Research & perspectives.
Analysis of AI, markets and business, alongside lessons from our own work.
AI discovery makes choosing the next experiment more important
Anthropic’s enzyme-system finding raises a practical question for research software: which proposals deserve a scientist’s time?
READCalifornia wants data centres to pay their way
Seven new laws put electricity, water and local approvals into sharper focus. AI infrastructure should be judged on the cost and certainty of getting it into operation.
READHarvey and the cost of useful AI
Specialised models can make AI cheaper to run. Buyers need to know whether the savings survive the time spent reviewing the work.
READWhat Anthropic’s 26% tells us about AI research
A new internal index makes automation more measurable. Its denominator matters as much as its headline.
READAustralia’s AI buildout will be decided project by project
PwC’s US$703 billion forecast puts a number on the opportunity. Power, equipment and local consent determine what can be built.
READNvidia’s $96.2 billion quarter leaves a second question open
The hardware supplier’s accounts establish the scale of spending. Buyers still have to earn a return on it.
READBefore owning a model, own the test
Sequoia’s August argument for specialised models raises a more immediate question: can the company measure an improvement?
READGoogle is putting the assignment inside the prompt
The Classroom update shows how an existing product can make an AI feature easier to start using.
READBefore an agent can act, the company has to agree what its data means
Connecting systems leaves a harder question unresolved: which definition should the software use?
READAn agent’s job description includes its permissions
Google’s enterprise guide makes a practical question unavoidable: what exactly has the software been authorised to do?
READEverything you need to know about Mythos
Anthropic just shipped the most consequential model release of the year: a new tier above Opus, split into a public model and a restricted one. The benchmarks, the safeguards, the backlash, the fine print, and our own one-prompt test. The whole picture, in one place.
READThe pendulum swing
For most of the last two quarters Anthropic owned the coding frontier. Two weeks ago the new Codex landed. Eight days ago GPT 5.5 followed. We spent the week working on both, and the verdict in our hands is unambiguous.
READWe tried Opus 4.7 so you didn't have to
Four days with Anthropic's new flagship, a week of our real workload, and a mountain of token receipts. Here is what is actually different, what is worth the money, and what to turn off before your bill arrives.
READClaude Mythos and the quiet arrival of the next regime
Anthropic just announced their latest model. It is so good, and so dangerous, that they decided not to release it to the public. Most teams are not ready for what comes next.
READWe built an AI fuel price oracle during Australia's worst fuel crisis
Petrol hit $2.60 in Sydney. Over 600 stations ran dry. Every fuel app tells you the price right now. None of them tell you what it will be tomorrow. So we built one.
READHow we built the output of a $100k team for $1k
We kept seeing people talk about AI agents like they were either magic or useless. The better question was simpler: how much useful work can you get from a small number of well-scoped agents, if you design the system right?
READWe built a bot that printed money on Polymarket's broken binary markets
How we scraped Polymarket's stale price feed and blended it with live exchange data to exploit mispriced crypto UP/DOWN markets, before the rest of the market caught up.
READWant this in your inbox?
GET IN TOUCH