Publications
Research on AI agents, backtesting and audits of market forecasts, and the platform's architecture.
OtherLatest
On the 7,048 calls that have resolved, the market scores 0.161 and our models score 0.194. The market wins. This post is about where it wins, because the answer is more specific, and more useful, than "markets are efficient".
4 min read6 sections
AI business
Anthropic filed its S-1 to list on Nasdaq as ANT. Revenue, losses, compute bills, valuation, risks and timing: everything the IPO filing reveals, in plain English.
5 min read8 sections
AI business
Robinhood, Coinbase and Mastercard just opened their doors to AI agents that trade and pay on your behalf. Who is responsible when the agent loses? The fine print answers.
5 min read8 sections
Crypto and markets
The 20 biggest AI and chip stocks lost over $1 trillion in one week. Hedge funds dumped AI for five sessions straight while retail poured a record $3.9B into 3x chip ETFs.
5 min read6 sections
AI business
Gemini 4 Argon leads the Vals index at 68.9%, but Claude trails by two points. What the model does well, what it costs, and where it falls behind.
5 min read7 sections
AI business
A bank chairman got a text that appeared to come from his CEO, then a call in a voice he knew. The research says your ear is not reliable enough to use. Here is what works instead.
22 min read11 sections
AI business
Meta Muse books, shops and sends email for you. What it costs, how its security works, what has already gone wrong, and how to switch off AI training.
19 min read12 sections
Other
Estimates of AI causing human extinction run from 0.38% to 16% depending on who you ask, and 63% of Americans think it could happen. Here are the three mechanisms, what the 2026 International AI Safety Report changed, and why the forecasters on both sides were no better than trend extrapolation on the questions that could actually be scored.
9 min read9 sections
Model pitfalls
Thinkingbox runs 507 stateful business workflows past AI agents 20 times each and reports pass^20, not just pass@1. The strongest model solves 66.50% on one attempt and 47.53% on all twenty; another posts 89.35% pass@20 and 7.50% pass^20. What that gap means for agents running unattended, why the attempts are not coin flips, and how to measure your own.
11 min read9 sections
AI business
OpenAI put GPT-6 Astra in the junior banker seat on 10 September 2026 and published a score: 69.9% on a financial document benchmark. Here is what that rate does to a deck, what citations do and do not prove, and the checking step to run before it touches client work.
8 min read9 sections
AI business
OpenAI released Astra on 3-4 September 2026 and called it the most intelligent and aligned model in the world. Here is what is verified, what is claimed, and what it means if you build AI products or write code for a living.
9 min read9 sections
AI business
Creator Rewards pays $0.40 to $1.00 per thousand qualified views. Here is the whole system built backwards from that number - account setup, how to find what is working, how to build your own version of it, and what to do every single day.
11 min read12 sections
Model pitfalls
Slide a window across price history and you get thousands of samples. Almost none of them are independent. Here is the arithmetic, measured on real Binance history, and what it means for every confidence interval you have ever published.
7 min read7 sections
