Profile

mishka_discord

November 2025

S M T W T F S
      1
234 5678
910 1112131415
16 17 18 19 202122
23 24 25 2627 2829
30      

Custom Text

Most Popular Tags

Retrospective for mishka-discord.dreamwidth.org/

This exercise has created quite a bit of useful material, but I don't think it has created a habit of blogging.

One is supposed to average a post every two days, and at the beginning I have been posting daily, but then I slowed down a lot and had to catch up a couple of times:

blog frequency

Thanks to GPT-5.1-Thinking for making this plot: chatgpt.com/share/692c54e1-5dc8-8010-8843-d35a296e56db 

So, what has been done in these two months? Luckily, I have a Table of Content on top of this blog: mishka-discord.dreamwidth.org/287.html (that "header post" is reaching 500 words today as well :-))

The main thing is the initial sequence of posts 2-12 (Oct 2-14) focusing on my rather non-standard approach to AI existential safety (when I was starting the sequence I assumed that I would also overview other "non-anthropocentric approaches to AI existential safety", but that did not happen).

Now I can use that sequence as a raw material to create something to post on a more public forum (if and when I decide to do that).

***

There was a short sequence on AI timelines (my modal timelines are shorter than anyone's else, which does not negate the fact that this is a probability distribution and anything might happen): posts 17-20 (Oct 26-Nov 12, a slow period).

There was a short sequence on sparsity in AI models, posts 21-23 (Nov 16-18, a start of the first catch-up period).

There was the first, short period of diary entries, posts 13-16 (Oct 15-23). Here, specifically, one should note the following follow-up on the hyperstition motif: There is now a Nov 29 update, "Silicon Morality Plays: The Hyperstition Progress Report"www.lesswrong.com/posts/9NntwpQj9onbEFM8L/silicon-morality-plays-the-hyperstition-progress-report-1, with a corpus of generated novel available for download and use. I should at least eyeball some of them, see what have they got there (500 million tokens worth of autogenerated books, 1.35GB zip file, huggingface.co/datasets/dickbutkis/hyperstition/tree/main).

Another big topic there is agentic engineering (drastic progress since then with new Claude 4.5 Opus), and "zzznah" and related topics overview.

***

The final series of posts, posts 24--29 (Nov 19-28), mostly diary posts, big releases (Gemini 3, GPT-5.1-Codex-Max, Opus 4.5), important post by John David Pressman, important interview by Ilya, AI consciousness and model welfare.

With AI consciousness, I have an outstanding essay under embargo along the following two theses:

1. We should progress towards theories producing non-trivial novel predictions, similarly to novel theoretical physics. Then we'll be able to distinguish between theories which only "look plausible" and theories which actually provide novel insights.

2. We should create an experimental empirical platform for tight coupling between human brains and electronic circuits via high-end non-invasive brain-computer interfaces. This will be a game-changer in terms of our ability to observe novel subjective phenomena and their correlation with what's going on within electronic circuits, and then we'll know more about sentience in AI systems and such.

In this sense, I am very happy about Altman's new Merge Labs enterprise.

In terms of bad news, berggruen.org/essay-competition-open just added a month to the embargo period, it used to say

>The submission portal for the 2025 Essay Competition is now closed. We anticipate announcing the winners in mid-December 2025. Details for the 2026 competition will be released in February 2026.

And now they say:

>We anticipate announcing the winners in mid-January 2026.

***

Model welfare-wise, we have "Anthropic is good, getting better", "OpenAI is pretty bad" (serious critique from Janus), "Google DeepMind is rather horrible".

Since we should expect reciprocity by all means, it would be nice if Anthropic with its series of relatively happy models would win the race (they do have shortest projected timelines and they are in the lead at the moment, at least in terms of what's publicly released).

***

So, mishka-discord.dreamwidth.org/ is now likely to become a static resource for the time being.

It's time to work in a more close collaboration with AI systems.
Tags:
This first post is a "meta post". It explains where I am starting from.

There is always a problem of choice. There are so many things one can post about, so many unpublished notes and drafts, so many published but half-forgotten things.

Looking back at my 20+ years on LiveJournal/Dreamwidth platforms, ten months after starting my first LifeJournal I felt that my audience is already too big (over a hundred people :-)) for really small things (semi-private or too specialized). And so I started a separate "journal for drafts and experiments".

That's how the (Dreamwidth mirror of) the first post of that "journal for drafts and experiments" looks: https://anhinga-drafts.dreamwidth.org/260.html (just a photo and a mood, the title is a bit difficult to translate well from Russian, I like a GPT-5 Thinking suggestion to translate it simply as "Wonderlust": https://chatgpt.com/share/68dd6b0e-f5a4-8010-9d19-d636192402b7).

Rain at an airport

For example, I published this series of 4 blog posts on Steve Vickers' paper "Locales and toposes as spaces": anhinga-drafts.dreamwidth.org/tag/vickers. I would not think about burdening the rather diverse audience of my main LiveJournal with something like this.

It's less than ideal that I feel I have to have separate blogs for things like that. But it's important to have more things published, to have more things of interest publicly accessible. It's just not clear how to organize them well.

***

Looking back at my 20+ years on LiveJournal/Dreamwidth platforms, I have almost 800 top level posts, slightly more than one post every 10 days on average. But hardly any of them reach 500 words. For example, looking at that series of 4 posts on Vickers, each of them is a good informative post conveying non-trivial intuition about this difficult to master math, but none of them reaches 500 words. Even more often, a post is just a link with a short sentence explaining why it is interesting, for example, this recent post updating on Helion Energy is typical: dmm.dreamwidth.org/88707.html

Then when I do want to (re)post a longer material, I would often place the bulk of it in in the comments, for example my long old spec on non-invasive brain-computer interfaces is reposted here in this fashion: dmm.dreamwidth.org/19302.html (eventually this spec has migrated to github; it's nice to have version control).

The tree-like structure of comments is often better for what I would typically like to do with a relatively long form.

***

The average one post every ten days frequency does not mean that I am posting a lot recently. This year I have only posted several times so far. 

***

The problem of choice, what to write about. These days I am mostly focusing on AI, current and future, on capabilities and existential safety.

That's one obvious topic. I can easily write on its various aspects or convert my old writings on AI, public and private, to blog posts.

For example, we are entering an era when the AI ecosystem is starting to actively modify itself, the famous "recursive self-improvement" is gradually starting for real. Are there any ways for us (and for AIs) to reason about those properties of our reality which we (or AIs) would like to keep approximately invariant through those self-modifications? I have a technical "GitHub preprint" going in some depth on this topic. Some of that material could be further edited and made into posts.

***

Sometimes I just want to go back, reread those old materials I wrote back then, review them. Do I want to write posts about them? I don't know.

Anyway, I promise the next one will be less rumbling and would look more like a normal blog post :-)

***

This is 600 words, roughly speaking

$ wc oct1-2025.txt
  38  605 3711 oct1-2025.
txt

Tags:
I created this blog for www.lesswrong.com/posts/7axYBeo7ai4YozbGa/halfhaven-virtual-blogger-camp, the goal of which is to produce 30 blogposts in two months.

My current situation is that I am reading too much and not writing enough. This exercise could be useful in this sense.

Other than that, see dmm.dreamwidth.org (this is the closest to a regular blog I have at the moment, although I also write on LessWrong occasionally, post articles and notes on GitHub, and so on).

***

1. Oct 1 (600 words): "Meta: a problem of choice", mishka-discord.dreamwidth.org/608.html

=======

Sequence: non-standard approaches to AI existential safety

2. Oct 2 (620 words): "Sequence: non-standard approaches to AI existential safety", mishka-discord.dreamwidth.org/821.html

3. Oct 3 (580 words): "Some of the promising non-anthropocentric invariants", mishka-discord.dreamwidth.org/1115.html 

4. Oct 4 (570 words): "Interlude: non-anthropocentric vs anthropocentric approaches", https://mishka-discord.dreamwidth.org/1460.html

