I now have a French-speaking account, where I will rant about #AI and various topics, avoiding English references as much as possible. I'll cross-post important things. Thanks for your attention!
Re-bonjour Twitter ! Il y a quelques mois, j'ai posté un fil en anglais à propos de l'importance de la communication scientifique en français, dans un domaine (pour moi l'#IA) où l'anglais domine. J'ouvre donc ce compte francophone. https://t.co/aDlXUrHUNI
I had the Metaculus (weakly or not) AGI timelines in the extras, they have lengthened since early 2025, right after the reasoning models boom, with a wider distribution.
The resolution will be a mess. I'll swap the charts for more relevant ones.
2025-02-25: giving a "What's going on in AI?" presentation at the local EA group.
2026-07-07: giving a "What's going on in AI, again?" presentation at the same local EA group.
Counting how many slides I'm updating.
All the non-cosmetic changes are in the supplementary materials, the various curves, the latest numbers... but I stand by my basic landscape description.
Is this a mark of good foresight, lack of precision? My hot takes were never written down! I have no receipts!
My face has been on the Internet for a while now, so I'm putting it on my profile. Adding a banner as well.
Posting more will be the harder task, but if I can apply the same "do the obvious fix" attitude, maybe...
I joined this website in 2010. I was 18 and did not want to post pictures of myself. Instead, for my profile I tried to be clever, and generated some text on a blue field, which was funny for at least 0.7 seconds and unremarkable.
Here it is:
"jrmyp" has ever been my handle in two places: GitHub, and my personal website (https://t.co/HGvd3pDs0P). In most other places (Wikimedia projects being the biggest exception), I'm Gyrodiot.
This has not made sense for a very long time. It's time to fix it.
@ja3k_ For context: ELIZA was an early (1960s!) conversational interface that sidestepped the problems of understanding what the user says by... asking questions parroting back what the user mentioned, through deterministic patterns.
This was enough for some users to engage!
@sebkrier There was a four-years-later from Drexler himself (https://t.co/gYWQ11nbOP), and his newsletter continues the work... should pieces be written again to comment it?
@BjarturTomas - Hit him with an extraordinary amount of evidence!
- We tried, his priors are too strong!
- But how can we make him update?
- We can't. Our only hope is for him not to make predictions. His beliefs don't pay rent, yet he might as well be the landlord by now.
@TetraspaceWest Past the intensity of red, past the unearthly glow of purple, past the white-hot that wasn't meant for air, comes the black of charred elements.
@FeltInTheHeart@Aella_Girl@ramez Because the test is a proxy for some properties you actually want but cannot measure easily, or where failure is too costly, e.g. "does not cause mass death".
Some alignment work is pessimistic on our ability to devise good enough proxies (keyword: Goodhart's Curse).
@NathanpmYoung You can't tell from the score but that Spain v. Cape Verde was one of the most engaging 0-0 matches I've seen in my life, especially in contrast to Germany v. Curaçao.
@liron A blunt instrument has been used. It does run counter to "they wouldn't dare constrain their local champions" arguments. Now frontier companies will have to do some effort to avoid future hits, now the instrument has been used once.
How much effort?