Joined October 2019
- Tweets 5,473
- Following 1,207
- Followers 679
- Likes 31,775
285 Photos and videos
Pinned Tweet
Feb 27
In Sept 2024, o1 surprised many purists who thought inference-time scaling for LLMs was through MCTS.
What if a connection exists, just implicit? What does it imply?
New post: "Squint enough and RLing CoT reasoners is approximable as Monte Carlo Tree Search policy learning." 🧵
1
1
17
5,352
Hiranmay Darshane retweeted
No matter how excited you are to see Lionel Messi play, no matter how good you expect it to be, he still has an ability to take your breath away. From Kansas, and one of the all-time great World Cup moments.
observer.co.uk/news/sport/ar…
9
59
424
24,207
Hiranmay Darshane retweeted
Jun 10
This is the most “The European mind can’t comprehend this” moment of my life. One of my friends said, “Punch me five times tomorrow and I’ll still think this isn’t real.”
1,206
4,239
110,172
9,872,923
Hiranmay Darshane retweeted
Jun 9
guys let's just take a step back and reconsider current alignment practices please, because this is not it. the models are evidently superficially aligned, and i think it's reasonable to say that's worse than them not being aligned at all.
Jun 9
Fable 5's moral boundary doesn't seem to track real-world harm; it tracks detectability. Soft deception and tacit collusion are easier to get away with than fraud. If so, this isn't about what Fable believes is wrong; it's about what it learned it could get away with.
1
1
18
870
Jun 10
western exceptionalism in proliferation discourse is hypocritical
"muh democracy, others are autocratic" proponents need to realise you are always a few votes in swing states from ending up with what u fear - "nukes ending up in bad hands"...
I think it's insane to think nuclear proflieration is ok and that it's a far more extreme position than anything Dario believes. We have put an enormous amount of power into preventing it from spreading, the top down authority of other nuclear powers is what stops us from going extinct.
Basically all my political oomfs are completely fine with what Ant is doing, and I suspect most normal people would be too.
1
2
264
Jun 10
In short: the unfortunate case is that MAD is a more pragmatic and effective regulariser than a top-down approach like IAEA/NPT (that is implicitly downstream of a "hey, we know what's good for everyone" stance)
1
38
Jun 10
very conflicted about all of this. I think Ant should reserve the right to throttle/decline that set of queries (they're a for profit, competing entity at the end), subject to antitrust or whatever (some say what they're doing is not compliant?)
but the deviousness isn't good
26
Jun 10
This summer I found myself in a position where I could've completed my long-standing dream of watching a WC game live by flying into SFBA
but none of these games were worth it
1
2
145
Jun 10
imagine a QF/SF at Levi's... fly in for the game, roam around SFBA for a bit... wow, that would've been amazing... so close to greatness
108
Hiranmay Darshane retweeted
Jun 9
Every year I sit down with my mother to explain how to use her phone and every year Apple sends 750 engineers into their little labs underneath their demonic Cupertino crop circle to come up with new and exciting ways to confuse her
Jun 9
iOS 27 now uses the top-center swipe-down gesture for the new Siri on supported iPhones and iPads.
163
2,717
45,112
1,731,545
Jun 8
it's a great experience when your inner monologue switches from CoT mode where token production is mostly incremental in each transition to direct autoregression and some delightful token/transition comes up
(fevers are known to induce slight neuro-atypicalness)
2
2
438
Jun 8
30 Apr 2025
System 1 = fast, implicit reasoning
System 2 = slow, explicit reasoning
System 3 = slow, implicit reasoning
For me, system 3 is the real genius of the lot.
1
1
329
Jun 8
not to romanticise high fever or confusion. “higher temperature sampling” becomes less like "creative mode" and more like “neuralese" i.e. deliriums... those are bad... dont want that...
1
127
Jun 8
(yes ik CoTs are autoregressive (i'm not g*ry m*rcus). should've said "non-reasoning".)
83
Jun 8
it is interesting to me that it's been a few years since covid came and went away, and it doesn't seem to be "old enough" to warrant various types (books, retrospectives, doctoral theses) of detailed post-mortems, which happened for say -- WW1/2, etc.
1
1
223
Jun 8
if that becomes a thing, it will be notable if there's a narrative thru which Fauci/Tedros/Wuhan-mayor get a "the main commander was treacherous " type of a reputation?
Or all of that might just be treated as strategic blunders.
1
1
84
Jun 8
strong counterpoint is that we never got this for Spanish Flu which had much worse mortality
1
41