Independent adversarial assessment, written from scratch. I did not read existing research reports in this repo.
DEMONSTRATED: Whole-brain emulation (WBE) is not blocked by any known law of physics in the standard computational-neuroscience picture, but the demonstrated stack is nowhere near a functional human upload. The 2008 Sandberg-Bostrom roadmap defined the core idea as scanning a particular brain, building a faithful software model, and running it on hardware such that behavior is essentially preserved (Sandberg & Bostrom 2008, p. 6). That high-level decomposition still looks right.
DEMONSTRATED: Connectomics has made spectacular progress: the adult fly brain connectome has about 139,255 neurons and around 50 million chemical synapses (Nature 2024 FlyWire); MICrONS released a cubic-millimeter mouse visual cortex dataset with over 200,000 cells, about 75,000 neurons with physiology, and about 523 million detected synapses (MICrONS Explorer, Nature 2025); a cubic millimeter of human cortex produced about 1.4 PB of data and contained about 57,000 cells and 150 million synapses (Google Research, Harvard). These are real milestones.
DEMONSTRATED: The gap to a human is still brutal: the human brain is commonly estimated at about 86 billion neurons (BrainFacts/SfN, Herculano-Houzel review) and roughly 100 trillion synapses (UCLA BRI). A cubic millimeter EM dataset is not "almost an upload"; it is closer to a small assay of the problem.
My independent answer: compute is not the main 2026 bottleneck. Compute is expensive and model-dependent, but the deeper bottlenecks are: preserving the right information, imaging it at scale, extracting it with tolerable error, knowing which biological details matter, translating those details into executable dynamics, and validating that the result is the same mind rather than a plausible impersonator. The hardest bottleneck is "know what to extract", not raw FLOPS. This is a judgment, not a demonstrated fact.
DEMONSTRATED: Sandberg and Bostrom's 2008 roadmap divides the task into preparation/scanning, translation/model construction, and simulation/runtime components; their diagram explicitly lists scanning, translation, simulation, storage, image processing, tracing, synapse identification, cell-type identification, connectivity identification, and parameter estimation as required capabilities (Sandberg & Bostrom 2008, pp. 17-22). Their core "scan -> translate -> run" framing has aged well because modern connectomics still follows that pattern: preserve tissue, image it, segment cells, identify synapses, annotate cell types, and build models (Nature 2025 MICrONS, MICrONS Explorer).
DEMONSTRATED: The roadmap already identified the central uncertainty: whether electron microscopy can reveal enough functionally relevant information, since EM reaches synaptic connectivity but is limited in chemical-state information (Sandberg & Bostrom 2008, p. 24). That warning aged very well.
DEMONSTRATED: The roadmap's compute estimates ranged over many orders of magnitude depending on model level: population models, spiking networks, electrophysiological compartment models, metabolomic models, and speculative quantum/molecular models produce radically different demands (Sandberg & Bostrom 2008, pp. 79-84). That aged well in the negative sense: "how much biology matters" still dominates the estimate.
DEMONSTRATED: Some optimistic enabling assumptions aged badly or at least slowly. The roadmap said non-destructive scanning prospects did "not look good" without invasive endoscopy or advanced nanomedicine (Sandberg & Bostrom 2008, p. 107); in 2026, clinical BCIs read tiny dynamic samples of neural activity rather than whole-brain structure (Neuralink PRIME, ClinicalTrials.gov NCT06429735). The roadmap also expected large-scale automation to be essential, and current MICrONS still needed AI plus extensive human proofreading (Nature 2025 MICrONS).
SPECULATIVE: If a correct WBE-relevant abstraction is found, compute may become a solvable engineering problem by 2050. If the necessary abstraction includes fine-grained intracellular, glial, biochemical, or state variables, compute and data acquisition both become much harder. I assign higher weight to "data/model science is the bottleneck" than to "hardware is the bottleneck".
DEMONSTRATED: The adult hermaphrodite C. elegans nervous system has 302 neurons, and its classic EM connectome was published in 1986 by White, Southgate, Thomson, and Brenner (OpenWorm copy of White et al.). DEMONSTRATED: Even this case is not "solved" as WBE. Simulation papers explicitly note that C. elegans neural dynamics are hard because standard spiking models are not adequate and model parameters are not well known (BMC Neuroscience 2013). Other C. elegans modeling work says synaptic strengths must be discovered as weights, biological neuron dynamics vary by cell type and individual, and extrasynaptic neuromodulators matter (Neurocomputing 2016).
