[ExI] Mind Uploads need to be near perfect

[BillK]

Following from the discussion Ben and Jason had about mind uploads, I asked
Grok about "good enough" or "perfect". Grok discussed the 80/20 rule, where
"good enough" and polishing it later was often the best option. But Grok
decided that mind uploads needed to be near perfect on the first attempt,
See discussion below.
BillK

Question -
For future human mind uploads, how near-perfect will be required?

Grok 4.20 Expert -

*Near-perfect fidelity will almost certainly be required* for human mind
uploads (whole brain emulation or WBE), but “near-perfect” is defined at
the *functional scale where mind emerges*, not at the impossible
atomic/quantum level. Absolute 100 % perfect replication of every single
molecule in every instant is unnecessary and probably irrelevant—thanks to
“scale separation” (the brain’s hierarchical nature lets you ignore finer
details once you’ve nailed the right layer).

Here’s the breakdown, based on the most authoritative roadmap (Sandberg &
Bostrom 2008, still the standard reference) plus modern neuroscience and
recent successes:
1. Technical fidelity tiers (what experts actually model)

The classic emulation success ladder:
Level What’s Captured Required for… Approx. Compute Consensus View
3–4 Spiking neurons + full connectome + synapse types/weights Species-typical
behavior, basic learning 10¹⁸ FLOPS Minimum viable “human-like”
5 Electrophysiology (ion channels, voltages, currents) + ~5×5×50 nm spatial
scan Individual personality, memories, skills 10²² FLOPS Probably
sufficient for most practical uploads
6–7 Metabolome + proteome (neurotransmitters, proteins, gene expression) Full
plasticity, long-term adaptation, qualia/consciousness 10²⁵–10²⁶ FLOPS Safe
“near-perfect” zone for convincing uploads
8–11 Protein complexes → single molecules → quantum Only if no scale
separation exists Insane (10³⁰+) Almost certainly overkill; quantum effects
in brain are negligible on relevant timescales

*Spatial resolution needed*: ~5 nm in key directions to see synapses and
thin axons (current electron microscopy on small brains already hits this;
fruit-fly whole-brain connectome + synapse prediction already produces
multiple real behaviors at 95 % accuracy). *Temporal*: <1 ms to capture
spikes and synaptic timing. *Chemical tags*: Which neurotransmitter,
receptor type, phosphorylation state—critical for “who you are.”

Kurzweil says neural/synaptic level is enough for uploads by ~2045.
Sandberg/Bostrom workshop consensus: levels 4–6 get you to “individual
brain emulation” (recognizably *you*, retains most memories/skills) and
likely “mind emulation” (subjective experience). Higher levels only if the
brain has no useful scale separation—which would make uploading impossible
anyway.

Recent proof: The 2024–2026 fruit-fly whole-brain emulation (125k neurons,
50M synapses) already drives real embodied behaviors when plugged into a
physics-simulated body. That’s level ~4–5 fidelity—and it *works*. Human
scale just needs the same approach, bigger.
2. Philosophical “is it really YOU?” requirement

This is where “near-perfect” gets stricter:

   - *Functionalist / pattern view* (most transhumanists, Koene, Kurzweil,
   Parfit-inspired branching identity): If the *information-processing
   pattern* is preserved at the right scale, the upload *is* a continuer of
   you. A perfect synaptic + dynamics copy is sufficient. Small approximations
   below noise level don’t matter—you don’t notice which exact molecules are
   firing right now anyway.
   - *Gradual replacement* (neuron-by-neuron swap): Highest
   continuity—widely accepted as preserving identity even by skeptics.
   Requires near-perfect tech but no destructive scan.
   - *Destructive scan-and-copy*: The upload wakes up thinking it’s you 100
   %, but the original dies. Many philosophers say it’s a perfect copy, not
   “you.” Psychological branching or psychological-continuity theories still
   say both branches can legitimately be “you.”
   - *Substrate chauvinist / quantum views* (Penrose, some dualists): You’d
   need literally perfect biology-level fidelity (or impossible quantum state
   transfer). Most experts dismiss this—decoherence kills quantum effects in
   <10⁻¹³ s.

*Bottom line on identity*: Near-perfect at level 5–6 + gradual method or
very high-fidelity non-destructive scan is what most serious thinkers would
accept as “safe enough” for personal continuation. Anything coarser risks
noticeable differences (amnesia, personality drift, or the dreaded “zombie”
worry).
3. Why “good enough” is *not* better than near-perfect here (unlike your
last question)

   - *One-shot high stakes*: Unlike a blog post or MVP, you can’t easily
   iterate on your own mind after upload. Errors compound (chaos + long-term
   plasticity). Early uploads will need over-engineering.
   - *Environment must also be near-perfect*: Sensory input and virtual
   body simulation must be accurate enough that the upload can’t tell the
   difference—otherwise psychosis or identity fracture. Minor distortions =
   serious mental consequences.
   - *Once mature*: Then “good enough + fast iteration” wins. Uploads can
   be copied, edited, sped up, and A/B tested ethically in simulation. The
   first successful ones just have to be close enough that the emulated person
   says “yep, still me.”

Practical threshold most experts converge on

   - Scan at ~5–50 nm spatial + chemical tagging + 1 ms dynamics.
   - Simulate at level 5–6 fidelity (synapses + electrophysiology + key
   chemistry).
   - Use gradual replacement or non-destructive methods where possible.
   - Validate until the emulation passes personalized Turing tests *and*
   internal self-report matches (under PAM—Principle of Assuming the Most).

That’s “near-perfect” in engineering terms: accurate to within natural
brain noise, preserving every functionally relevant bit. Absolute molecular
perfection? Unnecessary and wasteful.

In short: For mind uploads, *near-perfect at the synaptic-to-molecular
functional layer will be required*—because the mind is delicate and
identity is precious. But once you hit that threshold, the same principle
you asked about last time kicks in: iterated “good enough” versions will
rapidly outpace any biological human. The first upload has to be *really*
close; everything after that gets to be gloriously imperfect and improvable.

Motion (in silicon) will beat meditation (in carbon). But you still need to
nail the launch.
-----------------------------------
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20260321/b46bade2/attachment.htm>