Jesper L.

Karma: 99

I collect AI progress monitoring resources here

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

I promote kindness and epistemic humility.

Polymath with very short inferential distance. This makes me take certain things for granted that merit explanation.

Currently working in big corp. Collecting coordination and stakeholder management skills.

Peak theory interests ranked: Ontology, information theory, epistemology, ethics, physics, negotiation, social dynamics, biology.

Math I find most useful: Group theory, analysis, combinatorics, number theory

M.Sc.

Honest Ethics & AI – Part 1: The origins of morality

Jesper L.25 Apr 2026 14:15 UTC

2 points

0 comments13 min readLW link

(controlled-ascent.ghost.io)

Jesper L. 20 Apr 2026 16:50 UTC
−4 points
0
on: Jesper L.’s Shortform
I asked Claude Opus 4.7 to give me its impression and review of Machines of Loving grace. It did so and when asked, rated it ⁷⁄₁₀ initially, but eventually landed on ⁵⁄₁₀.
It made the downgrade following deeper analysis. It made the call that the essay as diagnostic of the field doesn’t bode super well for a positive near-term future.
This downgrade happened when I asked to critique the essay deeper and investigate the arguments and motives carefully.
I extended and added some critique to help it realize the final score, but I also asked it to vet and push back on the critique that I had offered. It did in fact do this, arguably quite well. I never offered a score of my own until it had landed on 5. (Mine is 6 if read as a speculative personal blog post, in context of when it was released and how, I agree on 5.)
Personally, the essay made me more pessimistic when I first read it than before reading it, and I had hoped for and assumed the opposite. It was interesting to see the newest Claude version converge on my own take. Older models differ.
The main feedback converged without any input from me beyond the request to analyse the arguments deeper, neutrally stated. I will not dive into substance if anyone wants to replicate.

Jesper L. 20 Apr 2026 16:35 UTC
1 point
0
in reply to: sophont’s comment on: Jesper L.’s Shortform
I am not the best person to answer, and I should maybe avoid excessive detail on a searchable forum anyway, but yes, potentially you could request plasmids that at least togther become lethal by a third party.
Plasmids are explicitly tested for in the systems card. Home kits also could allow some serious work done outside public scrutiny. I think that (semi-)automatic labs is the real concern and this is a real but very budding area of development.

Jesper L. 17 Apr 2026 16:20 UTC
8 points
3
on: Jesper L.’s Shortform
The AI biology threat vector is heavily bottlenecked by deployment capability and lab skills. Theory and detailed guidance alone is largely insufficient to provide a significant increase in threat level. (I say this as a somewhat experienced wet lab biologist, someone who has worked in cell biology labs, biotech labs, performed molecular biology tasks in vitro, and studied virulence of serious pathogens inside a biosafety level 2 lab)
That said, Claude Opus 4.7 is quite competent in chemistry and biology, and the first Claude model widely released that actually worries me, based on the release notes and systems card. https://www.anthropic.com/news/claude-opus-4-7
I will not pretend that I did not think we would see AIs perform this well on these tasks so fast.

Jesper L. 8 Apr 2026 20:28 UTC
2 points
0
in reply to: ryan_greenblatt’s comment on: ryan_greenblatt’s Shortform
If you believe they meant that they have a path to reliably make safe AI, then the point that Anthropic is making this claim in bad faith holds true. Importantly, it holds even if you don’t consider existential risk, or believe Anthropic is on any direct trajectory to ASI right now.
I read it to mean that they believe they have a reliable way to accelerate risk mitigation. That still doesn’t matter if the tail-risk increases beyond a critical threshold, but it is more honest.
It’s a good callout actually, it says something about how they reason.

Jesper L. 6 Apr 2026 18:31 UTC
1 point
0
in reply to: 1a3orn’s comment on: 1a3orn’s Shortform
It was actually one of the easier things to predcit if you know something about how everyday humans actually think (like, evolved heuristicd, etc.). The added base knowledge of how LLMs reason is easily obtained by laymen, and I am sure some pychologists and marketers saw this coming before ML people did.
Here is my one-shot explanation: AIs are really convincingly talking like humans, but unlike humans, their semantic ontologies are not derived from causally validated patterns, only correlation.
Oh yeah, also all commercial incentives are for them to discount truth and suck up to the user.

Jesper L. 26 Jan 2026 23:11 UTC
1 point
0
on: How AI Is Learning to Think in Secret
maybe our last hope for understanding what these things are doing before they turn into completely opaque superintelligence, please God let it keep working.”
*
We got unbelievably, preposterously, cosmically lucky with Chain-of-Thought.
Actually there is an important point to make here. We didn’t get so lucky, of course.
Rather CoT was the logical next step, because, as you just explained, it is more powerful to build chains of reasoning than optimize every single output ad infinitum.
We will get far, but never get to SI, with only an LLM, exactly because they lack integrated learnings and continuous states.
An LLM designing the start of a new SI-seed architecture paradigm is one (speculative) thing, but neither whole-models nor latent mesa-optimizers inside them, can accidentally turn into SI on their own (with or without agentic scaffolding)
You can scale for some time, probably even to strong AI using CoT. But for SI you need self-optimization that can scale. Functional, fluid intelligence requires reasoning over time, not just endless optimization. It requires continuity, integration, and, for some endeavours, stakes (to assess feedback).

Jesper L. 21 Jan 2026 20:06 UTC
1 point
0
on: GPT-5.2 Is Frontier Only For The Frontier
I am still to finish my own small tests and lm-evals on GPT 5, but I am already quite concerned about deception, and even functional sandbagging. How much of such behaviour have the models absorbed from their training by now?
I have poor insight into the Chinese models, but so far the GPT 5 series is the scariest one I have interacted with.

Jesper L. 14 Dec 2025 21:47 UTC
1 point
0
on: Eliezer’s Unteachable Methods of Sanity
I think you are severely underestimating how relatable and common your thoughts on this topic are (also to many journalists). In short, you underestimate people’s capacity to get this (probably because they are out-of-distribution for your way of structured reasoning in general, to borrow LW 2.0 lingo).

If I would make a guess, I think that (self-aware) people outside of LW and similar circles may be even more likely to relate to several of these points than people inside of LW. For example, “a sentence about snow is words, is made of words, but it is about snow” and “if I were to dwell on how it impacted me emotionally that the world was ending, I would be thinking about something which genuinely doesn’t matter to me very much…
…they’re not what I’m about.” is exactly the kind of stance that seems to define large chunks of generation Z right now. In truth, if I did not know who said this, my first guess would be some GenZ celebrity, in a viral video clip on social media.

Jesper L. 27 Nov 2025 21:51 UTC
2 points
0
in reply to: habryka’s comment on: Jesper L.’s Shortform
Thinking about this much harder, I would rewrite the post to clearly phrase it as damage prevention you can do fast and cheap from your home.
I notice I forgot to mention a first aid kit. First thing to correct.
Secondly, when thinking hard on what of your feedback might actually be useful, I realized that statistically speaking you may do more good by buying slip mats for wet floors, and potentially investing in premium ladders. Since fall accidents dominate injury statistics. Not something I thought about intuitively but makes sense.
https://injuryfacts.nsc.org/home-and-community/safety-topics/fire-related-fatalities-and-injuries/
https://rehis.com/news/preventable-accidents-in-the-uk-are-rising-and-deaths-reach-an-all-time-high-warns-new-report/
For general awareness, accidents are 3d leading cause of death in US, fire is top 3 common accidents in the home, and quick google confirms cast majority of fire deaths happen in the home. Which I knew from safety trainings too.

Jesper L. 27 Nov 2025 19:24 UTC
−11 points
0
in reply to: habryka’s comment on: Jesper L.’s Shortform
Oliver. You just suggested “just use a smartphone” as a flashlight in an emergency as an argument not to get a flashlight. And you say I don’t think things through? You make claims that fire alarms are useless. Data? If you hadn’t revived LW I would think you are a troll. And your odd reasoning using GPT seems mostly off topic.
I don’t think there is any negative utility here, perhaps low on investing in a fire blanket.
Also, I have lived through scenarios where I needed a flashlight, and I have used fire blankets. Like, I have used these in my own life. I have been in the wild too, and the military ( I am from Finland.) Pretty standard safety advice here, tbh. Not sure what your problem is. If you think post is useless, and we should just trust the system, fine, but don’t make such low quality critique?

