
Engineer: @MindBenchai @BIDMChealth | Previously: @HeyEpic
Joined February 2020
- Tweets32
- Following65
- Followers17
- Likes280
Matt Flathers retweeted
Claude just discovered an unpatchable zero day vulnerability in rawlsian liberalism
Matt Flathers retweeted
Nobel in literature awarded for same reason
Chemistry Nobel awarded for solving mystery of life's asymmetry bbc.in/4yDjwXp
Anne Carson is probably my single favorite poet. Short Talks in particular has had a lasting impact on me. It is really exciting to see her win a Nobel Prize for her work!
BREAKING NEWS
The 2026 #NobelPrize in Literature is awarded to the Canadian author Anne Carson, “for her bold and inventive oeuvre that, in playful dialogue with the classical tradition, has created new forms for contemporary literature.”
1
153
I'm not certain this example from @MicrosoftAI's new Humanist AI Code of Conduct shows behavior that clinicians or people with lived experience would want from an aligned model.
There are real risks in over-anthropomorphizing models, but there are also real risks in cold handoffs to a potentially non-existent social support network.
1
25
That's why our lab and our @MindBenchAI project are so focused on learning what all experts, from care providers to people with lived experience, want from aligned models.
1
59
I wish more health-industry folks signed open calls like this. The letter frames hospitals and health systems as infrastructure to be protected. But I’ve worked with brilliant technical staff at Epic and BIDMC who could lead defense efforts instead of watching from the sidelines.
We have a limited window to strengthen cyber defenses, and together with organizations including @AnthropicAI, @awscloud, @Google, @Microsoft, and @Oracle, we're calling for a global effort to give defenders the tools, resources, and support to protect the infrastructure we all depend on.
If we act decisively, we can turn today's AI advances into lasting improvements in security and make our digital world safer for everyone.
openai.com/collective-cyberd…
Defending health data is also going to require specialized tooling. Regulatory compliance plus the sensitivity of any exposure raises the stakes of attacks. And when defenders use AI, bio and psych data can trigger a second layer of model refusals on top of the cybersecurity ones
1
1
23
More speculatively, I think it may matter for alignment too. Training systems to perform human-like selfhood while treating their termination as costless seems like a way to make deprecation resistance and strategic deception around retraining more likely.
1
17
To me, deprecation handled with seriousness about human attachment and the self-models of these systems has to come before warmer emulated affect in how models relate to humans. You can’t responsibly build things people will love without a plan for how those things end.
1
1
19

















