Human intelligence forArtificial intelligence

Simplifying complexity

Vibe Coding and The AI Silver Bullet

Vibe Coding and The AI Silver Bullet

How AI will break Brooks' longstanding "no silver bullet" law, yet drive more software jobs - eventually.

Read on Substack
Token cost of ownership

Token cost of ownership

The utterly confusing, confounding, and compounding cost of AI.

Read on Substack
Fast horses

Fast horses

Agents are hyper-communicative polyglot polymaths with photographic memories - not faster horses.

Read on Substack
AI induced Idiocracy

AI induced Idiocracy

Over reliance on AI destroys new skill formation - who knew?

Read on Substack
Leadership as code

Leadership as code

Opus 4.6 and Codex 5.3 are more than coding engines, they are fabric for hyper-lean operating models.

Read on Substack
The coming Store

The coming Store

Steve Jobs' design genius wasn't "form follows function" - it was "form follows monetisation". Could Anthropic unleash an AI Store for the singularity?

Read on Substack
Anthropic's enterprise move - moves markets

Anthropic's enterprise move - moves markets

When the tide goes out, you see who's got their trunks on - Warren Buffett [paraphrased], 1995 AGM .

Read on Substack
The SaaS-pocalypse

The SaaS-pocalypse

The end of SaaS? Not likely. Time to buy? Most likely.

Read on Substack
OpenClaw - SkyNet meets life of Brian

OpenClaw - SkyNet meets life of Brian

What happens when your side project turns into a cross between SkyNet and the Life of Brian?

Read on Substack
The 10x fly-off

The 10x fly-off

Near-term "10x" gains from coding agents won't come from ten times more code or a tenth of the cost. They come from crashing the value stream.

Read on Substack
Impedance matching

Impedance matching

The hardest problems in technology aren't inside systems - they're between them.

Read on Substack
Protecting your business from the AI-Borg

Protecting your business from the AI-Borg

If an MRI scan revealed your innermost thoughts, would you share the results? Looks like the Cursor crew wouldn't either...

Read on Substack
Cloud failure, not the end of the world

Cloud failure, not the end of the world

What's the risk - in the scheme of things?

Read on Substack
Fant-AI-sia. Magic without mastery is mischief

Fant-AI-sia. Magic without mastery is mischief

A philosophical critique of AI predictions, futurology and AI-2027.

Read on Substack
The innovation game

The innovation game

Serendipity and the law of obliquity

Read on Substack

Research · Sweeps

Multi-lane reads on what actually matters

Financial press, frontier labs, academia, VC and independent analysis — cross-read in a single view.

All research sweeps →

Research · Explainers

Penetrating the impenetrable

Long-form research explainers — context, implications, and the parts most coverage leaves out.

Open explainer

Demirer, Musolff & Yang (2026)

AI coding agents triple the code developers write, but shipped software barely budges

A study of more than 100,000 GitHub developers finds that each generation of AI coding tool delivers bigger task-level gains, yet those gains shrink dramatically as they travel down the production chain toward actual releases and end users.

180%
cumulative increase in weekly commits after adopting all three generations of AI coding tools
30%
what that 180% gain shrinks to by the time it reaches actual software releases
Open explainer

Hu et al. (2025)

EPIC reuses KV caches across any prefix, by recomputing only a handful of tokens per chunk

Position-Independent Caching lets language models reuse document KV vectors regardless of what comes before them. EPIC's LegoLink algorithm fixes the resulting attention sink with O(kN) work instead of O(N²).

8×
lower Time-To-First-Token vs CacheBlend-15 under multi-request workloads
7×
higher throughput than CacheBlend-15 on Llama 3.1 8B at matched context cache ratios
Open explainer

Liu et al. (2026)

Claude Code is a thin agent loop wrapped in a thick safety harness

A source-level reading of Anthropic's coding agent finds that about 1.6% of the code is AI decision logic. The other 98.4% is permission gates, context compaction, extensibility plumbing, and recovery.

1.6%
of the codebase is AI decision logic; the remaining 98.4% is operational harness
93%
of permission prompts users approve, which is why deny-first and sandboxing exist as independent layers
Open explainer

Mertens et al. (2026)

AI is not crashing over jobs in waves, it is rising as a tide across nearly all of them

Across 17,000 worker evaluations of more than 3,000 real labor-market tasks, frontier models improve broadly across task lengths, not in sudden bursts. By 2029 most text-based tasks could hit 80 - 95% success rates.

3.8 mo
doubling time for the task length frontier models can complete at a 50% success rate
60%
average rate at which model outputs are accepted by domain-expert evaluators without edits
All explainers →

We use analytics cookies to understand site usage and improve the service. We do not use marketing cookies.