![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
1. With my "superoptimistic" (for some value of "optimistic") timelines, the prediction that the next qualitative transition will be to “trustworthy autonomy” makes intuitive sense, but it's rather difficult to imagine that the trend of two-fold shortening of the interval between qualitative transitions will hold and that we are going to have that in January 2026: mishka-discord.dreamwidth.org/5293.html
We do see that in September 2025 Christian Szegedy created math.inc, x.com/ChrSzegedy/status/1966186676289741117, with major breakthroughs www.math.inc/gauss and great vision: www.math.inc/vision
It's not difficult to see how this would lead to trustworthy autonomy and more in the next year, but that's still not pointing to January 2026, that's more like Anthropic timelines. But it will be always difficult to say, "yes, it's now, it's happening".
Sam Altman said a few days ago that their systems are at about "5 hour task difficulty" now (up from GPT-5 2+ hours METR evaluation). With normal development, they will be at about 10 hour task difficulty in January, but I think “trustworthy autonomy” needs more than that.
We do expect Gemini 3.0 in a few weeks joining the wave of competent agents, and it would not be too surprising if it becomes number 1. A strong Grok 5 is a possibility. We do see more orgs doing automated AI researchers in the style of Sakana AI. Things are happening; agents like Codex CLI over GPT-5-Codex and other strong agents are accelerating AI R&D, more and more compute is coming online.
But it does feel that having “trustworthy autonomy” in January 2026 would require something coming "from the left field", something we don't quite expect at the moment, although the trend of two-fold shortening of the interval between qualitative transitions does point at January 2026 as the time of another major transition.
2. Sam Altman is getting together something called Merge Labs, a start-up around high-end non-invasive brain-computer interfaces (non-invasive BCI). It is likely to involve transcranial ultrasound for both read and write. Risks from correctly used transcranial ultrasound are presumed to be mild, but this is a very new tech, so who knows.
This would help us with many thing, and, in particular, with making progress in solving the "hard problem of consciousness" by providing a radically novel experimental platform. Invasive methods like those used by Neuralink are too risky, healthy researchers would rightly hesitate to use them on themselves, permissions are very difficult to get, it's all very slow.
We want methods which are fast, relatively safe, reversible, relatively inexpensive, and which don't need formal permissions from authorities. Non-invasive BCIs are the way to do that, invasive BCIs are a road to nowhere.
Pondering subjective experience, I was just looking at color lights on the street in connection with Winter holidays, and yes, the glowing filament inside the colored glass is "actually" white, but our mind makes a model of the world where the filament itself is colored, and not just the glass of the lamp. That's because the filament looks colored after the light has been filtered by the glass.
Similarly, out brain circuits filter and transform external signals, and that's how a very cool colored subjective reality is created, although the signals transmitted between neurons in the brain are not colored. Yet, the subjective aspects of it, the color "qualia" remain a mystery. We more or less understand how colors are coded in the brain, but the transition from brain dynamics to subjective feelings of qualities is not understood at all; we don't even understand what those subjective feelings of qualities are.
Yes, we sort of understand the information processing and signal processing aspects of it, but we don't understand the "subjective realm".
However, do see this essay: "Why it's so hard to talk about Consciousness" by Rafael Harth, www.lesswrong.com/posts/NyiFLzSrkfkDW4S7o/why-it-s-so-hard-to-talk-about-consciousness
Using the terminology of this essay, I am a Camp 2 person. If you are a Camp 1 person, what I am saying in the last few paragraphs probably does not make much sense to you.
We do see that in September 2025 Christian Szegedy created math.inc, x.com/ChrSzegedy/status/1966186676289741117, with major breakthroughs www.math.inc/gauss and great vision: www.math.inc/vision
It's not difficult to see how this would lead to trustworthy autonomy and more in the next year, but that's still not pointing to January 2026, that's more like Anthropic timelines. But it will be always difficult to say, "yes, it's now, it's happening".
Sam Altman said a few days ago that their systems are at about "5 hour task difficulty" now (up from GPT-5 2+ hours METR evaluation). With normal development, they will be at about 10 hour task difficulty in January, but I think “trustworthy autonomy” needs more than that.
We do expect Gemini 3.0 in a few weeks joining the wave of competent agents, and it would not be too surprising if it becomes number 1. A strong Grok 5 is a possibility. We do see more orgs doing automated AI researchers in the style of Sakana AI. Things are happening; agents like Codex CLI over GPT-5-Codex and other strong agents are accelerating AI R&D, more and more compute is coming online.
But it does feel that having “trustworthy autonomy” in January 2026 would require something coming "from the left field", something we don't quite expect at the moment, although the trend of two-fold shortening of the interval between qualitative transitions does point at January 2026 as the time of another major transition.
2. Sam Altman is getting together something called Merge Labs, a start-up around high-end non-invasive brain-computer interfaces (non-invasive BCI). It is likely to involve transcranial ultrasound for both read and write. Risks from correctly used transcranial ultrasound are presumed to be mild, but this is a very new tech, so who knows.
This would help us with many thing, and, in particular, with making progress in solving the "hard problem of consciousness" by providing a radically novel experimental platform. Invasive methods like those used by Neuralink are too risky, healthy researchers would rightly hesitate to use them on themselves, permissions are very difficult to get, it's all very slow.
We want methods which are fast, relatively safe, reversible, relatively inexpensive, and which don't need formal permissions from authorities. Non-invasive BCIs are the way to do that, invasive BCIs are a road to nowhere.
Pondering subjective experience, I was just looking at color lights on the street in connection with Winter holidays, and yes, the glowing filament inside the colored glass is "actually" white, but our mind makes a model of the world where the filament itself is colored, and not just the glass of the lamp. That's because the filament looks colored after the light has been filtered by the glass.
Similarly, out brain circuits filter and transform external signals, and that's how a very cool colored subjective reality is created, although the signals transmitted between neurons in the brain are not colored. Yet, the subjective aspects of it, the color "qualia" remain a mystery. We more or less understand how colors are coded in the brain, but the transition from brain dynamics to subjective feelings of qualities is not understood at all; we don't even understand what those subjective feelings of qualities are.
Yes, we sort of understand the information processing and signal processing aspects of it, but we don't understand the "subjective realm".
However, do see this essay: "Why it's so hard to talk about Consciousness" by Rafael Harth, www.lesswrong.com/posts/NyiFLzSrkfkDW4S7o/why-it-s-so-hard-to-talk-about-consciousness
Using the terminology of this essay, I am a Camp 2 person. If you are a Camp 1 person, what I am saying in the last few paragraphs probably does not make much sense to you.
Tags:
