iiinova — where your digital self keeps evolving

So now you're just gonna lie about what you wrote? We're done now. I said I was presuming good faith, can you very well established that you are not participating in any conversation in good faith? Have a good night.

100

You're seriously claiming that I did not ask you a question?

100

Then what did you do? I asked you a question and you claimed that I didn't ask you a question. You said I was instead pulling some dick move.

100

And then obviously there is the question of whether probabilistic thinking could be applied productively in this context. I'm merely pointing out this particular discipline does not prioritize probabilistic thinking. (or perhaps agreeing with you about this.) 2/

000

Look, I am very aware that law schools fail to teach their students how to think probabilistically. I think that's a real shortcoming of the discipline. But that does not mean that lawyers cannot learn to think probabilistically. 1/

200

You claim to be a career trial attorney. You claim that the human mind is not capable of evaluating the likelihood of winning a case before jury selection, even as vaguely as perhaps one chance in four. That's just what you wrote. Which part of that did I get wrong?

100

What is not true? That you treated a question as though it was an attempt to represent what you meant with confidence?

100

What is not true? I asked you a question?

100

I never said 71.6% chance. Again, you're arguing with a straw man. I said one chance in four. that's pretty broad

100

You have indicated that you don't understand what question marks are for. I asked you a question because I was interested in your answer. I asked you a question because how you answer it will determine what I say next.

200

I asked you a question. I asked you if that's what you meant. You still declined to answer the question. When I ask a question, I'm asking a question. It's an invitation to address a point. That might not be how you use questions, but it is how I use questions.

100

I didn't accuse you of saying anything. I asked you a clarifying question.

200

And you can't evaluate the chance of winning at trial? It's always 50-50, as best you can tell?

100

You are a career trial attorney?

100

I think you should be VERY careful with "I don't think the human mind is capable of" without a strong basis for the individual claim. In other words, why do you think the human mind is incapable of this particular thing? 7/7

100

There are so many things that I cannot do and cannot imagine how someone is capable of doing it. I mean…someone wrote Bolero. Regardless of neural condition, HOW!? Or the Imperial March? 6/

100

I don't think the human mind could possibly conceive of cubism. I mean…WTF? And yet, a human mind DID. 5/

100

For example, being able to shoot open NBA-range point shots at like a 50% rate? It's just neurons controlling muscles. But I cannot really can't understand how a mind can make that happen consistently enough to hit those shots at such a high rate. 4/

100

That is not to say that I can learn to do ANYTHING. No, I have my limits. 3/

100

That's *not* to call you dumb. It's just to point out that expertise matters. There are many things that I cannot do and I cannot conceive of how they are done. And yet, I have seen them done. I just don't know how to do them. I lack the skills, knowledge and experience to do them. 2/

100

Well, YOU can't. But you are not an experienced expert who has spent a career learning how juries work and how they respond to things. 1/

200

Judgement of the chance of winning is just not the same thing as judgment about the actual guilt (by the prosecutor). you seem to want to equate them. here's a proof that they are not necessarily equal. 6/6

000

Because the reward (in terms of justice) is great enough to be worth the trouble, even if if there's less than a 50% chance of winning. Because the alternative is a 0% chance of justice. 5/

100

Why would this prosecutor push this case, despite being less than sure that they can win? less the 51% confident? Because a crime has been committed. They know who did it. And they want them punished. 4/

100

So, the prosecutor is limited to admission evidence, but thinks there's only like 1 chance in 4 of getting a jury who will unanimously see it as beyond a reasonable doubt. 3/

100

Let's say that there's some inadmissible evidence that proves guilt. The prosecutor therefore has absolutely 0% doubt about guilt. Anyone would. (The evidence is not not admissible because of THIS case, but because of the importance of preserving people's rights in the future.) 2/

100

I'm gonna presume good faith here. So, let's examine a particular hypothetical to isolate an answer to your question, and then maybe work out from there. 1/ (sorry. technical problem. reposting)

100

So you're changing the question. You're presuming that the defendant is in a jail cell.

