JOKE SCORING IN FIVE MINUTES WITH DEEPSEEK-R1
Since Deepseek has released distilled Qwen models under a generous MIT license, it’s simpler than you might expect to create a quick assistant for small reasoning tasks on your own hardware. I could see a researcher spinning up an instance of this to sort through and classify documents, for example, without worrying about uploading them to a 3rd party or having an AI vendor monitor your activity.
DEEPSEEK-R1: IMPLICATIONS
The following prompt was supplied to a distilled instance of Deepseek-R1:32B, running on my desktop machine with a single GPU (an NVIDIA RTX A4500 with 20GB of VRAM). Its response (including the <think> prelude) follows. –jim
NOMIC: PLAYING SELF-MODIFYING GAMES WITH LLMS
In the summer of 2024, preparing to spend time at the Berkman Klein Center as a new fellow, I spent some time thinking about AI and law. I started doing research for a piece (never finished) about encouraging ChatGPT to play Nomic, a classic pen-and-paper game that explores the underpinnings of democracy and law made famous by Douglas Hofstatder’s Scientific American column in 1982.
INTERNET HISTORY: NEXT STEPS
Thanks to everyone who responded to my earlier post about the need to preserve and curate the history of Internet measurement for future historians.
THINKING ABOUT INTERNET HISTORY
I could have happily been a librarian.
I spent the glorious summer of 1988 as an intern in my university’s library system, learning what it was like to work in the various departments. I got to sample it all: the patient craft of the book restorers, the voices of the dead in Manuscripts and Archives, the exotic tastes of the special collections managers, the ever-present fear of halon suffocation in the sealed glass cages of the rare book collection.
INFERENCE AND THE UNRELIABLE NARRATOR
When you said ‘AI’ in the 1990s, you meant symbolic AI .. and decades of work and many millions of dollars went into the construction of the ontological foundations atop which edifices of symbolic reasoning would one day be built. There were elaborate projects like Cyc, which grew to include millions of pieces of hand-curated information about the world, and ‘heuristic level modules’ to support multilevel inference:
CHATGPT ON INTERNET MEASUREMENT
An unedited dialogue with ChatGPT.
<p>Please generate a compelling and exciting first paragraph for a blog about Internet measurement, and describe the specific measurements that the blog is based on.</p>
INTERNET INFRASTRUCTURE PREDICTIONS FOR 2023 (PART 2)
In part 1, I gave two predictions: that the Ukrainian Internet would prove resilient as Russia’s attacks on infrastructure continue, and that Central Asian connectivity could become more diverse in response. In this post, I’ll look at how conflict could affect the broader Internet, and predict how that may change the conversation around Internet infrastructure centralization.
INTERNET INFRASTRUCTURE PREDICTIONS FOR 2023 (PART 1)
Looking back, I suspect 2022 will be remembered as a year in which conflicts and crises shaped our collective attitudes toward infrastructure and the Internet in interesting ways.
NEW COLLECTION, OLD PRESENTATIONS
I finally found a quiet afternoon to pull together a set of my old presentations from the Renesys and Dyn Research days, and reconstruct the timeline. Thank goodness for the Internet Archive’s Wayback Machine. I’ve also embedded videos when they exist, so if you start at the beginning, you can watch me gain weight as the years pass.