Your brain runs at about 10 bits a second, and no wire changes that.
The obvious next move is to skip the keyboard and wire the brain straight into the machine. A cable can carry billions of bits a second, so the wire has bandwidth to spare. One number makes all of that moot: the brain emits about 10 bits per second of deliberate meaning, whether those bits leave through your fingers, your mouth, or a chip in your skull. That is the cap, and no wider wire raises it.
That ceiling decides the future of knowledge work. Last time, I argued that attention is the finite, un-buyable bottleneck on agent work. Here, it runs into the hardware built to route around it, and the hardware loses.
Type all day and you emit under a byte a second
The unit is information, not keystrokes. English compresses to about 1 bit per character once you account for how predictable it is (Shannon 1951). Type at the median 51.6 wpm (Dhakal 2018) and you run 51.6 ÷ 60 × 5.5 × 1 ≈ 5 bits/s. A full workday of typing carries less unique information than a few seconds of video. Speech barely beats it where it counts: the meaning rate is 10-13 bits/s (Zheng & Meister 2025). The linguistic signal itself clocks ~39 bits/s across 17 languages (Coupé 2019), but once you strip out what’s predictable, the novel meaning is far less.
So wire your brain straight in
Right now, a thought takes a detour: you think it, turn it into words, and then push those words out through your fingers or your voice. Delete that middle stretch, and you get thought → computer, or thought → another brain. A chip on the motor cortex isn’t limited by finger speed. On paper, the channel widens by orders of magnitude. As a wire, the promise is real. So check whether your fingers or your voice were ever what held us back.
The eyes already deliver a billion bits a second. You can’t process it.
Your eyes push about 10⁹ bits/s at your brain, but the meaning you actually work with is only about 10 bits/s (Zheng & Meister). That is eight orders of magnitude lost somewhere between the retina and the thought. Reading runs a little wider at 28-45 bits/s, and listening sits around 13, but they are in the same order of magnitude - that is the words arriving, not the meaning landing, which stays near the ~10 bits/s we started with. Taking meaning in is nearly as narrow as pushing it out. Both sit far below the wire.
The proof is already strapped to your face. Your eyes are the highest-bandwidth input channel you will ever own. A wider input wire - a chip, a feed, a sharper screen - runs straight into the wall the eye already hits. The bottleneck sits downstream, in the funnel from sensation to meaning - the same funnel, whichever way the data flows.
One scope note this piece rests on: this is deliberate, conscious meaning you could report, not the unconscious sensorimotor traffic your brain runs in the background. Call it ~10 bits/s, good to a small factor. And the source is what’s slow, not the channel. Switch from fingers to voice - a far faster output - and the deliberate rate barely moves (~13 bits/s).
Hard thinking barely moves the meter. It still wrecks you.
Producing those bits does cost something, but it isn’t calories. Your brain burns about 20% of your resting metabolism around the clock. Even thinking as hard as you can lifts the working region’s blood flow by 5% or less. Whole-brain measurements during intense mental work show no change at all (Raichle & Gusnard 2002). Your hardest hour of thought and an hour of daydreaming cost almost the exact same fuel.
So why does a morning of deep work leave you wrecked? What runs out is attention. Focus is a serial spotlight that lands on one thing at a time. Every minute you spend here is a minute stolen from everything else. That opportunity cost is what you feel as effort (Kurzban 2013). An implant changes none of it. Wiring thought to the machine still forces you to think the thought. It also adds a new tax: with no muscles or vocal cords standing between your head and the wire, you have to consciously hold back the half-formed drafts and private noise you never meant to send. Both draw from the one budget you can’t refill. This step - cheap energy but expensive attention - is the shakiest link in the chain. It’s a reasoned bound, not a direct measurement.
The best brain implants already run at mouse speed
The receipts settle it. Neuralink’s cursor metric went 4.6 → 8.0 → over 10 bits/s, a measure of cursor-control speed, compared to an able-bodied-mouse baseline of ~8-10 (Neuralink 2026). The best speech brain-computer interface (BCI) hit 62 wpm, about 5 bits/s (Willett 2023). Every device lands in the ordinary typing-and-mouse band, nowhere near orders of magnitude above it. Fluent speech still outruns every BCI on Earth. Blame the sparse electrodes if you like, but even a perfect high-density wire meets the same ceiling: how fast you can consciously form one distinct intent after another. That brings us right back to the ~10 bits/s we started with.
