Lots of great research predicting the effect of AI on wages, unemployment, the labor income share and more. This first microsimulation experiment shows the possibilities of translating those market effects into household resources through public policy.
CIO can prep some reliable examples (which is probably most examples at this point with the best models)
Check out the tool Representative McDonald Rivet cited when releasing the bill: policyengine.org/us/working-parents-tax-relief-act
Slides from my panel discussion at the CIO4Good conference: policyengine.org/slides/cio4good-talk
Slides from my workshop, co-facilitated with @brhkim.bsky.social: policyengine.org/slides/cio4good-workshop
Mission slide
If I were a CIO slide
How it started on AI slide
Agentic workflow diagram
Yesterday I told 70 nonprofit CIOs to lock their CEO in a room and make them use coding agents.
AI gives enormous new resources to pursue a mission. No other way to internalize it and expand your thinking of what's possible.
Max in front of the AI for the Economy Forum poster
Great conversations with researchers and technologists at the inaugural Google / MIT FutureTech AI for the Economy Forum today. Big opportunity uniting approaches to accurately model the impact of AI on the economy, and the impact of public policy in shaping that relationship.
Alumni Notes section including PolicyEngine
Grateful to the Berkeley IEOR magazine for sharing our work localizing evidence-based policymaking @policyengine.org
ieor.berkeley.edu/wp-content/u...
straude.com/leaderboard
Straude leaderboard
Hooked up Straude today and discovered I'm #1 this week globally — $20K in token volume at list prices, 3.3× ahead of #2. About 40B tokens in 7 days between Codex and Claude Code.
Here @jasondebacker.bsky.social, @roxanagonzalez.bsky.social and I elicit LLMs' views on how people respond to tax policy by assigning them personas. As we increasingly consult AI for policy advice, their assumptions on these dynamics will alter predictions. Much more to learn along this frontier.
OpenAI's new document for these tax and transfer policies:
- Higher taxes on capital gains and corporate income
- Taxing sustained AI-driven returns and automated labor
- Wage-linked incentives
- A Public Wealth Fund distributed directly to citizens
- Automatic stabilizers
openai.com/index/indust...
Image of Max Ghenis at CIO4Good with session descriptions
Speaking at the CIO4Good Summit next Tuesday in Arlington — two sessions on AI in nonprofits.
9am panel on building data + AI foundations: governance, culture change, getting AI working across an organization.
10:15am hands-on workshop on AI-driven software development for nonprofit teams.
Great to see this from the Living Wage Institute team. I met their co-founder in NYC last week and learned more about their work giving employers a clearer picture of taxes and credits affecting their hourly workforce. Glad to see PolicyEngine powering the tax calculations behind it.
Built an interactive viz of Dell (2010) on Peru's mining mita — colonial forced labor abolished 200+ years ago that still shows up in child stunting, consumption, and road density today.
The map morphs into a scatter plot as you scroll. Districts become data points.
maxghenis.com/mita
To clarify I mean OpenMessage has no servers, it operates on your own machine so you can use the AI tools of your choice with its MCP. Agree that getting it working with a local LLM would be killer.
OpenMessage desktop app showing a group chat with an AI-drafted reply, alongside a terminal where Claude uses MCP tools to read conversations and compose messages.
In February, I launched OpenMessage — an open source local AI-connected MacOS app for Google Messages.
Just redesigned it and added WhatsApp support. Now my daily driver for both chat surfaces.
Free and open source: openmessage.ai
Berkeley and Hoover economists make several diverging assumptions around California's billionaire tax, resulting in dramatically different revenue estimates. We've put them all in a tool you can explore - updating daily from the Forbes list.
We looked at every general-purpose chip on the market, and none of them could compute a marginal tax rate fast enough. So we made our own.
farness is free, open source (MIT), and works anywhere: Claude Code plugin, Codex agent, MCP server, Python library, CLI.
If your agent gives advice, make it forecast first.
farness.ai
But Study 2 showed the advantage mostly disappears on held-out probes. Honest null result included.
farness.ai/paper
Also wrote a paper on the method behind farness: "stability-under-probing."
Probe an LLM's answer with challenges, measure how much it moves. Structured prompts move less -- they start closer to where both conditions end up.
Tested on Claude and GPT across 11 scenarios.
farness.ai/paper
Also used it to pick new bedsheets based on Whoop data.
KPIs: sleep efficiency delta, deep sleep delta, cost
My Sheets Rock: +1.5% efficiency, +1.0% deep sleep, $210
Brooklinen Percale: +1.0%, +0.5%, $179
Do nothing: +0%, +0%, $0
One of my 63 farness.ai decisions logged since December.
Another: @policyengine.org "Should we publicly announce our self-serve API option now?"
Broad launch: 10 activations, 9 support tickets
Soft launch: 6 activations, 4 tickets
Delay: 2 activations, 1 ticket
Soft launch won -- best activation-to-support ratio. farness made the tradeoff explicit.
Real example from @policyengine.org: "What frontend framework should we use?"
Vite SPA: 55% "things just work", 45% workaround burden
Next.js: 90% things just work, 15% workarounds
The decisive KPI wasn't dev velocity -- it was how often things worked without hacks.
farness.ai homepage
I built farness.ai — a decision framework for AI agents.
Instead of "should I do X?", it forces a forecast: KPI, confidence interval, base rate, disconfirming evidence, review date.
Works with Codex, Claude Code, and any agent that speaks MCP.
pip install farness
Claude Code now defaults to medium effort. You can persistently set that to high with `/effort high`, or change it to max for the current session with `/effort max`.
To set it persistently to max, add to ~/.claude/settings.json:
"env": {
"CLAUDE_CODE_EFFORT_LEVEL": "max"
}
As far as I can tell, this is the first time a congressional committee has built a public tool on an external open-source economic model. It's also the JEC's first public tool powered by microsimulation — their other interactive tools are data visualizations of government statistics.
Side by side charts using adaptive and snap125 ticks.
Ask an AI model to make a common chart in a React app, and it’ll pick Recharts (27k GH ⭐).
The tick marks bothered me, so I fixed them. That change is now live in the latest release.
Use `niceTicks="snap125"` to activate it.
My PR: github.com/recharts/rec...
Related: github.com/recharts/rec...
Ziming Hua speaking with MIT FutureTech
My colleague Ziming Hua also spoke about our multi-agent AI policy encoding system: policyengine.org/us/encode-policy-multi-agent-ai
