I am delighted to be keynoting the IR-for-good track at @ecir2026.eu
I will be talking about what I learned from studying generative AI's use during the biggest election year in history. (Spoiler: the problem is not just deepfakes.)
The talk is called gemini hegemony, expect a little 🌶️
For AI to have positive effects in academia specifically, though, researcher incentives will definitely have to change — knowledge frontier expansion is exactly where AI does worst (see phys.org/news/2026-01...)
I once worked for a prof whose entire PhD was on digitizing books of economic info. As data science became more accessible, a lot of it moved to journalism and OSINT.
One idea: fewer RAs, more community AI clinics (the law school model) to help civil society do the research they couldn't before.
Anthropic screens its hires on values, so Dario has to speak to a workforce that joined because they explicitly believe in building safe AI. That his viewpoint was sent as a staff memo seems like a key detail in that context.
...So now the moral onus sits with the warfighter workforce instead of the AI one.
And then there is the OpenAI response. Anthropic drew an explicit line against fully autonomous systems. Sam put a human in the loop. That's what any good scholar of sociotechnical systems would recognize as a "moral crumple zone"---someone to blame when autonomous systems go awry.
A really high fraction of Anthropic's workforce has safety-aligned values because they've been screening for it since day one. Whatever your perspective on EA ideology, this is a wildly interesting case of workforce values clashing with state power.
Hegseth calling it "defective altruism" is the strangest and most telling moment in this whole saga.
When you interview for Anthropic, the most important question is "why anthropic?" It's a culture screener: the answer is safety, alignment, being one of the good guys.
Hegseth calling it "defective altruism" is funny but also pretty telling - anthropic screens every hire on "why anthropic" and the answer is its safety culture and EA-adjacent values. This is a direct attack on a big part of the bay's techie culture. I can't wait to read your follow-up.
up next, agent swarms (long-term repeat interactions, with memory)
If you're trying to build successful and pro-social agent swarms, here are four lessons from moltbook (...on what *not* to do)
lol are u a bot u kinda write like a bot?
I had an editor at a respectable social science journal sit on a paper about AI adoption (written in early 2024!) for seven months. When I emailed about it he said they had only been able to get one reviewer and I should send it somewhere else.
Now I mostly publish in computer science conferences.
might you be able to write about how you are thinking about tradeoffs between comprehensiveness and simplicity? i don't yet have a good mental model for this in agentic engineering
Huh, this is useful for thinking about the moltbook spectacle. "Look everyone, the agents are inventing secret languages" <- a bunch more humans set their agents to invent secret languages.
I know this is silliness but I'm convinced that fact that all the top pro-AI people are on X is a big factor in why alignment is going so badly
for data science, may i recommend hub.decision.ai - learned about it from the @pymc-labs.bsky.social folks, incredible modeling and stats skills for big improvements in accuracy
Are other people's AIs really into Ursula K. Le Guin and Borges or is that just mine?
I wonder if we will look back on the social media era in the same way we now look back on hitch-hiking - what a wild thing we all put our brains through
most fears about advanced technology are actually fears about capitalism, to paraphrase ted chiang
as it turns out, capitalism + frontier tech is indeed a scary combo
buckle up everyone, it's gonna be a wild decade.
Here's hoping some of the digital/plurality ideas start finding their way into builder/policymaker culture on this side of the Pacific. The future of democracy might depend on it.
A couple years ago, @audreyt.org and @glenweyl.bsky.social and a big collective of brilliant people published the book "Plurality: the future of collaborative technology and democracy". Takahiro Anno, founder of Team Mirai, read it before running for governor of Tokyo.
But who is influencing the builders? There's a tell in the NYT article: a whiteboard with what sure looks like @audreyt.org plurality poem
Nothing is more important in shaping the future of tech than the builders themselves---so it's fascinating to see the unexpected electoral success of Japan's Team Mirai, a party of software engineers who are using AI-powered transparency and deliberation tools
www.nytimes.com/2026/02/22/w...
In case you are curious why AI feels so societally misaligned these days, remember that all the Important Model Builders have decided to continue drenching their brains in the twitter
5 PRs from an AI agent that fixed all 500 issues in 5 minutes
Is this really so different from contracting chegg? I think we all knew this was coming when gig-work-for-homework was the first industry to be decimated by chatgpt.
On the bright side at least it's not just the rich kids cheating anymore. Democratization! /s
"I have a strong personal policy of not publishing AI-generated writing under my own name." <- same. in fact i've been enjoying the challenge of trying to make my words as human as possible. ty for the red/green tdd tip, excited to try!
I set my openclaw to give me a six-month learning path with regular spaced-repetition quizzes. Learning with it has been a true delight. Maybe the problem isn't the openclaws, but the systems within which we are deploying them...
What does the social network of bluseky look like? @theo.io maps 3.4m users.
Found via @flowingdata.com, which has been my source for the best data visualizations on the internet for nearly a decade
