David Manheim
@davidmanheim.alter.org.il
Humanity's future can be amazing - let's make sure it is.
Visiting lecturer at the Technion, founder https://alter.org.il, Superforecaster, Pardee RAND graduate.
Visiting lecturer at the Technion, founder https://alter.org.il, Superforecaster, Pardee RAND graduate.
I'm here at #AIES2025, and still worry quite a lot about this.
The deep skepticism about AI systems ever being generally capable, or even human-level in specific domains, doesn't seem to have changed over the past few years.
The deep skepticism about AI systems ever being generally capable, or even human-level in specific domains, doesn't seem to have changed over the past few years.
October 22, 2025 at 11:32 AM
I'm here at #AIES2025, and still worry quite a lot about this.
The deep skepticism about AI systems ever being generally capable, or even human-level in specific domains, doesn't seem to have changed over the past few years.
The deep skepticism about AI systems ever being generally capable, or even human-level in specific domains, doesn't seem to have changed over the past few years.
September 7, 2025 at 9:30 AM
From the other site - not a complete explanation, but correct as far as it goes:
"the current configuration of economics/ wealth distribution is pretty solidly optimized to drive the wealthiest people in society batshit insane, which - to some extent - explains a lot of things you see around you"
"the current configuration of economics/ wealth distribution is pretty solidly optimized to drive the wealthiest people in society batshit insane, which - to some extent - explains a lot of things you see around you"
September 7, 2025 at 9:29 AM
From the other site - not a complete explanation, but correct as far as it goes:
"the current configuration of economics/ wealth distribution is pretty solidly optimized to drive the wealthiest people in society batshit insane, which - to some extent - explains a lot of things you see around you"
"the current configuration of economics/ wealth distribution is pretty solidly optimized to drive the wealthiest people in society batshit insane, which - to some extent - explains a lot of things you see around you"
...and finally:
September 3, 2025 at 7:19 AM
...and finally:
Looks like this didn't include articles 1-9.
September 3, 2025 at 7:18 AM
Looks like this didn't include articles 1-9.
We're clearly hitting a wall - because if AI progress is really exponential, why aren't the numbers going up faster?
Checkmate, AI industry!
Checkmate, AI industry!
September 2, 2025 at 9:39 AM
We're clearly hitting a wall - because if AI progress is really exponential, why aren't the numbers going up faster?
Checkmate, AI industry!
Checkmate, AI industry!
Yet another place where @garymarcus.bsky.social is right that LLMs don't have a correct world model.
(The filter exerts pressure on the water level. It doesn't understand that. But then, I'd bet most humans wouldn't realize this either.)
(The filter exerts pressure on the water level. It doesn't understand that. But then, I'd bet most humans wouldn't realize this either.)
August 31, 2025 at 7:59 AM
Yet another place where @garymarcus.bsky.social is right that LLMs don't have a correct world model.
(The filter exerts pressure on the water level. It doesn't understand that. But then, I'd bet most humans wouldn't realize this either.)
(The filter exerts pressure on the water level. It doesn't understand that. But then, I'd bet most humans wouldn't realize this either.)
Partly agree; multiple metrics can help, but the issues that make metrics fail aren't so simple, and the incoherence of chosen metrics is a deeper problem than just infighting, and multiple metrics doesn't necessarily help.
See the full article here: www.cell.com/patterns/ful...
See the full article here: www.cell.com/patterns/ful...
August 19, 2025 at 7:10 AM
Partly agree; multiple metrics can help, but the issues that make metrics fail aren't so simple, and the incoherence of chosen metrics is a deeper problem than just infighting, and multiple metrics doesn't necessarily help.
See the full article here: www.cell.com/patterns/ful...
See the full article here: www.cell.com/patterns/ful...
I'll fill that one out as soon as you do one for me... ;)
August 14, 2025 at 10:01 AM
I'll fill that one out as soon as you do one for me... ;)
If done well, regulation and liability for current medical AI could reduce mistakes and mitigate ethical concerns.
But if your concern reliably leads to more people being dead because doctors aren't using new technology, you're doing ethics wrong!
But if your concern reliably leads to more people being dead because doctors aren't using new technology, you're doing ethics wrong!
August 14, 2025 at 10:00 AM
If done well, regulation and liability for current medical AI could reduce mistakes and mitigate ethical concerns.
But if your concern reliably leads to more people being dead because doctors aren't using new technology, you're doing ethics wrong!
But if your concern reliably leads to more people being dead because doctors aren't using new technology, you're doing ethics wrong!
It no longer lets me add the column, so that's good for the future. I "fixed" it by doing "clear cache and logout" - which removed all columns, but didn't actually log me out.
