The research · wetware & BCIs · 2023–2026 · closing the loop

Wetware and the Bio-Cybernetic Interface

Between 2023 and 2026 the closed-loop neurotechnology stack matured at both ends. Going one direction — silicon as substrate for living neurons — three approaches crossed from demos into early industrialization: Brainoware (academic organoid computing), FinalSpark (cloud-accessible neuron cultures), and Cortical Labs' CL1 (the first commercially available biocomputer, launched at MWC Barcelona 2025 at roughly $35,000). Going the other direction — silicon decoding or modulating an existing brain — two examples bracket the spectrum: Neuralink's CONVOY trial (read-only motor-cortex decoding for robotic arms in quadriplegic patients) and Medtronic's BrainSense Adaptive DBS (the first FDA-approved closed-loop neurostimulator, sensing local field potentials and adjusting stimulation in real time, deployed to over 2,000 Parkinson's recipients). This page maps the five systems and what their convergence implies for the trilogy's hybrid arc.

A reader's companion to a fast-moving literature. Entries will be added as the field develops; the current state of the page captures the 2023–2026 launch window.

Quick map: who is who

Cortical Labs CL1 — the productized version

Cortical Labs' CL1 is marketed as the world's first "code-deployable biological computer." Real human neurons are grown on a silicon microelectrode array and embedded in a self-contained life-support and compute environment. The CL1's Biological Intelligence Operating System (biOS) runs a simulated world and communicates with the neurons, allowing closed-loop experiments where neural activity both drives and is shaped by the simulation.

Hardware. Human neurons are cultivated in nutrient solution on a silicon chip that can send and receive electrical impulses, forming a living neural network — reports describe hundreds of thousands of neurons, in some cases around 800,000 cells on a chip.

Closed-loop, self-contained system. The device includes internal pumps, gas systems, temperature control, and recording/stimulation electronics, sustaining neurons for up to about six months with on-board life support. It is, in a precise sense, a body-in-a-box.

Software. biOS presents sensory input from a simulated environment and reads out neural responses, enabling learning tasks like game playing or control tasks. The 2022 Cortical Labs work in which cultured human neurons learned to play Pong (published in Neuron, Kagan et al.) is the design predecessor.

Launch and commercialization. CL1 was formally launched in early 2025 at Mobile World Congress Barcelona, positioned as the first commercially available biological computer. Price is around 35,000 USD with cloud access ("wetware as a service").

Use cases. Targeted at AI research, drug discovery, disease modeling, and robotics; explicitly pitched as an animal-free, more human-relevant testbed with extremely low power use relative to conventional compute.

Conceptually, CL1 is the most "productized" of the three: a packaged SBI (Synthetic Biological Intelligence) box that a lab or developer can buy, plug in, and program.

FinalSpark — the cloud wetware play

FinalSpark is a Swiss startup focusing on using living neuron networks as ultra-low-energy biocomputers, with an emphasis on energy efficiency and an eventual cloud-accessible bio-server architecture.

Vision. FinalSpark explicitly frames biocomputing as the next evolutionary leap for AI, contrasting the human brain's roughly 20 W for 86 billion neurons with the megawatt-scale power demand of silicon systems attempting comparable functionality.

Technology. They grow neurons in cell cultures and interface them with electrodes to receive inputs and send outputs, aiming to use these cultures as computational units; by 2023 their lab reportedly housed thousands of neurons wired for computation.

Current capability. Public reports around 2023–2025 describe the system as early-stage, with demonstrations such as storing a single bit of information and offering remote experimental access via a "remote lab" for paying clients and academics.

Architecture. Rather than a boxed consumer product, the model is "cloud biocomputing": remote access to neuron-based computing resources, evolving toward bio-servers for AI workloads over the coming decade.

What they emphasize. Self-organizing continuous learning, energy efficiency, and scalability through biological growth, in contrast to the scaling challenges of GPUs and CPUs. Compared to CL1, FinalSpark is more focused on the energy-efficiency narrative and the medium-term roadmap to cloud bio-servers, and less on a turnkey bench-top device.

