I study (and teach) how people behave and how they impact others, organizations and markets.
United States
Joined November 2008
- Tweets14.5K
- Following73
- Followers20.1K
- Likes1.8K
SlingMD frees your Outlook email: Markdown for Obsidian or AI tasks, attachments, threads and follow-up tasks. Classic Outlook on Windows.
I’m not involved and don’t know the team. I bought them a coffee.
SlingMD: l.sslevine.com/RP4A
Coffee: l.sslevine.com/T0l2
This is the paper I've always wanted to write, now in PNAS.
In our experiments, exchange collapsed when the best partners were of another race. A brief intervention, helping people speak as a team and state needs and offers, closed the gap.
@PNASNews pnas.org/doi/10.1073/pnas.25…
'I Teach at a Business School. It’s Time to Shut Them Down,' writes Chris Marquis.
l.sslevine.com/gi7Q
This is the paper I've always wanted to write, now in PNAS.
In our experiments, exchange collapsed when the best partners were of another race. A brief intervention, helping people speak as a team and state needs and offers, closed the gap.
@PNASNews pnas.org/doi/10.1073/pnas.25…
On October 17, 1933, Einstein arrived in America as a refugee. He never returned to Europe. He already had a Nobel Prize. As we debate admitting foreign scholars, remember: most promising researchers arrive before fame does. ias.edu/about/mission-histor…
Making gender diversity work won us a Best Paper Prize! A costly choice of team helped identify cooperative men and form teams that performed well. Our team: two men, two women, three continents, ERC and NSF grants. We took the research question personally.
Making gender diversity work won us a Best Paper Prize! A costly choice of team helped identify cooperative men and form teams that performed well. Our team: two men, two women, three continents, ERC and NSF grants. We took the research question personally.
'I Teach at a Business School. It’s Time to Shut Them Down,' writes Chris Marquis.
l.sslevine.com/gi7Q
On October 17, 1933, Einstein arrived in America as a refugee. He never returned to Europe. He already had a Nobel Prize. As we debate admitting foreign scholars, remember: most promising researchers arrive before fame does. ias.edu/about/mission-histor…
An agent’s output wasn’t what I expected. I asked it to retrace its steps, unsure whether it had erred or why. It found the mistake, admitted it, and corrected the work. Agents can err. They can also admit error and repair their own work. That’s worth testing.
Twenty-seven words, one equation, one number, one reference. Is Friedrich Lenz’s 1951 paper the shortest peer-reviewed publication? What's more, it is a marvel of scholarly restraint: knowing where an observation ends can be as valuable as knowing where a theory begins.
An agent’s output wasn’t what I expected. I asked it to retrace its steps, unsure whether it had erred or why. It found the mistake, admitted it, and corrected the work. Agents can err. They can also admit error and repair their own work. That’s worth testing.
𝐃𝐨 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭𝐬 𝐤𝐧𝐨𝐰 𝐰𝐡𝐚𝐭 𝐭𝐡𝐞𝐲 𝐤𝐧𝐨𝐰? Sol could not know why Terra failed. It could diagnose the visible failure and repair the process. Not proof of awareness—perhaps a useful form of self-audit.
Twenty-seven words, one equation, one number, one reference. Is Friedrich Lenz’s 1951 paper the shortest peer-reviewed publication? What's more, it is a marvel of scholarly restraint: knowing where an observation ends can be as valuable as knowing where a theory begins.
𝐃𝐨 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭𝐬 𝐤𝐧𝐨𝐰 𝐰𝐡𝐚𝐭 𝐭𝐡𝐞𝐲 𝐤𝐧𝐨𝐰? Sol could not know why Terra failed. It could diagnose the visible failure and repair the process. Not proof of awareness—perhaps a useful form of self-audit.
𝐓𝐡𝐞 𝐋𝐋𝐌 𝐨𝐫𝐚𝐜𝐥𝐞. GPT-4 at r = 0.92, matching human forecasters. Holds for studies never in the training corpus — ruling out memorization. 2x overestimation. Same tool screens anti-vax messaging; first-order guardrails don't stop it. l.sslevine.com/VdoD
𝐈 𝐩𝐥𝐚𝐲 𝐭𝐡𝐞 𝐟𝐢𝐞𝐥𝐝. Preference is not a methodology. We've always run models side-by-side. Today, GPT-5.6 Sol joins the stack. The real question is never which model wins; it's what they accomplish when you stop asking them to compete.
𝐓𝐡𝐞 𝐋𝐋𝐌 𝐨𝐫𝐚𝐜𝐥𝐞. GPT-4 at r = 0.92, matching human forecasters. Holds for studies never in the training corpus — ruling out memorization. 2x overestimation. Same tool screens anti-vax messaging; first-order guardrails don't stop it. l.sslevine.com/VdoD
