Newsletter: Based on estimates of their burn rate and historic analyses, I hypothesize that OpenAI will collapse in the next 12-24 months unless it raises more funding than in the history of the valley and creates an entirely new form of AI.
https://t.co/ImtkIbgl3s
Trump’s “act like a blustering jerk” doctrine has successfully persuaded most of the world to prefer China to the United States of America.
https://t.co/Cff9omxwxy
Neither here nor there but the Marshallese flag that flashes when the automated captions include a reference to a New York City Marshal, the enforcement officers, is a pretty funny failure mode 🇲🇭
20 months no rent paid.
7 orders to show cause.
Apartment vacated and completely destroyed.
An HPD inspector would issue a dozen violations or more to the property owner for the damage caused by the tenant in this unit which the owner was legally barred from entering while the renter destroyed it.
Must be capitalism’s fault right?
NoMa McDonald’s for breakfast, one of the famous DC museums in the morning (Spy Museum), lunch from a food truck on the mall before seeing a classic DC monument in the afternoon (Eisenhower Memorial), followed by a walk through historic Chinatown. Dinner at Ben’s Chili Bowl
Star Axis is complete. Fifty years in the making, land artist Charles Ross has finished his monumental observatory on a mesa in New Mexico. A staircase, aligned with Earth's axis, climbs to a open lens that tracks Polaris and the 26,000-year cycle of Earth's gravitational wobble.
Read the beautiful feature in the NYT: https://t.co/vXco4bZcNB
Substack has one of the highest concentrations of skilled human writers on the internet, and we think that's something special.
Starting today, users can request a Pangram check on any Substack post over 100 words directly on the platform.
🏆 S C O R I G A M I 🏆
🇫🇷 France 4–6 England 🏴
3rd Place Match
THIS SCORELINE HAS NEVER HAPPENED IN 96 YEARS AND 1,067 MATCHES OF WORLD CUP HISTORY.
ADD IT TO THE BOARD.
34th unique score. Only in the World Cup.
https://t.co/zaGnN238gG
#FIFAWorldCup#FRAENG#LesBleus #ThreeLions
People often evaluate Pangram's false positive rate using pre-ChatGPT text. This is where our 1 in 10,000 FPR number comes from. But we often face the question:
"How can you prove that Pangram isn't just memorizing pre-2022 text that's already in its training set?"
To some degree, this is a fair criticism. If you're checking Moby Dick, any pretrained base will have seen this text, and therefore could have memorized that this text is human-written.
I talked to someone who wanted to test this hypothesis by testing Pangram on tokens that didn't exist on the internet. Super interesting. Of course we want to see the results. So I gave them credits, they found a bunch of old books that didn't exist anywhere, and scanned them. Almost three million words in total.
What they found was actually quite interesting! Three AI-assisted segments and one AI-generated segment. But when they looked at their data, they found that it the segments Pangram flagged were actually hallucinated by Mistral OCR. Mistral OCR had hallucinated entire tables onto blank pages, and these were the only segments that Pangram flagged as AI.
No false positives in about three million words / 8,000 segments. Pretty cool!
I made a visual viewer for HTTP range requests when exploring a Cloud Optimized Geotif from a remote server
As you pan around, it shows you exactly which tiles are being pulled and exactly where the bytes are stored
We couldn’t find a usable set of boundaries for 10,000+ global cities, so we (@devdatalab) developed our own and made them open for research.
We were kind of shocked this didn’t exist already. A thread 1/