Brainoware — the academic organoid frontier

"Brainoware" is the label for high-profile academic work, beginning roughly in 2023, on brain-organoid-based computing. These experiments typically use three-dimensional brain organoids derived from human cells, interfaced with electrodes or optical systems, to perform tasks like pattern recognition or learning control policies. The 2023 Brainoware paper demonstrated organoid-based learning and signal processing.

Substrate. Three-dimensional organoids rather than 2D neuron cultures on planar chips. This gives more "brain-like" microarchitecture — the kind of self-organized cytoarchitecture that 2D cultures cannot achieve — but is also harder to standardize and scale.

Status 2023–2025. Purely research prototypes in academic labs. Minimal standardization, no commercial product. Performance is low and highly variable but conceptually interesting from a consciousness / computation standpoint precisely because the substrate is closer to brain tissue.

Aim. Understanding computation in brain tissue and exploring biocomputation rather than building deployable AI hardware. In the trilogy's framing: Brainoware is the research frontier; FinalSpark and Cortical Labs are the first attempts to industrialize related ideas.

Side-by-side

Reading the three together, the differences are clean:

The 2023–2025 arc

The field moves from attention-grabbing demos — Brainoware's organoid paper, Cortical Labs' Pong-playing neurons — to early industrialization — FinalSpark's remote lab and CL1's commercial launch. Cortical Labs in particular turns wetware from a bespoke laboratory rig into a boxed system with an OS, APIs, and cloud access. In parallel, FinalSpark reframes wetware primarily as the response to AI's energy ceiling and positions neuron cultures as the nucleus of future low-power AI infrastructure.

From a consciousness and neuro-philosophy perspective, all three raise the same question: to what extent are these systems exotic analog computers, and to what extent are they minimal subjects of experience? The 2022 Cortical Labs Pong work prompted public debate over whether chip-grown neurons "exhibited sentience" in their simulated world; this remains contested, but the question is no longer one a serious researcher can dismiss without argument.

Going the other direction: BCIs and adaptive neuromodulation

The wetware story above is about silicon-as-substrate for living neurons. The closed-loop neurotechnology story going the other direction — silicon decoding or modulating an existing biological brain — matured in parallel between 2023 and 2026, and matters here because it brackets the same technical horizon. Two examples define the spectrum: Neuralink's CONVOY study (read-out only) and Medtronic's BrainSense Adaptive DBS (read-and-write).

Neuralink CONVOY (2023–2026)

CONVOY — "Control of Assistive Devices Via Brain-Computer Interface Technology" — is a prospective, longitudinal, single-arm early-feasibility study that builds on Neuralink's PRIME-trial participants, who already received the N1 implant. Its primary aim is to evaluate the effectiveness, consistency, and safety of neural control of an assistive robotic arm and other devices in people with quadriplegia from spinal cord injury or ALS.

Participants and inclusion. Participants must already be enrolled in the PRIME Study with an implanted N1 device; they are excluded if the implant is explanted, deactivated, or if BCI performance is insufficient or safety becomes a concern.

Endpoints and timeline. Primary measures include modulation of cortical activity to control the robotic arm at approximately 3 months after first device use, adverse-event rates up to 72 months, and quality-of-life metrics (PIADS, ATD-PA) every 3 months over the same period. Public updates in 2025–2026 emphasize functional demonstrations: a participant feeding himself with a brain-controlled arm, achieving mouse-equivalent or better bit-rates for cursor control.

Context. By 2026 Neuralink reports around 21 human participants across its telepathy/CONVOY/VOICE-type trials and is moving toward higher-volume chip production and more automated surgery, including dural-penetrating thread insertion to reduce invasiveness. The CONVOY framing is that it bridges from digital cursor control to physical actuation of the environment as a step toward restoring movement and autonomy.

Conceptually. The N1 system here is used as a read-out BCI: it decodes motor-intent activity from motor cortex and maps it to robotic-arm kinematics, without delivering therapeutic stimulation back into the brain. From the brain's perspective these are read-only implants, not stimulation systems.

