 19 investors including McKinsey and Bain, and 150 engineers from Tomoro. The model provider just ate the consulting layer.
NBER data shows 90% of firms saw no measurable productivity change from AI. Gartner shows zero correlation between AI layoffs and AI ROI. The layoff narrative is not connected to the results.
SAP Sapphire 2026: 200+ AI agents, Anthropic Claude powering HR/procurement/supply chain, and a bet that governance — not models — is the winning layer.
Heuristic Learning uses coding agents to maintain and improve software-based policies through feedback loops — matching Deep RL baselines without training a single neural network.
1,308 models trained to show that the Chinchilla rule of ~20 tokens per parameter is wrong. The universal rule is ~60 bytes per parameter.
Agent-World builds 2,000 realistic training environments with 19,822 tools — and shows environment diversity beats parameter scaling for agentic tasks.
Monte Carlo's survey of 260 enterprise leaders reveals: 63% found unauthorized data access, 36% can't roll back a failing agent, 70% expect to rebuild systems they already shipped.
Google Cloud Next 2026: Vertex AI becomes Gemini Enterprise Agent Platform. A2A protocol at 150 orgs in production. 200+ models including Anthropic Claude. A Google insider's perspective.
A new paper unifies agent memory, skill discovery, and rule learning into a single Experience Compression Spectrum — and shows agents with Level 2 skills outperform Level 1 memory by 68.5 percentage points.
DeepSeek shipped V4 the same day as GPT-5.5. 1.6T parameters, 1M context, Apache 2.0, $3.48 per million output tokens. Sparse attention cuts KV cache to 10% of V3.2.
The largest systematic comparison of agentic AI frameworks reveals that engineering discipline — memory management, retry policies, failure handling — matters more than architecture.
A new paper shows 84.6% token savings through knowledge compounding. The insight: tokens should be capital investments, not consumables.
A new paper shows agents improving themselves without model updates. But the bigger insight is the full optimization surface most enterprises aren't using.
A new paper tested when to collapse multi-agent systems into a single agent. The answer is counterintuitive: it's not about the task. It's about the metric.
When your Copilot agent sends an email, it does it under your identity. Your credentials. Your permissions. Microsoft's Agent 365 won't fix this until December 2026.
Jensen Huang told Lex Fridman he thinks we've achieved AGI. The same week, ARC-AGI-3 launched. Humans score 100%. Best frontier AI scored 0.26%. Both are right.
The AVO paper gave a coding agent access to CUDA docs and Blackwell B200 specs. It produced 40 kernel versions in 7 days — outperforming cuDNN by 3.5% and FlashAttention-4 by 10.5%.
Andrej Karpathy hasn't typed a line of code since December. He spends his days expressing his will to agents. The bottleneck just changed.
This is the role I wish existed when I was still at Google: applied research close enough to production that the work actually ships.
If the agent runtime is Linux, then the runtime is a commodity. Every enterprise will run the same one. The value migrates up — to the intelligence layer.
Their new 'proprietary' model was Kimi K2.5, an open-weight model from Moonshot AI, fine-tuned with RL. A developer found it in 24 hours. The internet called it a scandal.
Midjourney runs on models they own. Their inference costs drop every quarter. Your enterprise runs on Claude and GPT. Your costs go up with usage. Your moat is zero.
Donald Knuth, the most respected computer scientist alive, was stunned when Claude solved an open problem he'd been working on for weeks. The implications go beyond math.
Andy Grove wrote the most important sentence in management. In 2026, that equation quietly changed. Your organization now includes AI agents.
We just gave an AI agent unrestricted CLI access to our staging environment. The security team found out from a Slack alert — not a governance review.
You have deployed AI. It works. But it does not compound. That is not Stage 3. That is Stage 2 — and most enterprises are stuck there.
The technology we built to replace offshore labor is about to cost more than offshore labor. A Gartner prediction that should worry every enterprise AI buyer.
In 2024, we built drag-and-drop agent builders. In 2026, developers just describe what they want. The config UI era lasted 18 months.
One lesson about leadership has stayed with me throughout my career. Leadership is not about doing one thing. It is about making one team.
Everything they have done in 2026: $20B revenue run rate, number one on iOS, the Pentagon. A breakdown of how Anthropic is winning — and the risks ahead.
I publish on LinkedIn and update this site as I build.