But after that, it stays at the same URL - which does not work when loaded manually, so you need to navigate away.
But after that, it stays at the same URL - which does not work when loaded manually, so you need to navigate away.
August 14, 2025 at 7:48 AM
It no longer lets me add the column, so that's good for the future. I "fixed" it by doing "clear cache and logout" - which removed all columns, but didn't actually log me out.
But after that, it stays at the same URL - which does not work when loaded manually, so you need to navigate away.
But after that, it stays at the same URL - which does not work when loaded manually, so you need to navigate away.
Next requests:
- collapsable left bar
- allow customizing column widths, to fit more than 2 on the screen
Also, bug report - the column shown here is broken, and I can't remove/close it.
- collapsable left bar
- allow customizing column widths, to fit more than 2 on the screen
Also, bug report - the column shown here is broken, and I can't remove/close it.
August 13, 2025 at 9:19 PM
Next requests:
- collapsable left bar
- allow customizing column widths, to fit more than 2 on the screen
Also, bug report - the column shown here is broken, and I can't remove/close it.
- collapsable left bar
- allow customizing column widths, to fit more than 2 on the screen
Also, bug report - the column shown here is broken, and I can't remove/close it.
Sycophancy isn't really addressed by current OpenAI approaches.
This is seen, for example, contrasting types of errors found - not phrasing, but actually substantively different analysis, discussing different issues - when asking the two substantively identical below questions.
This is seen, for example, contrasting types of errors found - not phrasing, but actually substantively different analysis, discussing different issues - when asking the two substantively identical below questions.
August 12, 2025 at 7:36 AM
Sycophancy isn't really addressed by current OpenAI approaches.
This is seen, for example, contrasting types of errors found - not phrasing, but actually substantively different analysis, discussing different issues - when asking the two substantively identical below questions.
This is seen, for example, contrasting types of errors found - not phrasing, but actually substantively different analysis, discussing different issues - when asking the two substantively identical below questions.
Sounds like a co-founder of OpenAI is effectively publicly admitting they have no expectation that there will be meaningful human oversight and/or control of the systems they will deploy.
(Contra their recent commitment under the EU AI act; "AI-powered tools" are not meaningful *human* oversight.)
(Contra their recent commitment under the EU AI act; "AI-powered tools" are not meaningful *human* oversight.)
August 11, 2025 at 12:26 PM
Sounds like a co-founder of OpenAI is effectively publicly admitting they have no expectation that there will be meaningful human oversight and/or control of the systems they will deploy.
(Contra their recent commitment under the EU AI act; "AI-powered tools" are not meaningful *human* oversight.)
(Contra their recent commitment under the EU AI act; "AI-powered tools" are not meaningful *human* oversight.)
No, moving a specified ethnicity out of a region is actually definitionally ethnic cleansing. And I'm generally happy to defend Israel from ridiculous claims - but this isn't one of those.
ihl-databases.icrc.org/en/customary...
ihl-databases.icrc.org/en/customary...
July 19, 2025 at 6:27 PM
No, moving a specified ethnicity out of a region is actually definitionally ethnic cleansing. And I'm generally happy to defend Israel from ridiculous claims - but this isn't one of those.
ihl-databases.icrc.org/en/customary...
ihl-databases.icrc.org/en/customary...
Very happy to announce that my org has helped get commitment from the Israeli Ministry of Health to finalize plans to mandate salt iodization in Israel.
Following the meeting this past Wednesday, they have one month to present the final plan to the health committee. Wish us luck!
Following the meeting this past Wednesday, they have one month to present the final plan to the health committee. Wish us luck!
July 18, 2025 at 1:38 PM
Very happy to announce that my org has helped get commitment from the Israeli Ministry of Health to finalize plans to mandate salt iodization in Israel.
Following the meeting this past Wednesday, they have one month to present the final plan to the health committee. Wish us luck!
Following the meeting this past Wednesday, they have one month to present the final plan to the health committee. Wish us luck!
People think of respiratory disease as a fact of life, especially for kids - and assume that it doesn't really matter.
In fact, @meltemdaysal.bsky.social et al showed children's minor illnesses has significant impact for decades. Reducing "routine" disease is a big deal!
cepr.org/voxeu/column...
In fact, @meltemdaysal.bsky.social et al showed children's minor illnesses has significant impact for decades. Reducing "routine" disease is a big deal!
cepr.org/voxeu/column...
July 10, 2025 at 2:19 PM
People think of respiratory disease as a fact of life, especially for kids - and assume that it doesn't really matter.
In fact, @meltemdaysal.bsky.social et al showed children's minor illnesses has significant impact for decades. Reducing "routine" disease is a big deal!
cepr.org/voxeu/column...
