Woke up this morning to the news that Stephen Hawking died, and by the time I'd made coffee my Twitter feed was already wall-to-wall black holes and time-brief-history jokes and grainy photos of the guy at Cambridge in the 80s. Fine. Deserved, obviously. But I want to write about something a little narrower, because I think it's the part nobody outside the accessibility-tech world pays much attention to: the software that actually let him talk.
By the end, Hawking wasn't using anything exotic to communicate. He had one working muscle, a twitch in his cheek, and a sensor mounted on his glasses that picked it up. That signal drove a cursor across a keyboard on screen, scanning row by row, and he'd click to select a letter or a whole predicted word. In 2011 he was down to about one word a minute doing this. One. Try having a conversation at that speed — you'd give up on small talk entirely, and honestly I think he mostly did.
What fixed it, or at least made it livable, was a 2014 collaboration with Intel called ACAT (Assistive Context-Aware Toolkit). Intel built it with a UK company called SwiftKey doing the word-prediction layer, the same predictive text engine that was on a lot of Android keyboards at the time. Hawking's typing rate roughly doubled once the prediction got good, because the system learned his vocabulary (a lot of physics jargon, apparently, which generic dictionaries choke on) and started guessing his next word correctly a huge chunk of the time. Intel open-sourced the whole ACAT platform on GitHub back in 2015 specifically so other people with motor neuron conditions, not just famous physicists, could build their own version without paying a fortune for proprietary AAC hardware. I remember reading about that release and thinking it was one of the more quietly important open-source drops of the decade, and it got about a tenth of the coverage of, I don't know, some new JavaScript framework nobody uses two years later.
That's the bit I keep chewing on today. A man communicated an entire physics career, and a lot of jokes, through a cheek muscle and some word-prediction software, and the tooling behind that is sitting in a public repo right now for anyone to fork. Meanwhile half of tech Twitter argues about tabs versus spaces like it's a moral crusade. Priorities, I guess.
Small tangent, only sort of related: today is also Pi Day, 3/14, and it's always bugged me a little that Americans are the only ones for whom that date format even makes the joke work. Write it 14/3 like the rest of the world does and the pun just evaporates. MIT leans into it hard though; they've released their regular admissions decisions on Pi Day at 6:28pm for years now (6:28, as in tau, because somebody in admissions has a sense of humor), and every March a small subset of stressed 17-year-olds refreshes a portal instead of doing literally anything else with their evening. Weirdly Hawking dying on Einstein's birthday, which is also Pi Day, is the kind of coincidence that if you put it in a novel an editor would cut it for being too on the nose.
Anyway. If you want to actually look at the ACAT stuff instead of just reading a paragraph about it, it's still up on GitHub under IntelSoftware/context-aware-toolkit, or it was last I checked a couple years back. Worth a poke around even if you're not building assistive tech, because the eye-gaze and cursor-scanning logic in there is a genuinely clever bit of engineering for a hard, unglamorous problem — the kind of problem that doesn't get a keynote.
I don't have a tidy way to close this out. Go read A Brief History of Time again, or don't, everyone already owns a copy they haven't finished. I'm going to go stare at my own todo list, which currently has "fix contact form on the old portfolio site" sitting on it for the third week running, and feel appropriately small about my priorities compared to a guy who ran a full career off a cheek twitch and a good prediction model.