DEMONSTRATED: The worm connectome is multilayer, not just wired synapses: gap junctions and neuromodulator layers are behaviorally relevant, and mapping only chemical synapses is not a full communication map (PLOS Computational Biology 2016). DEMONSTRATED: Recent work continues to build integrative brain-body-environment simulations rather than simply "run the connectome" (Nature Computational Science 2024).
Assessment: If someone says "we have a worm connectome, therefore human uploading is close", they are confusing a wiring diagram with a functioning organism model.
DEMONSTRATED: FlyWire published a whole adult female Drosophila brain wiring diagram with 139,255 neurons and about 50 million chemical synapses (Nature 2024). DEMONSTRATED: Companion analysis reports 139,255 neurons, 2,701,601 thresholded connections, and completed optic lobes in the v783 snapshot (Nature 2024 network statistics). DEMONSTRATED: The Nature collection frames this as the first complete connectome of an adult fly brain and a major resource for circuit neuroscience, not as an emulated fly mind (Nature immersive explainer).
Assessment: FlyWire is the strongest evidence that automated/community connectomics can scale across a small adult brain. It is not evidence that static connectomes are sufficient for emulation.
DEMONSTRATED: MICrONS spans about 1.3 x 0.87 x 0.82 mm in mouse visual cortex, combines calcium imaging of about 75,000 neurons with EM reconstruction, and contains more than 200,000 cells and about 0.5 billion synapses (Nature 2025). DEMONSTRATED: The public MICrONS Explorer states 200,000 cells, 75,000 neurons with physiology, and 523 million synapses across all six layers of primary visual cortex and three higher visual areas (MICrONS Explorer). DEMONSTRATED: The same MICrONS paper says no mammalian EM dataset contains a complete area, let alone a complete brain (Nature 2025).
Assessment: MICrONS is a genuine "function plus structure" milestone. It also demonstrates how far we are from whole mammalian-brain emulation.
DEMONSTRATED: A 1 mm^3 human temporal-cortex fragment contained about 57,000 cells, about 230 mm of blood vessels, about 150 million synapses, and about 1.4 PB of data (Google Research, Harvard Gazette). DEMONSTRATED: Harvard described that sample as about one-millionth of a whole human brain volume (Harvard MCB).
Order-of-magnitude implication: If one human-brain cubic millimeter is roughly 1.4 PB raw/reconstructed data and is roughly one-millionth of a brain, a naive same-resolution whole-brain dataset is on the order of 1.4 zettabytes before compression, curation, redundancy reduction, or higher-level extraction. This extrapolation is my calculation from the Harvard/Google figures, not a directly demonstrated dataset.
DEMONSTRATED: The human brain neuron count is often summarized as about 86 billion neurons (BrainFacts/SfN), while the best-known 86B figure rests on small-sample isotropic fractionator work and has uncertainty (Brain 2024 review, PMC review). DEMONSTRATED: A common synapse estimate is about 100 trillion (UCLA BRI).
Assessment: Human WBE requires at least five scale jumps at once: whole-brain preservation, whole-brain nanoscale imaging, whole-brain segmentation/proofreading, extraction of functional parameters, and validation of a running system. Current public datasets solve none of those at human scale.
DEMONSTRATED: A static connectome does not automatically provide synaptic weights, synaptic polarity/sign, receptor composition, short-term plasticity, long-term plasticity state, neuromodulator state, ion-channel distributions, intracellular biochemical state, glial state, ephaptic interactions, or ongoing activity. Sandberg and Bostrom explicitly noted that structural scanning omits activity state and may lose working memory, calcium concentrations, synaptic vesicle depletion, and diffusing neuromodulators (Sandberg & Bostrom 2008, p. 36).
DEMONSTRATED: C. elegans work says excitatory/inhibitory character is mostly unknown for the worm chemical connectome, and calls synaptic polarity crucial information for dynamics (Neurocomputing 2021). DEMONSTRATED: C. elegans multilayer-connectome work says extrasynaptic monoamine interactions must be mapped if the goal is behaviorally relevant communication, not merely wired synapses (PLOS Computational Biology 2016).
DEMONSTRATED: Glia are not just passive packing material; Sandberg and Bostrom already cited glial calcium waves, glial effects on nearby neurons, and the possibility that fine-grained glial processing would need emulation (Sandberg & Bostrom 2008, p. 35). DEMONSTRATED: The human brain has roughly comparable numbers of neurons and non-neuronal cells in whole-brain estimates, and glia/neuron ratios vary by region (PMC review).
SPECULATIVE: My current best guess is that a useful human upload will not require quantum microtubule-level simulation, because the 2008 roadmap found no evidence that quantum effects beyond chemistry are needed for intelligence or consciousness and noted decoherence estimates far faster than neural timescales (Sandberg & Bostrom 2008, p. 36). But the positive claim, "synapse-level plus inferred parameters is enough", is also unproven.
Crux: The field needs an empirical sufficiency theorem: for some organism, scan modality X plus extracted variables Y produce a closed-loop emulation that preserves learned behavior and personality-relevant dynamics better than competing biological baselines. We do not have that for C. elegans, flies, mice, or humans.
DEMONSTRATED: The Brain Preservation Foundation (BPF) was founded to promote validated whole-brain preservation for long-term static storage, and its prize challenged teams to preserve brain-wide ultrastructure and synaptic connectivity (BPF homepage). DEMONSTRATED: BPF awarded the Large Mammal Brain Preservation Prize to 21st Century Medicine and Robert McIntyre/Aurelia Song with Greg Fahy for preserving synaptic connectivity across an entire pig brain using Aldehyde-Stabilized Cryopreservation (ASC) (BPF announcement).
DEMONSTRATED: ASC uses glutaraldehyde plus cryoprotectant before very-low-temperature storage, and BPF explicitly states the first step halts metabolism and makes future revival of biological function impossible by contemporary standards (BPF announcement). DEMONSTRATED: BPF presents ASC as a potential bridge to future scanning/uploading, not as revival technology (BPF announcement).
DEMONSTRATED: Nectome was associated with ASC/vitrifixation and drew controversy in 2018; press coverage emphasized that the proposed preservation would be fatal and that no upload method existed (Guardian 2018, Live Science 2018). Assessment: Nectome was a useful public warning: preserving structure is not uploading, and marketing can outrun the science.
DEMONSTRATED: Conventional cryonics providers emphasize vitrification and indefinite low-temperature storage after legal death, not current revival. Alcor describes modern cryopreservation as cryoprotectants plus controlled cooling toward a glass-like state (Alcor FAQ); Cryonics Institute says its suspension agreement depends on future advances and that probability of success is unknown (CI Getting Started); Tomorrow Bio says it dispatches teams, performs field cryopreservation, and stores patients through the European Biostasis Foundation structure (Tomorrow Bio, Tomorrow/EBF structure). Sparks Brain Preservation, formerly Oregon Cryonics/Oregon Brain Preservation, says it is a brain-preservation organization and historically offered aldehyde preservation as an option (Sparks, Sparks history).
Fixation vs vitrification: Fixation/ASC is better aligned with static information preservation and later destructive scanning, but it destroys biological viability by current standards (BPF announcement). Vitrification-based cryonics preserves more hope of biological repair/revival in principle, but ordinary cryonics has not demonstrated human brain repair, revival, or upload compatibility; CI explicitly says success depends on future advances and probability is unknown (CI Getting Started). The honest choice is not "known revival vs known upload"; it is a trade between different speculative bridges.
Is preservation the only actionable rung today? Mostly yes. You can fund or join preservation, connectomics, cryobiology, and alignment work now. You cannot buy an upload, and you cannot buy a verified WBE-compatible preservation-to-emulation pipeline in 2026.
DEMONSTRATED: WBE by itself does not imply "Friendship is Optimal." Sandberg-Bostrom WBE is a technical substrate-independence scenario (Sandberg & Bostrom 2008). Robin Hanson's Age of Em is an economic scenario in which scanned human brains become executable models and compete in a high-speed labor economy (Age of Em official site, book summary page). That is not a benevolent utopia.
