Jabberwacky

» see jabberwacky.com as it lived, on the wayback machine

~ the goodbye ~

by the time i was old enough to have gone looking for a chatbot for fun, cleverbot had already taken over from its ancestor. i knew jabberwacky as a name that appeared next to eliza and racter in the little museum of "chatbots you can talk to" on early 2000s ai curiosity pages. i never actually went to jabberwacky.com and typed at it.

rollo carpenter is a british ai developer who started the project as a hobby in 1982, hard-coded onto a sinclair zx81. that early version was not learning at all: it was hand-scripted response patterns dressed up as conversation. in 1988 carpenter restarted the project under the name "thoughts", this time with an actual learning loop that used every human input to build a growing response database. the web launch in 1997 opened the training corpus to anyone who wanted to chat with it.

the technique was contextual pattern matching. it stored every conversation and, on receipt of a new input, searched its history for the closest previous conversation and replied with what a real person had said in the analogous spot. no natural language understanding in the modern sense and no transformer weights, just enormous chat logs and a lot of fuzzy retrieval. what came out was uncanny in short bursts and incoherent over more than four turns. that was the whole shape of the thing.

cleverbot launched in october 2008 with the same basic mechanism but a cleaner interface and a browser url that spread everywhere. jabberwacky was left running, drawing occasional traffic and the odd student project. the domain has stayed up in a semi-active state through the 2010s and into the 2020s, but development effectively stopped around 2011. the lineage ran through cleverbot from then on.

~ the chronicle ~

~ the loebner years ~

the loebner prize was an annual imitation-game contest with cash prizes for the bot that judges found most human. jabberwacky won it in 2005 with a character version called "george", and again in 2006 with "joan". back-to-back wins put carpenter's approach at the center of what public-facing conversational ai looked like in the mid-2000s. that was a small center, but a real one.

the wins mattered because they proved the retrieval-over-huge-corpus approach could hold a short conversation well enough to fool judges under time pressure. this was almost fifteen years before large language models made the same trick work at a much larger scale with much better math underneath.

~ how it worked ~

the whole architecture was a big pile of past conversations plus a similarity search. when you typed something at jabberwacky, it looked through years of prior chats for the closest match to your input, then replied with what someone else had said in reply to that match. this meant the bot's voice was actually the average voice of everyone who had ever chatted with it. what you got back was strange and often mildly rude, sometimes very funny.

the failure mode was that context past a few messages fell apart. jabberwacky did not remember the conversation you were in. it remembered the general shape of conversations it had seen before. two exchanges in, you were talking to a different person's chat history than the one you were mid-thread with.

~ the cleverbot inheritance ~

cleverbot took the same approach with more logs behind it, a cleaner web ui, and a friendlier url. within a year of cleverbot's 2008 launch, most of the "you have to try this chatbot" links people were emailing pointed to cleverbot.com rather than jabberwacky.com. the parent stayed up. it just stopped growing.

the cleverbot brand later spawned iphone apps and a paid api for developers. it even had a documentary appearance in the 2011 short film "do you love me". jabberwacky was mentioned in the credits and in ai timelines. it never appeared in anyone's app store.

~ what it left behind ~

the retrieval-based chatbot lineage that runs from jabberwacky through cleverbot influenced everything from smarterchild on aim to microsoft tay a decade later. the idea that a bot should learn from its conversations was carpenter's early bet, and it turned out to be an interesting idea with a hard failure mode: the bot only ever gets as coherent, or as offensive, as the population that talks to it.

modern transformer-based chatbots do not work anything like this. they are trained once on a huge frozen corpus and then produce statistical continuations. the jabberwacky family was the last flowering of the older lineage before that shift. reading the transcripts now feels like reading a species that did not make it out of the mesozoic.

~ the domain today ~

jabberwacky.com still resolves as of 2026. the site is a semi-static page maintained by rollo carpenter's icogno company, alongside the cleverbot properties. talking to it in the browser still works most days. the underlying logs and matching engine are the ones that were running in the late 2000s, which is a strange kind of digital fossil to have kept alive.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »