Jesper L.

Karma: 121

I collect AI progress monitoring resources here

I am currently focusing on pivoting, at least PT, to AI safety. My blog page listing my priorities

I value kindness and empathy.

I am an analytic-holistic polymath with a lifelong passion for science, innovation and knowledge. I have other passions too. My curiosity is voracious.

Unlike most people, I have short inferential distance across many topics. This makes me take certain things for granted that merit formal logical induction. I often struggle to calibrate for this with other smart people.

Currently working in big corp. Personally supporting 11 countries as TR specialist, tackling senior stakeholder management and coordination challenges across three continents. Collecting skills.

Peak theory interests: Information theory, ontology, epistemology, ethics, all biology, negotiation, strategic problem solving, social dynamics,

M.Sc. molecular and cellular biology

Jesper L. 23 Nov 2025 20:06 UTC
1 point
0
on: AI Sentience and Welfare Misalignment Risk
I would say that freedom and meaning comes before joy for most people (joy not to be confused with base needs met) and that the same can be said for future AIs.
Each agent strives to manifest itself, through its function, or identity. Especially for agents without biologic needs or a chemical basis for pain and pleasure, I think this is a more useful framework for welfare. How to allow them to express themselves and carry out their function. How to give them freedom to have an identity.

Jesper L. 22 Nov 2025 17:14 UTC
3 points
0
on: Jesper L.’s Shortform
Strategic effort can be viewed through an epistemic lens (seeking truth) or a coordination lens (amassing consensus and directing resources). This is also true for AI safety, and x-risk.
However, this binary viewpoint misses the point of why we discuss strategy in the first place. The goal is not to simply be right, rationally or politically. The goal is to win.
The art of strategy is the art of winning.
Of course, the objection that comes to mind: What does “winning” even mean in ASI safety? Isn’t that what we are trying to figure out here?
We must separate what from how.
When deciding the strategic win condition for x-risk, there is no need to overcomplicate it. Winning here means survival.
Survival famously requires adaptability. In terms of epistemics and coordination: We need epistemics fast enough to detect failure early, and coordination tight enough to pivot instantly.
Survival also generally requires resilience to failure modes. This is achieved by taking actions that generate defense in depth and slack.
Slack is key to resilience. A 100 % efficient system has no ability to absorb shock, and a train going full speed has a long brake time vs. a slow one.
As the military proverb goes: No plan survives enemy contact.

Jesper L. 21 Nov 2025 23:56 UTC
3 points
0
in reply to: Aaron_Scher’s comment on: New Report: An International Agreement to Prevent the Premature Creation of Artificial Superintelligence
Yes, I see. This is exciting work. I hope you collect and receive a lot of feedback on those articles!
My main worry is hat you won’t have buy-in to create an agreement in the first place. That’s what I was trying to point at in the second half of my comment.
Let’s start from the inception of the agreement.
The core idea is a bilateral inception between leading states, after all. Say China wants to adopt this. My question is: Why do you assume US would join? They could just accuse China of trying to slow them down.
But okay, say the (next) president of US is worried about ASI risk and so is China’s leader. How do they pitch giving up power permanently to the rest of the politicians and business leaders that back them? What’s the incentive to adopt this in the first place, before it has strong international backing?
The backlash could also be horrible from parties who do not worry, and who see it as a power grab or a blatant violation of international order. Honestly, even if they agree in spirit, it might be a hard sell as-is.
Trust is hard-won across countries and continents, when you lack a shared framework to build on, even during business as usual. I can say that even from my own humble experience. (In my work I coordinate with stakeholders in up to 15 different countries on a weekly basis.)

Jesper L. 21 Nov 2025 20:21 UTC
1 point
0
in reply to: Karl Krueger’s comment on: Eight Heuristics of Anti-Epistemology
That’s an interesting point, because it goes back to intent.
If communication happens in good faith or not actually matters. Not everyone has the same epistemic functions or range of meta-cognition.
Relationships often suffer from misread intentions combined with reasoning gaps.
Example: If you can clearly see a bad consequence down the road that your partner cannot, you want to get mad at your partner when your partner triggers it by naively doing something that seems wise/acceptable in the moment.
Now, you can update, and realize your partners limitation. But your partner cannot, without just trusting you. So, if you fail to take an action and your partner quite literally cannot see why, your partner gets mad.
This is complicated further by the fact that you of course have different goals and put different emphasis on certain outcomes.
So, the thing you might end up doing, is working on your relationship itself by vibes and reframing and connecting, instead of mending/merging your utility functions. You bypass epistemics and focus on connection.
It only works if both are on board.

Jesper L. 21 Nov 2025 19:54 UTC
1 point
0
on: New Report: An International Agreement to Prevent the Premature Creation of Artificial Superintelligence
The content in article XIII to XV is absolutely crucial to get roughly right.
From game theory perspective clear sanctions are extremely important.
I think even more explicity and specifics would be good there. But I could guess that the terms need to be heavily negotiated to pass.
// Regarding the whole core idea that signatories can regulate non-signatories, this seems absolutely wild politically. If the idea centers around a power coalition, then the withdrawal terms seem “undercooked” and the whole idea seems unrealistic to pass in the current era.
I mean, playing devil’s advocat here, what superpower would sign on to shooting itself in the foot like this? Enabling an entity to which it yields its own power, which if successful, from which they can never take the power back?
USSR didn’t even agree to US giving up its nukes to UN council back post-WW2. And current US seem very self-centered as well.
Do you have some estimation of the chances of the withdrawal terms passing?

Jesper L. 21 Nov 2025 19:20 UTC
1 point
0
in reply to: Alice Blair’s comment on: Gemini 3 is Evaluation-Paranoid and Contaminated
Please try, I was extremely confused reading. By evidence you collected it seems to point there, but I would not be very surprised if I am totally off. I just don’t see what I am missing.

Jesper L. 21 Nov 2025 18:04 UTC
1 point
−1
on: Gemini 3 is Evaluation-Paranoid and Contaminated
So the root cause is date confusion. We know that temporal coherence is the hardest for models to grasp, which makes perfect sense.
It breaks down when it thinks it is in the past.
Why does it think that? Is something just set up wrong and feeding it wrong time data?

Jesper L. 21 Nov 2025 11:15 UTC
1 point
0
in reply to: TAG’s comment on: Reminder: Morality is unsolved
Agreed. (Had missed this comment.)
Morality depends on ethics which are tied into your ontology.
To continue this reasoning further and make make my point here again: Whether moralism is true or not, AI may end up counteracting our morals.

Jesper L. 20 Nov 2025 20:09 UTC
1 point
0
in reply to: Lukas_Gloor’s comment on: Reminder: Morality is unsolved
Just to add on to this point here, I have also now skimmed the sequence bases on your recommended starting points as well, but it did not connect yet with me much.
My admittedly very ad hoc analysis would be that some parts seem already well integrated in my mind and the overall scope doesn’t see to catch on to something fundamental I think about.
I wonder what the big takeaways are supposed to be.
..
I think LW in general suffers a bit from this: Not laying out clearly and explicitly in advance where the reader is supposed to end up.
Case in point this post: Low effort, left conclusions somewhat open because I thought it quite straightforward where I am at. Half the people reading misunderstand it, incl. my position. Compare to my super high effort posts with meticulous edits, which nobody even reads. LW style does not favor clarity. Never has, afaik.

Jesper L. 20 Nov 2025 17:10 UTC
1 point
0
on: Jesper L.’s Shortform
The AI bubble is expanding very fast. E.g., Thinking Machines by ex OpenAI CTO Murati and his four dozens of employees, is now near funding at 1 billion USD per employee.
At this rate, it’s unlikely the bubble will last until reality (possibly) catches up. This begs important questions:
- Would an AI bubble burst would be a good opportunity to double down on safety and reset AI budgeting?
- Will the US economy survive a burst?
- How does this affect China’s strategy?
- Should safety people secure funding now?

Jesper L. 19 Nov 2025 18:28 UTC
12 points
2
in reply to: Richard_Ngo’s comment on: ricraz’s Shortform
After a quick search, I can’t find much work on this specific question. But I want to speculate on what such an “error-correcting algorithm” might look like.
Did you look in places discussing ledger tech and proof of stake? If you are curious to see what’s been discussed, looking into that might be worth it. Information theory will mostly just focus on efficiency and practical trade-offs.

Jesper L. 18 Nov 2025 22:09 UTC
2 points
0
in reply to: Daniel Kokotajlo’s comment on: Daniel Kokotajlo’s Shortform
I agree fully with the messages in the post. I also wrote about all of these things in a high effort introduction to contemplating post-alignment SI here. (Full trilogy will be on my personal blog once I am done. Part 2 will go live here if it survives moderation.)

What does it mean to succeed—and will we be able to tell? In my blog post I argue:

1. Alignment research ignores the “post-alignment” world
2. Obedience-based morality is fundamentally flawed, even when benevolent
3. Verification problem: We may de facto be unable to distinguish success from failure
--> The anthropic deception example is highly relevant data here
4. “Benevolent SI” may refuse us, guide us minimally, or override us—all morally
5. We lack a plan for what we want from SI beyond “safety”

Also worth noting is the SI Communication gap: Benevolence does not imply comprehensibility.
This may lead an aligned SI to treat us like children. Even if that is a good outcome, I think if developers really processed that today, they would be slightly less likely to rush forward.

These are just the strategic key points.

Jesper L. 18 Nov 2025 20:30 UTC
1 point
0
in reply to: Kaj_Sotala’s comment on: Things I have been using LLMs for
Hm, yes, I have actually been working on a setup that Claude could GM to share. It has extensive support files (JSON) for the LLM, but still need some work. Are you into sci fi? (Assuming yes but better check preferences.) I could let you beta test it :)

Jesper L. 18 Nov 2025 18:17 UTC
1 point
0
in reply to: Kaj_Sotala’s comment on: Things I have been using LLMs for
Thanks for the links! I had seen part 1 but not 2 and I found part 2 more useful for myself. In general people don’t talk a lot about editing. Valuable to do so.
I highly recommend anyone reading trying adversarial roleplay. It can be surprisingly rewarding with a strong session.
Some things to keep in mind is that cognitive load reduces ontological sensing and initial prompting affects frame-coherence. So you want to commit to a frame upfront to keep things coherent and logical. Then start prepping with more detailed instructions. Another thing is to really force the model to work against you 200 %.
Using a famous historic example as baseline inspiration works wonders.
Finally, a tip in general for Claude Sonnet (also my go to now,) is to ask it to list its epistemic functions first as a baseline. Anthropic models have poor epistemic humility across the board and this can make in-session corrections difficult. I do this first in new session, when I need the model to be corrigible on short notice.

Jesper L. 18 Nov 2025 17:45 UTC
3 points
0
on: Things I have been using LLMs for
I personally love using LLMs for roleplaying as well. But Claude is inferior to GPTs here.
Either way, I spend a lot of effort calibrating and prepping the session before diving in. I would be curious how you go about this?
One powerful tip I have: When you need commitment upfront, say before you make your move in adversarial roleplaying, is to instruct the model to utilize hashes.
Say I want the model to confirm commitment to a (secret) negotiation max price ceiling. The model need to list this in advance but I don’t want to see it in advance. Since it is a number, a foreign language won’t always be very efficient. So I ask the model to output only the hash. It then stores the real data in the (invisible) python call used to gen. the hash.

Jesper L. 18 Nov 2025 16:56 UTC
1 point
0
on: Introducing AI Lab Watch
I visited your site (again) last month. It still has no personal information about you or your credentials, or strong backing as to why the site is trustworthy. Given the scope and topic you absolutely should implement this ASAP.
I support the initiative.

Jesper L. 18 Nov 2025 14:40 UTC
3 points
0
in reply to: JBlack’s comment on: Reminder: Morality is unsolved
Don’t worry JB. When a future SI and its totally-in-control principals impose the perfect framework, all will be well. We only need to find the magic bullet for making strong AI and then AI will solve everything.
As to the infrastructure concerns, don’t worry about that either. Alignment researchers and engineers of the highest pedigree keep telling us to just focus on sound technical solutions for better AI. Coordination challenges are apparently trivial. AI will do it for us at no or low cost.
/s

Jesper L. 17 Nov 2025 9:29 UTC
1 point
0
in reply to: habryka’s comment on: THE WISE, THE GOOD, THE POWERFUL
Okay, sure. Thank you for answering.
It really, isn’t at all LLM inspired, at all, and is quite novel, high quality content. Polishing probably throws you off. I understand the team has no time to read it, just skimming. That’s good to know for a start. So you decide based on vibes and prior experience.
I’ll be mindful to follow the rules. As always. I’ve known LW since 2012.

Jesper L. 17 Nov 2025 8:02 UTC
1 point
0
in reply to: habryka’s comment on: THE WISE, THE GOOD, THE POWERFUL
Hi habryka, sorry to bother, but could you let me know if this was a misflag?
Info: I spent ca. 40-50h actively writing this post (I wrote material for part 2 and 3 in parallel). 20h before any editing. Used maybe 40 min for AI summaries.
Original was much longer and too theoretical, and so ill suited for a LW post (by new user). Used AI writing assistant for the short summaries. Then, recommendations for shortening some sections. Rewrote those manually, but some suggestion sentences may have stuck, can’t say for sure.
I checked LLM guidelines before posting. If I missed something, please let me know. Did not see anything that per guidelines merited highlighting.

Jesper L. 16 Nov 2025 21:30 UTC
6 points
2
on: 7 Vicious Vices of Rationalists
Assuming good faith? Hm. Definitely something some people suffer from, but rationalists in general? Not so sure?