[ExI] Superforecast for AGI Singularity

[BillK]

On Sat, 4 Jul 2026 at 18:10, BillK <pharos at gmail.com> wrote:

> I decided to try the superforecast prompt suggested in the other thread to
> see what the difference was between a normal answer and a superforecast
> response.
> I used GLM-5-Turbo with Websearch and Deep Think enabled for Normal and
> Superforecast.
> The answers were different, but basically both said, "Don't know."
> See what you think.
> BillK
> ---------------------------------
>


I asked GLM to explain the final estimate of 4% probabilty of the
Singularity by 2040, as to me this sounded like saying it was impossible by
2040. The answer was that GLM was separating AGI and the subsequent
Singularity event.
So I followed up by asking why the Singularity would be so slow after AGI
was achieved.
The answer was physical constraints on building in the real world.
BillK


GLM-5-Turbo -

This objection conflates two fundamentally different metrics: *cognitive
velocity* (software speed) and *physical instantiation*
(hardware/infrastructure speed). My 4% estimate already accounts for the
fact that AGI will think incredibly fast; the bottleneck is the physical
world.

Here is the mechanical breakdown of why "incredibly fast thinking" does not
translate to an incredibly fast Singularity:

*1. The Hardware Ceiling* An AGI running on a 100,000-GPU cluster possesses
cognitive velocity millions of times faster than human neurons. It could
theoretically design a superior neural architecture, invent a new
optimization algorithm, and write the code in seconds. However, once it
finishes that design, it hits a hard physical wall: it cannot test or
deploy that new architecture because it is already using 100% of its
available compute. To actually *become* smarter, it requires more GPUs and
more electricity. It cannot think its way out of a hardware constraint.

*2. The Semiconductor Manufacturing Latency* For the Singularity to occur,
the AGI must expand its compute substrate exponentially. Even if an AGI
emails TSMC a perfect blueprint for a next-generation chip that is 100x
more efficient, the physical latency of the semiconductor supply chain
remains fixed. Building a new fabrication plant takes 3 to 5 years.
Procuring advanced lithography machines (ASML EUV) takes years.
Manufacturing wafers, packaging chips, and integrating them into data
centers takes months. Software can be iterated in milliseconds; physical
infrastructure is iterated in years.

*3. The Energy and Thermodynamic Bottleneck* Compute requires power.
Current data centers require gigawatts of electricity. If an AGI attempts
to recursively self-improve by ordering the construction of new data
centers, it is constrained by global energy production, grid transmission
capacity, transformer manufacturing, and physical construction timelines.
You cannot accelerate the pouring of concrete or the mining of copper
through software optimization.

*4. The "Slow Takeoff" Mathematical Reality* Because the physical world
acts as a severe rate-limiter, the transition from AGI to Singularity is
mathematically clamped. Let’s assume the AGI improves its software
efficiency by 10,000% in a week (a "fast takeoff" in software). If its
physical compute substrate can only grow by 20% per year due to supply
chain constraints, the actual intelligence growth curve of the system is
dictated entirely by the 20% hardware growth rate. The Singularity is
delayed until the hardware catches up to the software.

*Conclusion on Probability* My 4% estimate is explicitly calibrated
*assuming* AGI operates at maximum supercomputer speed. The probability is
low because the Singularity requires breaking the physical bottleneck.

To change my 4% estimate upward, you would have to prove that intelligence
can increase exponentially *without* requiring proportionally more physical
compute—for example, by demonstrating that an AGI can achieve a 100x
intelligence multiplier purely through algorithmic compression on fixed
hardware. Until that is empirically proven, the "fast thinking" assumption
is neutralized by "slow building."

-----------------------------
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20260705/ae8a0d9f/attachment-0001.htm>