The most advanced brain implant on Earth performs about as well as the mouse next to your keyboard. The wire reached human I/O speed. It didn’t transcend it.
Three things a brain implant fixes - none is speed
The wins are real, but they are all orthogonal to a healthy worker’s throughput. Restoration: for someone paralyzed or blind, the channel is broken rather than slow, and 5 bits/s beats zero. Latency: you get the same bits hands-free, though composing them remains the expensive part and typing is already automatic. Parallelism: you can control a few motor channels at once, even while cognition stays stubbornly serial.
The disagreement is live. Sauerbrei & Pruszynski 2025 put the brain above 10 bits/s, but that’s unconscious traffic, not the deliberate meaning you’d send down a wire. Meanwhile, Wiederhold 2025 clocked memorization at ~40 bits/s. But there, the wall was reading speed - another limit inside the head rather than the channel. Both studies push the number, but neither moves the bottleneck off the brain and onto the bottleneck of the wire. Every win works around the funnel instead of widening it.
The trick was never a wider pipe. It’s less to push through it.
There are two kinds of interface, and only one hits the ceiling. If you push more bits into the funnel of your attention - via a wider wire, a faster feed, or an always-on stream - you hit the wall, because the cspscity of the funnel is fixed.
If you take work off the funnel - using an agent that does the task or a filter that triages for you - nothing about it bounds you. You spend 10 bits saying what you want and a few more approving what comes back. The thousands of bits of work in between never touch your funnel. A wider wire moves all those bits toward you. An agent keeps them off you. That’s why agents change the work while a brain implant stays marginal. The question was never how to get more into your head faster. It’s how much you can keep out of it entirely.
So an always-on feed is the wrong move. It shoves unrequested bits back into the funnel, taxing your scarce 10 bits/s with endless vigilance. The winning interface is a filter with an off-switch. It can’t make any single decision cheaper, since that cost is fixed, but it cuts down on how many decisions reach you at all. It remains silent by default, spending your scarce attention only on the bits that matter.
That filter can absolutely be hardware - smart glasses, an earpiece, or even the implant this whole piece has been doubting. It earns its place by cutting friction, putting the agent just one glance or one thought away - summoned the moment you want it, and silent otherwise. On-demand is the win.
The wire’s been fast enough for years. The bottleneck was always the 10 bits a second of thinking capacity.
Key takeaways
- The brain moves ~10 bits/s of deliberate meaning in each direction, in and out, whatever the channel. That number sets the ceiling, not the wire.
- The best brain implants already run only at mouse speed - they match human I/O, they don’t beat it.
- With bandwidth maxed, the frontier is triage, not transport: the interface that wins is a filter you summon on demand - glasses, an earpiece, or even an implant - never an always-on feed.
Where the numbers come from
All sources accessed 2026-07-13.
- Brain output ~10 bits/s; sensory input ~10⁹ bits/s; reading/listening rates: Zheng & Meister, Neuron, 2025
- Speech signal ~39 bits/s across 17 languages: Coupé et al., Science Advances, 2019
- English entropy ~1 bit/char: Shannon, Bell System Technical Journal, 1951
- Typing median 51.6 wpm: Dhakal et al., CHI, 2018
- Brain ~20% resting metabolism, near-flat under load; local change ≤5%, no whole-brain change: Raichle & Gusnard, PNAS, 2002
- Felt effort = opportunity cost, not fuel depletion: Kurzban et al., BBS, 2013
- Speech BCI 62 wpm: Willett et al., Nature, 2023
- Neuralink cursor 4.6 → >10 BPS; mouse ~8-10 BPS: Neuralink, “Two Years of Telepathy”, 2026
- 10 bits/s is a lower bound (unconscious processing faster): Sauerbrei & Pruszynski, Nature Neuroscience, 2025
- Memorization to ~40 bits/s, reading-limited: Wiederhold et al., bioRxiv, 2025
Leave a Reply