Jesper L. 27 Nov 2025 10:56 UTC
−10 points
0
in reply to: habryka’s comment on: Jesper L.’s Shortform
While skepticism on utility is wise, I strongly disagree with most of your comment. Fire accidents at home is by far the leading majority of fire related deaths in many countries, and fire accidents is a major cause of accidental civil deaths.
If we look at US as a whole (not where I live but you do), fire accidents are few versus drug overdose and gun violence. If I post: Don’t do drugs, is this useful advice? Tell me the low-hanging fruit there. Fires you CAN protect yourself against but many don’t.
Falls and poisoning are other common accidents worldwide. But, “don’t fall” is not very helpful, is it? And I don’t know how to proactively prevent posinoning for adults apart from not overdoing drugs. Those are context dependent. And traffic accidents are its own topic, everyone know about it.
Your claims on fire alarms lack backing. And using fire blankets for extinguishing small fires is a sound strategy although you may be right its not super effective. Point is, prevented home accidents do not show up in stats, it is akin to survivor bias.
As for using your phone and its limited capacity to work as a torch, this is just dumb. I mean, where do I start? Having a designated, powerful battery powered flashlight in an emergency cannot compare to the flashlight capacity of a smartphone, that may or may not be charged. And your phone is otherwise needed for other things during a longer blackout.

Jesper L. 26 Nov 2025 21:32 UTC
1 point
−1
on: Jesper L.’s Shortform
“Common sense” advice that is highly rational and easily overlooked by A LOT of people, including smart people

This is your regular reminder.

1. Buy fire alarms, a fire blanket, a strong light torch for blackouts, and a first aid kit (this is the bare minimum low-cost emergency kit that can save you a lot of hassle and potentially your life).

1.2 Do NOT have heavy power-draining appliances connected via extension cord. Fridges and washing machines should never sit in an extension cord unless you know what you are doing and that it can handle the strain.

A leading cause of accidental death in many countries, especially if we discard traffic accidents, relate to fires. If you live in a city, it’s probably much more common than you think to die or suffer serious injuries in a fire there. Power blackouts and power surges cause a lot of accumulating risks and issues too, but are harder to pinpoint in statistics. Minor accidents that dominate at home are cuts and infections.
2. Get private insurance, especially a good home insurance
3. Invest some savings in low-cost index funds, regularly, if you invest in nothing else
4. Take care of your BASIC health PROACTIVELY: Basics is to eat enough, sleep now and then, and watch your mental health. I constantly fail this one. I only notice danger levels when I am already crashing out. Admittedly more important for certain people, but many of them would be found on LW.

5. Have someone you can call in a crisis. Really, make sure you get this down. We are 8 billion people and far too many of us has a fragile or non-existent social support network.

Jesper L. 26 Nov 2025 14:00 UTC
1 point
0
in reply to: Raymond Douglas’s comment on: mattmacdermott’s Shortform
What does “pictures in order” mean? Also, damn.

Jesper L. 25 Nov 2025 23:58 UTC
1 point
0
in reply to: programjames’s comment on: Maybe Insensitive Functions are a Natural Ontology Generator?
If so, I would be genuinely curious to hear your ideas here. This might be an actually powerful concept if it holds up and you can formalize it properly. I assume you are an engineer, not a scientist? I think this idea deserves some deep thinking.

Jesper L. 25 Nov 2025 17:38 UTC
1 point
0
in reply to: Kaj_Sotala’s comment on: Things I have been using LLMs for
Great! Um, to be honest I am a little bit of a perfectionist, even for lower effort stuff, but as soon as I have a viable versioning of the files needed I will DM them to you :) Something to think of as a (hopefully) fun surprise later on!

Jesper L. 25 Nov 2025 17:20 UTC
1 point
0
in reply to: programjames’s comment on: Maybe Insensitive Functions are a Natural Ontology Generator?
Thank you, just knowing you are strictly coming from a ML perspective already helps a lot. This was not obvious to me, who have approached these topics more from a physics lens.