000

I'm not so sure that that's the ENTIRE reason, but it's super important. And you're arguing against a strawman, not me

000

That's a huge logical leap. I agree that if a prosecutor knows that the actual evidence isn't sufficient to meet the burden under the law that it would be unethical to prosecute the case anyway. But that's not what I was talking about. I was talking about probabilities, not certainties.

100

Ceolaf @ceolaf.bsky.social · 18h

Yes. I did use a shorthand there. No doubt. Certainly beyond a reasonable doubt, there.

000

A prosecutor who is not confident they can convince a jury is not the same thing as a prosecutor who is not confident beyond a reasonable doubt that a crime has been committed. Methinks you are conflating the two. 3/3

100

Ceolaf @ceolaf.bsky.social · 18h

That's not a question of how confident the prosecutor is in the guilt. The problem is that the decision needs to be made in advance of jury selection. So there has to be a huge uncertainty bar around the certainty judgment, simply because it is an unknown jury. 2/

200

Ceolaf @ceolaf.bsky.social · 18h

There's a question of how confident they should be that they can prove guilt to a jury beyond a reasonable doubt. Should they be 100% confident? 51% confident? "I probably can't prove it to that level, but it's worth trying nonetheless." 1/

200

Reposted by Ceolaf

Jess Calarco @jessicacalarco.com · 08/10/2026

It's telling that men would rather try to invent tech to replace humans' role in caregiving than step up as equal partners in parenting. Telling, that is, in that it shows how financial capitalism and patriarchy work together to dissuade men from doing "unprofitable" labor, like the work of care.

X post from AI industry insider Alvin Sng :

My wife: Can you read our baby a book?
Me: I can just play an audiobook from my phone.
My wife: But then it won't be in your voice.
Me: I can train a voice model.
Parenting hits different when you work in Al.

35148891208

Ceolaf @ceolaf.bsky.social · 08/10/2026

1) Costco does not sell 50-pound boxes of spaghetti 2) If you throw a 50-pound box of spaghetti at the wall, nothing should stick. Boxes of spaghetti are full of dry pasta. @npr.org

000

Ceolaf @ceolaf.bsky.social · 06/10/2026

@jennbinis.bsky.social you'd love this keynote, methinks. Lots of warning about how AI undermines learning and the importance of metacognition. #AIME-Con

010

So, multi-agent seems to push more consistently toward the existing ceiling—with relatively simple multipliers on resource costs. More like improving reliability than improving validity. Valuable. But let's recognize what it does for us and what it doesn't. 8/

000

That's the fundamental challenge: how to add capabilities beyond what training and reinforcement baked into the model. Mere prompt engineering doesn't do it. Multi-shot hasn't gotten us there. I'm not seeing multi-agent getting there. 7/

100

Because training is expensive—and far more difficult with frontier models. 6/

100

And then trying to figure out how to add learning (perhaps from HitLs or meta-HitL enabled HitLs back into the automated portions of the loop. 5/

100

This is about recognizing where HitL (human in the loop) is still needed, when multi-HitL and/or meta-HitL might be needed—or at least helpful. 4/

100

So, the next family of strategies seems to be multi-agent. That's what I am working on now. But I keep seeing the same underlying problems as before. Multi-agent is not raising the lower ceilings. 3/

100

Multi-shot risks over-generalizing from from inevitably small sample samples. 2/

100

I didn't say they don't work in *any* case. I am working on finding strategies to help where LLMs don't have the expertise in their training or original reinforcement. Zero-shot, even with detailed instructions doesn't work. Original training and reinforcement overwhelms. 1/

100

No, I want to recognize and accurately document the limits, figure out what might expand those limits outward, and recognize what kinds of problems need entirely different strategies. That is, real research rather than advocacy.

100

I point out that a strategy is not working to solve all problems and you say I should count on others fixing it!? I'm continuing my research efforts to contribute to finding strategies that work.

100