10 Bits Per Second

The brutal mismatch between how fast you think and how fast you can do anything about it

In 1999, Bill Gates published Business @ the Speed of Thought. The premise was electrifying: technology would soon let organizations act on information as fast as people could think about it. Build a “digital nervous system,” Gates argued, and the lag between insight and action would collapse. Information would flow like reflexes. Decisions would happen in real time.

It was a beautiful vision. The one thing Gates didn’t account for: computers have gotten millions of times faster since 1999. We haven’t.

The 10-bit bottleneck

In 2024, researchers Jieyu Zheng and Markus Meister at Caltech published a study in Neuron that put a number on human cognitive throughput. After applying information theory to decades of behavioral research — typing, reading, gaming, solving Rubik’s cubes, playing chess — they arrived at a figure: roughly 10 bits per second.

That’s the speed of you. Not your eyes, not your ears — those take in about 1 billion bits per second. But the conscious part, the part that thinks and decides and acts? Ten bits. Per second.

For perspective: your home Wi-Fi probably runs at 100 megabits per second. Your brain’s conscious output is ten million times slower. A 1959 Bell telephone modem — the kind that screeched when it connected — could transmit 100 bits per second. Your behavioral throughput wouldn’t even saturate that.

When Gates wrote about business at the speed of thought, we didn’t yet know how slow thought actually was. The speed limit wasn’t technology. It was us.

Your ports are worse than you think

The 10-bit bottleneck is your processing speed — how fast your conscious mind can make decisions and form outputs. But even getting information into and out of that processor is painfully constrained. Let’s walk through the actual hardware.

Output — getting thoughts out of your head:

Your fastest output channel is speech, at roughly 150 words per minute. Because English has about 1 bit of entropy per character, that translates to approximately 13 bits per second. Typing, the way most of us interact with computers all day, is much worse — the average person types at about 40 words per minute, roughly 3.3 bits per second. And handwriting? About 13 words per minute. Just over 1 bit per second. You are literally outputting information slower than a telegraph operator in 1860.

Input — getting the world into your head:

Reading is your fastest conscious input channel. The average adult reads at about 238 words per minute for non-fiction, with skilled readers reaching 350. That’s roughly 20–29 bits per second. Listening tops out at about 275 words per minute before comprehension collapses — which is why podcasts at 2x speed feel like the ceiling. Your senses take in a billion bits per second of raw data, but after your brain compresses, filters, and discards, you’re consciously working with maybe 10–30 bits of it.

Now compare that to the machine sitting in front of you. A basic SSD reads data at 500 megabytes per second — about 4 billion bits per second. A Thunderbolt 4 cable transfers at 40 gigabits per second. An LLM can ingest an entire novel in seconds and generate output at 100+ tokens per second. The gap between human I/O and machine I/O isn’t a percentage difference. It’s a factor of millions.

The bandwidth hierarchy

Here’s the picture, ordered from fastest to slowest human output:

Inner speech (what you think in your head) → estimated up to 4,000 WPM, though this is poorly measured and the Caltech research suggests that usable, decision-relevant throughput is still only about 10 bits/s regardless.

Speaking → ~150 WPM / ~13 bits/s

Typing → ~40 WPM / ~3.3 bits/s

Handwriting → ~13 WPM / ~1.1 bits/s

Notice the brutal compression at every stage. You think something rich and layered. By the time it reaches your mouth, most of the nuance is gone. By the time it reaches your fingertips, you’re down to a trickle. Each transition is lossy. Each port is narrower than the last.

And here’s what makes it worse: we’ve built almost our entire knowledge-work infrastructure around the slowest channel. We type emails. We type reports. We type Slack messages. We type prompts into AI chatbots. We have spent decades optimizing keyboards and screens when the port that’s three to four times wider — your voice — is sitting right there, mostly unused for productive work.

A 2016 Stanford study confirmed what the numbers already suggest: speech-to-text is three times faster than typing, with a 20% lower error rate. Speaking to a machine isn’t just faster. It’s closer to the natural pace of your thinking, which means less is lost in the compression between thought and expression.

We write like we text

This bandwidth problem shows up clearly in how we interact with AI. We have been given the most powerful text-processing tools in history, and our dominant mode of using them is the chatbot: short prompts, a few sentences at most. Data suggests the typical ChatGPT message is 200 to 500 characters — about the length of a text message. We’re not writing structured briefs or rich problem statements. We’re pecking out quick one-liners and hoping the machine fills in the gaps.

This isn’t laziness. It’s physics. Writing a detailed, well-structured prompt is hard because writing anything detailed and well-structured is hard. It requires pushing your thoughts through the narrowest output port you have — your fingers on a keyboard — and sustaining that effort long enough to produce something coherent. Most people can’t or won’t do it, for the same reason most people don’t write long-form anything: it’s cognitively expensive and we were never very good at it in the first place.

The irony is sharp. AI systems get dramatically better when you give them more context, more structure, more detail. But the human on the other end of the conversation is bandwidth-limited to roughly 3 bits per second of typing, doesn’t enjoy writing, isn’t particularly good at it, and defaults to the shortest possible message. We’ve built a Ferrari and we’re feeding it fuel through an eyedropper.

What Gates got right — and what comes next

Gates wasn’t wrong about the vision. He was just early, and he underestimated where the real bottleneck would be. The “digital nervous system” he described — information flowing instantly to wherever it’s needed — exists now. We have it. The internet, cloud computing, APIs, and AI have made information flow essentially frictionless between machines.

The bottleneck that remains is biological. It’s the 10-bit-per-second human sitting in the middle of a gigabit world, trying to get thoughts out through fingers that move at 3 bits per second, relying on a memory system that hallucinates, and communicating in a format — natural language — that is ambiguous by design.

The next wave of productivity won’t come from making machines faster. They’re already fast enough. It will come from widening the pipe between the human and the machine. That means voice-first interfaces, because speaking is your widest output port. It means AI that can interpret messy, fragmented human input and ask clarifying questions — because our output will never be clean. It means tools that capture what you’re thinking before the lossy compression of writing it down — through conversation, through ambient recording, through interfaces we haven’t built yet.

And until that future arrives, the single most valuable thing you can do is also the simplest: open your mouth. Talk to your tools. Dictate your notes. Speak your prompts. You’ll get three times more information out of your head per second, with fewer errors, and with less cognitive strain.

Your brain is a 10-bit processor trapped behind 3-bit ports. Use the widest one you’ve got.

Sources and references

The 10 bits/s finding: Zheng, J. & Meister, M. (2024). “The Unbearable Slowness of Being: Why do we live at 10 bits/s?” Neuron. https://arxiv.org/html/2408.10234v2

Caltech press coverage: https://neurosciencenews.com/brain-thinking-sensory-speed-28292/

Paradromics analysis of the paper (BCI implications): https://www.paradromics.com/blog/the-speed-of-human-thought-and-its-implications-for-brain-computer-interfaces

Speech-to-text 3x faster than typing (Stanford, 2016): https://news.stanford.edu/stories/2016/08/stanford-study-speech-recognition-faster-texting

Average reading speed meta-analysis: Brysbaert, M. (2019). “How many words do we read per minute?” Journal of Memory and Language. https://www.sciencedirect.com/science/article/abs/pii/S0749596X19300786

Listening comprehension ceiling (~275 WPM): Lund, R.E. (2021). “Reading rate and most efficient listening rate are highly similar.” Psychonomic Bulletin & Review. https://pubmed.ncbi.nlm.nih.gov/34516216/

Average handwriting speed: https://tctecinnovation.com/blogs/daily-blog/comparing-handwriting-speeds-among-different-age-groups

Shannon entropy of English (~1 bit/character): Shannon, C.E. (1951). “Prediction and Entropy of Printed English.” Bell System Technical Journal.

Bill Gates — Business @ the Speed of Thought (1999): https://www.goodreads.com/book/show/41617

AI cognitive overload (“AI brain fry”): Bedard, J. et al. (2026). “AI brain fry: Managing cognitive overload in the age of artificial intelligence.” Harvard Business Review, covered in Psychology Today. https://publichealth.gmu.edu/news/2026-03/ai-and-rise-cognitive-overload

ChatGPT typical message length: https://www.reddit.com/r/ChatGPT/comments/1o9ghsd/howactiveare_you/