In fact, @meltemdaysal.bsky.social et al showed children's minor illnesses has significant impact for decades. Reducing "routine" disease is a big deal!
cepr.org/voxeu/column...
Second, we also propose a maturity model for supervision, so we can suggest how firms can improve, or show when safety-washing occurs.
Level 1: vague safety claims + no docs = performative.
Level 5: lifecycle-wide supervision, transparent mitigations, and adaptive review.
Level 1: vague safety claims + no docs = performative.
Level 5: lifecycle-wide supervision, transparent mitigations, and adaptive review.
July 9, 2025 at 7:14 AM
Second, we also propose a maturity model for supervision, so we can suggest how firms can improve, or show when safety-washing occurs.
Level 1: vague safety claims + no docs = performative.
Level 5: lifecycle-wide supervision, transparent mitigations, and adaptive review.
Level 1: vague safety claims + no docs = performative.
Level 5: lifecycle-wide supervision, transparent mitigations, and adaptive review.
So what can be done?
First, we need to know what is being done - so we provide a schema for documenting supervision claims, linking risks to actual control or oversight strategies.
If someone says "oversight" without explaining all of this, they are irresponsibly safety-washing.
First, we need to know what is being done - so we provide a schema for documenting supervision claims, linking risks to actual control or oversight strategies.
If someone says "oversight" without explaining all of this, they are irresponsibly safety-washing.
July 9, 2025 at 7:14 AM
So what can be done?
First, we need to know what is being done - so we provide a schema for documenting supervision claims, linking risks to actual control or oversight strategies.
If someone says "oversight" without explaining all of this, they are irresponsibly safety-washing.
First, we need to know what is being done - so we provide a schema for documenting supervision claims, linking risks to actual control or oversight strategies.
If someone says "oversight" without explaining all of this, they are irresponsibly safety-washing.
One key point is that you can't just slap "human in the loop" on a system and call it safe.
If you do it naively, you'll get failures for oversight and control - like a rubber-stamp UI, an overwhelmed operator, and/or decisions too fast for humans to understand or fix.
If you do it naively, you'll get failures for oversight and control - like a rubber-stamp UI, an overwhelmed operator, and/or decisions too fast for humans to understand or fix.
July 9, 2025 at 7:12 AM
One key point is that you can't just slap "human in the loop" on a system and call it safe.
If you do it naively, you'll get failures for oversight and control - like a rubber-stamp UI, an overwhelmed operator, and/or decisions too fast for humans to understand or fix.
If you do it naively, you'll get failures for oversight and control - like a rubber-stamp UI, an overwhelmed operator, and/or decisions too fast for humans to understand or fix.
So to start, we offer precise definitions:
- Control is real-time or ex-ante intervention. A system does what you say.
- Oversight is policy or ex-post supervision. A system does what it does, and you watch, audit, or correct.
And critically, preventative oversight requires control.
- Control is real-time or ex-ante intervention. A system does what you say.
- Oversight is policy or ex-post supervision. A system does what it does, and you watch, audit, or correct.
And critically, preventative oversight requires control.
July 9, 2025 at 7:12 AM
So to start, we offer precise definitions:
- Control is real-time or ex-ante intervention. A system does what you say.
- Oversight is policy or ex-post supervision. A system does what it does, and you watch, audit, or correct.
And critically, preventative oversight requires control.
- Control is real-time or ex-ante intervention. A system does what you say.
- Oversight is policy or ex-post supervision. A system does what it does, and you watch, audit, or correct.
And critically, preventative oversight requires control.
In AI policy, “human oversight” is invoked constantly:
• In the EU AI Act
• In industry risk docs
• In safety debates
But there’s often a fundamental confusion: people conflate oversight with control. That confusion ignores key challenges, and hides where oversight can fail.
• In the EU AI Act
• In industry risk docs
• In safety debates
But there’s often a fundamental confusion: people conflate oversight with control. That confusion ignores key challenges, and hides where oversight can fail.
July 9, 2025 at 7:11 AM
In AI policy, “human oversight” is invoked constantly:
• In the EU AI Act
• In industry risk docs
• In safety debates
But there’s often a fundamental confusion: people conflate oversight with control. That confusion ignores key challenges, and hides where oversight can fail.
• In the EU AI Act
• In industry risk docs
• In safety debates
But there’s often a fundamental confusion: people conflate oversight with control. That confusion ignores key challenges, and hides where oversight can fail.
Given that @bc-y.bsky.social is here rather than active on twitter, I should tag him and given him the opportunity to respond to my concern that this was HARKing.
May 21, 2025 at 10:47 AM
Given that @bc-y.bsky.social is here rather than active on twitter, I should tag him and given him the opportunity to respond to my concern that this was HARKing.