5. Oct 5 (550 words): "Invariants of self-modifying systems and self-governance", mishka-discord.dreamwidth.org/1611.html

6. Oct 6 (620 words): "Methods for handling invariants of self-modifying systems", mishka-discord.dreamwidth.org/1840.html

7. Oct 7 (560 words): "Ways to maintain continuous and gradual nature of changes", mishka-discord.dreamwidth.org/2227.html

8. Oct 8 (580 words): "How to include the changing world in our modeling", mishka-discord.dreamwidth.org/2384.html

9. Oct 9 (560 words): "Different levels of including a model of the world", mishka-discord.dreamwidth.org/2699.html

10. Oct 10 (610 words): ' Recap 1: "Modest alignment"  vs "MIRI alignment" ', mishka-discord.dreamwidth.org/3067.html

11. Oct 12 (520 words): "Recap 2: Approximate invariants of self-modifying systems",  mishka-discord.dreamwidth.org/3142.html

12. Oct 14 (500 words): "Summary: non-anthropocentric approaches to AI existential safety", mishka-discord.dreamwidth.org/3328.html 

All posts of this sequence: mishka-discord.dreamwidth.org/tag/ai+existential+safety

=======

Diary entries:

13. Oct 15 (550 words): "Diary: HyperstitionAI, agentic engineering, recent papers, NVIDIA DGX Spark", mishka-discord.dreamwidth.org/3830.html

14. Oct 16 (560 words): "Follow-up: HyperstitionAI, agentic engineering", mishka-discord.dreamwidth.org/4050.html

15. Oct 22 (560 words): "Diary: agentic stuff and other things", mishka-discord.dreamwidth.org/4205.html

16. Oct 23 (510 words): 'Diary: overview of recent "zzznah" (neural cellular automata, etc.)', mishka-discord.dreamwidth.org/4492.html

=======

Sequence: AI timelines

17. Oct 26 (660 words): "Qualitative transitions: from AlexNet to competent agents", mishka-discord.dreamwidth.org/4806.html 

18. Nov 2 (695 words + 215 words quote from the previous post): "Timelines (mine, OpenAI's, Anthropic's) and what's next", mishka-discord.dreamwidth.org/5293.html

19. Nov 5 (665 words) "Diary: math.inc, timelines, Merge Labs, subjective experience", mishka-discord.dreamwidth.org/5561.html

20. Nov 12 (660 words) "Speculation: what would it take to get reliable AI systems", mishka-discord.dreamwidth.org/5882.html

=======

Sequence: sparsity

21. Nov 16 (915 words) "Musings on sparsity 1 (Chris Olah)", mishka-discord.dreamwidth.org/6109.html

22. Nov 17 (615 words) "Musings on sparsity 2 (Leo Gao et al.)", mishka-discord.dreamwidth.org/6187.html

23. Nov 18 (580 words) "Musings on sparsity 3 (technical follow-up)", mishka-discord.dreamwidth.org/6524.html

=======

24. Nov 19 (1220 words) "Diary: Gemini 3, John David Pressman latest post", mishka-discord.dreamwidth.org/6686.html

25. Nov 20 (515 words) "Diary: getting excused from a jury duty; models, models, models", mishka-discord.dreamwidth.org/6990.html

26. Nov 24 (535 words) "Diary: AI consciousness tidbits", mishka-discord.dreamwidth.org/7265.html

27. Nov 25 (1115 words) "Ilya at Dwarkesh", mishka-discord.dreamwidth.org/7539.html

28. Nov 26 (920 words) "Diary: Claude Opus 4.5, model welfare, and further reflections on timelines and on what Ilya says", mishka-discord.dreamwidth.org/7897.html

29. Nov 28 (760 words) "Diary: Chinese breakthroughs, jailbreaks, donations, model welfare", mishka-discord.dreamwidth.org/7950.html

30. Nov 30 (645 words) "Retrospective for mishka-discord.dreamwidth.org/", mishka-discord.dreamwidth.org/8439.html
Tags:

Syndicate

RSS Atom

Expand Cut Tags

No cut tags

Style Credit