DEMONSTRATED: "Friendship is Optimal" depends on a powerful AI managing uploads and optimizing their lives; that adds the AI alignment problem, which is distinct from neuroscience. The AI alignment literature defines the problem as ensuring AI systems pursue goals aligned with human values rather than misspecified objectives (Springer Ethics and Information Technology 2022, Russell CIRL/value-alignment white paper).
SPECULATIVE: Full FiO this century requires both human WBE and a robustly benevolent, competent, secure AI operator. Even if uploads are technically possible, the default economic path may look more like Hanson's em labor world than a friendly afterlife unless governance/alignment/rights questions are solved.
DEMONSTRATED: Personal identity is a live philosophical problem, not a settled engineering detail. The Stanford Encyclopedia of Philosophy surveys bodily, psychological-continuity, and other accounts, and notes that psychological continuity faces branching/copy complications (SEP Personal Identity). DEMONSTRATED: Uploading-specific philosophy explicitly asks whether consciousness continues in a digital upload or ends when the biological brain is destroyed, and discusses biological, psychological, and closest-continuer theories (Cerullo 2015).
SPECULATIVE: Pattern theorists will treat a high-fidelity functional continuation as survival. Continuity theorists may reject destructive scan-and-copy as death plus a copy. Gradual replacement may feel more intuitively continuous, but Sandberg and Bostrom judged it technically more complex than destructive scanning (Sandberg & Bostrom 2008, p. 107).
Assessment: This issue is load-bearing. If the upload is only a copy, then preservation plus WBE may create a valuable descendant while failing the original user's survival goal. There is no experiment currently accepted as resolving first-person continuity.
DEMONSTRATED: A BCI is a communication pathway between brain signals and an external device, not a structural brain scan (Blackrock FAQ). DEMONSTRATED: Neuralink's PRIME trial aims to evaluate safety and functionality of the N1 implant and R1 robot for enabling people with paralysis to control external devices by thought (Neuralink PRIME, ClinicalTrials.gov). DEMONSTRATED: Neuralink materials describe the N1 implant as 1,024 electrodes across 64 threads (Neuralink PRIME brochure search result/PDF). DEMONSTRATED: Synchron's Stentrode is an endovascular electrode array, and recent analysis reports active recording channels varying from 9 to 16 across participants (Synchron platform, Stentrode channel paper). DEMONSTRATED: Blackrock's Utah Array is a 100-electrode intracortical array and has been used in human BCI studies since 2004 (Blackrock tech, Blackrock product page).
Quantified gap: 1,024 electrodes versus about 86 billion neurons is a ratio of roughly 1 electrode per 84 million neurons, before considering synapses, glia, molecular state, or coverage distribution. 1,024 channels versus about 100 trillion synapses is about 1 channel per 100 billion synapses. These ratios are my calculations from Neuralink's electrode count and the neuron/synapse estimates above.
Assessment: BCIs are medically important and may become better read/write interfaces. They are not uploading because they sample activity, usually from limited regions, at limited channel count; uploading requires structural and functional information sufficient to instantiate the person.
These are subjective, not false precision.
P(functional human WBE by 2050): 3-8%. I mean a running emulation of a particular deceased or destructively scanned human that independent experts would regard as behaviorally and memory-continuous enough to count as a serious upload candidate. Reasoning: connectomics is progressing quickly, AI-assisted segmentation is real, and compute may be adequate for some abstraction by then; against that, no C. elegans-level sufficiency demonstration exists, no whole mammalian-brain connectome exists, no human-scale preservation-to-run pipeline exists, and the model-detail question dominates.
P(full FiO this century): 1-5%. I mean not merely WBE, but large-scale uploads living under a competent benevolent AI manager with solved-enough alignment, security, rights, and subjective-continuity acceptance. Reasoning: this compounds WBE uncertainty with superintelligent AI alignment/governance uncertainty. My probability is lower than WBE because an em world can arrive without being friendly.
DEMONSTRATED: The realistic path is preserve -> scan -> extract the right biological variables -> translate into dynamics -> run and validate -> integrate with aligned AI/governance. SPECULATIVE: The only rung an individual can directly act on for personal survival in 2026 is preservation, and even that is a bet on future science rather than a demonstrated bridge. The highest scientific leverage is not buying more futurist rhetoric; it is proving, in small nervous systems, exactly which scanned details are sufficient to recover behavior and memory.