//
So, addressing your implementation ideas, this approach is practically speaking pretty neat! I lack formal ML background to properly evaluate it, but it seems neat.
Now, I will try to succinctly decipher the theory behind your core idea, and you let me know how I do.

You propose compressing data into a form that preserves the core identity. It gives us something practical we can work with.

The elbow has variables that break symmetry to the left and variables that hold symmetry to the right. This is an important distinction ~~between~~ from* noise and signal that I think many miss.

*mended, edit
This is all context dependent? Context defines the curve, the $B e t a$ parameter.
// How did I do?

Note: I should say at this point, understanding fundamental reality is my lifelong quest (constantly ignored in order to live out my little side quests) and I care about this topic. This quest, is what ontology means in the classical, and philosophical sense. When I speak about ontology in AI-context, I usually mean formal representations of reality, not induced ones. You seem to use AI context but mean induced ontologies.
The ‘ontology as insensitivity’ concept described by johnswentworth is interesting, and basically follows from statistical mechanics. But it is perhaps missing the inherent symmetry aspect, or something replacing it, as a fundamental factor. You can’t remove all symmetry. Everything with identity exists within a symmetry. This is non-obvious and partly my own assertion, but looking at modern group theory, this is indeed how mathematics define objects and so I am supported within this framework.

If we take wentworth’s idea and your elbow analogy, and try to define an object within a formal ontology, within my framework that all objects exist within symmetries, then we get:

$Concept = \frac{Total Reality}{Symmetries (The Tail)}$
The “Elbow” doesn’t mark where reality ends and noise begins. It marks the resolution limit of your current context.
- To the left of the elbow: Information that matters (Differences).
- To the right of the elbow: Information that doesn’t matter (Equivalences/Symmetries).
Your example was a hand-written digit “7”. The Tail is the symmetries. You can slant the digit, thicken the line, or shift it left. These are the symmetries. As long as the variation stays in the “tail” of the curve, the identity “7” is preserved. (Note that the identity is relative and context dependent).

The Elbow: This is the breaking point. If you bend the top horizontal line too much, it becomes a “1“. You have left the chosen symmetry group of “7” and entered the chosen symmetry group of “1”.

Jesper L. 25 Nov 2025 12:07 UTC
1 point
−1
in reply to: Jesper L.’s comment on: Gemini 3 is Evaluation-Paranoid and Contaminated
Alice I see you had difficulty to understand this. I tried to make it quite clear but feel free to ask. Basically, there are assertions here not mentioned in your post
1. Older Gemini models knew their identity, 3.0 does not
2. 3.0 Has a different thinking mode/architecture than the others. Zvi has also summarized this in his reporting
3. It is well known (and obvious) that LLMs have no sense of time since they don’t operate continuously.
Due to the knowledge cutoff being Jan 2025, but deployment being November 2025, these things conspire to generate the confusion and paranoia we see in the agent.

Jesper L. 25 Nov 2025 11:19 UTC
1 point
0
in reply to: programjames’s comment on: Maybe Insensitive Functions are a Natural Ontology Generator?
So, I feel like you just got deeper into the weeds here, thinking aloud. This seems interesting. I am trying to parse, but there is not enough formal context to make it make sense to me.
My main question was anyway, what w/could you use it for? What is the scope/context?
(Making some light banter) Maybe you are american, so I need to “debate” you to make it more obvious. “James, this is all a nice theoretical concept, but it seems useless practically. In its current form, I don’t see how it could be used for anything important”.

Jesper L. 24 Nov 2025 23:42 UTC
3 points
0
in reply to: programjames’s comment on: Maybe Insensitive Functions are a Natural Ontology Generator?
First: Yes, this post seems to essentially be about thermodynamics and either way it is salient to immediately bring up symmetry. So agree on that point.
Symmetry, thermodynamics, information theory and ontology happen to be topics I take interest in (as stated in my LW bio).
Now, James, for your approach, I would like to understand better what you are saying here, and what you are actually claiming. Could you dumb this down or make it clearer? What scope/context do you intend for this approach? How far do you take it? And how much have you thought about it?

Jesper L.

Hon­est Ethics & AI – Part 1: The ori­gins of morality

Honest Ethics & AI – Part 1: The origins of morality