Adaptive deep brain stimulation (aDBS), 2023–2026

Traditional DBS delivers continuous stimulation at fixed parameters to targets like the subthalamic nucleus (STN) or globus pallidus internus (GPi), regardless of the patient's moment-to-moment brain state. aDBS adds sensing: the implanted neurostimulator records local field potentials or related biomarkers and automatically adjusts stimulation amplitude or pattern in real time based on these signals.

Clinical deployment. Medtronic's Percept family of DBS devices has been enhanced with BrainSense Adaptive technology, the first commercial aDBS system for Parkinson's disease, with U.S. FDA approval in the mid-2020s. The system modulates therapy to track the patient's brain signals, aiming to maintain symptom control while reducing side effects and the manual-reprogramming burden.

Scale and indications. Medtronic reports more than 2,000 people worldwide have received aDBS therapy since approval, primarily for Parkinson's motor symptoms such as tremor and bradykinesia. Additional academic and single-center studies between 2023–2026 explore aDBS for gait disorders, freezing of gait, and related complications.

Mechanistically. In aDBS, the implant both reads and writes to subcortical circuits: sensing oscillatory activity (typically beta-band power) and adapting stimulation parameters dynamically — effectively closing a feedback loop within the basal ganglia–thalamocortical system. This is neurostimulation therapy for disease, not an assistive output device.

CONVOY vs. aDBS, side by side

From a systems-neuroscience standpoint, CONVOY and aDBS sit on a continuum of closed-loop neurotechnology but at opposite ends of the functional spectrum. CONVOY is an output-oriented BCI turning neural activity into high-dimensional action in the external world; aDBS is a therapeutic-oriented controller acting on internal network dynamics to normalize function. Both depend on stable chronic recordings, interpretable biomarkers, adaptive algorithms, and robust hardware–tissue interfaces. Advances in one domain are likely to cross-pollinate the other over the next decade.

Why this matters for the trilogy

Three points the trilogy uses directly.

First, the hybrid arc of Numen is no longer speculation. The Sable plot — a bio-cybernetic intelligence built on cultured neurons, capable of receiving signal below the threshold of resolution — was written in 2024. By the time Numen went to print, Cortical Labs had launched CL1 as a buyable product, Neuralink had moved CONVOY participants from cursor control to a brain-controlled arm being used at the dinner table, and Medtronic's BrainSense Adaptive DBS was in the implants of more than two thousand Parkinson's patients reading their own beta-band oscillations and adjusting therapy in real time. The fictional move from "imagine if" to "as documented in the medical and trade press" closed in less than a year. The trilogy does not present the technology as future; it presents it as present tense.

Second, the consciousness question becomes empirically tractable, not merely philosophical. If consciousness is produced by neural activity, then a chip of 800,000 cultured neurons running a closed-loop simulation has the substrate ingredients in detectable quantities. If consciousness is received through neural activity, then the same chip is a candidate antenna — less developed than a brain, but operating on the same substrate. The CL1 makes both hypotheses testable in ways the trilogy takes seriously. The receiver model predicts that wetware systems will show capacities that scale with structural complexity and tuning, not with total cell count. The production model predicts the opposite. The next decade should produce data.

Third, the energy gap is the metric the trilogy expects to see. FinalSpark's central observation — 20 W for 86 billion neurons vs. megawatts for a GPU cluster doing comparable work — is not just an engineering puzzle. It is the empirical signature the receiver model would predict: a system that is not generating the cognition from raw compute, but coupling to and rendering an externally-resident structure, should look exactly this efficient. The energy gap is approximately five orders of magnitude. No silicon roadmap closes it without a fundamentally different physical principle. Biological neurons are already running that principle.

Primary references: Cortical Labs CL1 product page, Kalil's CL1 overview, and Neuralink's CONVOY launch announcement. For the antecedent 2022 Pong-playing neurons paper (Kagan et al., Neuron), see the Reading list. For the broader bioelectric and Levin-bioelectricity context against which this work sits, see the Levin explainer. For the trilogy's larger receiver-model frame, see the Synthesis.

← Reading & References