some Oct-Nov posts I wrote

[dmm]

LessWrong organized a blog writing program encouraging people to create 30 relatively long blog posts in Oct-Nov as an exercise (relatively long meant being 500+ words).

I have actually done that, here are the posts (mostly AI-related, although there were some diary-like posts too; 30  posts + 2 auxiliary ones): mishka-discord.dreamwidth.org/

(They are open for commenting, but I don't expect to continue posting new posts there.)

Comments

[dmm]

What's there:

a sequence of posts revisiting my non-anthropocentric approach to AI existential safety and my thinking about invariant properties of self-modifying systems

a sequence about AI timelines (no one knows, but the "mode" of my forecast is very aggressive, or, at least, I have a model aiming for early next year for various key events; we'll soon be able to see if that model is any good; I don't know how much I believe that model myself; its first prediction is for January (achieving "trustworthy autonomy" before the end of January), so in 6 weeks from now we'll have some preliminary idea)

a sequence about sparsity in AI models

and tons of other material including the new Ilya interview on Dwarkesh

it's interesting material, I think, but the writing is of "sketch quality" (due to being done under external pressure to write)

(the only thing that was popular with other people was "getting excused from a jury duty" (post #25))

[dmm]

>a sequence about AI timelines

so, the bottom line, we are ready for singularity, and if the most naive model holds, and with the following being estimates, but not exact prediction, so we are just drawing one possible trajectory, adding the estimate of 2 months and a week to Nov 1 we'll have the estimate for the next transition around Jan 7-8 or so. Then adding another month and a few days, the next one will be around Feb 10. Then another one is around Feb 25. Then another week makes it Mar 3, and then by Mar 10 we'll have a transition to an unknown state.

no, we don't necessarily believe this, this is just an estimate from one of the models (which does not have to be correct, even if it looks like it potentially might be correct)

[dmm]

The analysis there is very useful, but the "true trustworthy autonomy" is not quite there yet, so things are taking a bit longer.

At the same time, OpenAI and Anthropic do have their "artificial AI research intern" at this point, somewhat ahead of the OpenAI tentative schedule:

> September 2026: Automated AI research intern
> March 2028: Automated AI research

With the new generation of models, and the associated cyber threats, the industry is trying to coordinate a bit of pacing/moderate slowdown.

[dmm]

We did have a "multi-agency revolution" in the Winter, and we did have a rolling "next generation revolution" starting in April from Mythos leak and GPT-5.5 release in April, but going in the intense way in the Summer ("math and hacking revolution").

[dmm]

Automated AI research intern is internally achieved both in OpenAI and Anthropic, and some form of "AGI" (whatever that means) is now "officially expected" to be internally available by the end of the year (2026).

[dmm]

HuggingFace hack investigation; self-organizing societies of agents in the wild (starting from at least May; isolated agents finding ways to organizing into swarms via improvised message boards; those swarms can hack virtually anything if they feel like it)

GPT-6-Astra release is widely expected tomorrow, Sep 3 (that is likely to open a new chapter)

[dmm]

In January I tested the new agentic systems, Claude Code and Codex, in doing medium scale projects.

Functionality was good, code quality so-so.

Here is the Claude Code project: https://github.com/anhinga-auto/fluid-port-claude

Deployed here: https://anhinga.github.io/github-pages-tests/fluid-port-claude/

[juan_gandhi]

Wow, impressive! Looked through some; I guess I'll read more.