Sabitlenmiş Tweet
first time writing an article about AI, regarding something I have been thinking about quite a lot.
open.substack.com/pub/gravity42/…
English
Chayanka_42
800 posts
Chayanka_42
@42_gravity
PhD student, Center for High Energy Physics, IISc Bangalore | Non-perturbative QFT • Tensor Networks • Cosmology | AI x Science
Katılım Haziran 2019
506 Takip Edilen884 Takipçiler
My "For You" tab has suddenly gotten markedly better.
Nikita Bier@nikitabier
We're rolling out a small tweak to boost visibility of your posts to your mutuals (people who you follow back). We noticed this data was missing from the algo and it made your friends appear less in your replies. This resulted in the reply section feeling more like a battleground with people you don't recognize. This should also help clusters form around interests more easily, which many people have asked for.
English
These blackboard style talks are really inspiring, enough details to keep people curious but not so technical that people clock out
Source: @cursor_ai YT channel
youtu.be/Z5M33oh-SAU?si…
YouTube
Ben Geist@b_geist
Spoke at @cursor_ai’s conference, compile, yesterday hosted by @dwarkesh_sp on my research how introducing information in the latent space into LLMs can make the underlying models more efficient. Haven’t written on a chalkboard in years, was a great conference and experience 😄
English
@b_geist @cursor_ai @dwarkesh_sp was a really nice talk, I hope more people work in similar areas as its going to have very real consequences in the world out there
English
Spoke at @cursor_ai’s conference, compile, yesterday hosted by @dwarkesh_sp on my research how introducing information in the latent space into LLMs can make the underlying models more efficient. Haven’t written on a chalkboard in years, was a great conference and experience 😄
English
the new ChatGPT Work on mobile feels like a superpower
the fact that it can also connect to Sites incentivizes experimentation and playing around with ideas
- Ask it to explain a complex math concept interactively using Sites
I'll share more examples later
Tibo@thsottiaux
Thank you to the 7M active users who are now using Codex and ChatGPT Work. We have added a banked reset to everyone's account to celebrate the milestone. You can apply the reset in the desktop app or on web and it will replenish the weekly usage for you. Have fun out there.
English
@nblqbl thanks for putting it up on Github, makes it so much easier to play around
English
you can download the code and do the analysis for your favorite arXiv here: github.com/nabiliqbal/arx….
(perhaps ironically i would probably not have put in the time to learn the arXiv api etc. to make these plots without claude to do the work for me)
N/N
English
As @ThomasVanRiet2 points out, there are significantly more papers being written than before, almost certainly due to LLMs.
I looked at a few of the arXiv categories, e.g. my beloved hep-th:
1/N
Thomas Van Riet@ThomasVanRiet2
28 new hep-th papers during the summer holiday period. We entered a new age.... (to all non-Europeans here: holiday is a period in which you take a break from professional activities)
English
Chayanka_42 retweetledi
@Bilzebor @modernalkemist @42_gravity That's the point, neural networks were considered a dead end and its origins are from the 1940s. Geoffrey Hutton was adviced not to pursue them and still did.
Point is you can't count string theory out completely, we still need researchers to continue
English
@ArtificialAnlys GPT 5.6 Pro not available till now in the scoring?
English
GPT-5.6 Sol (max) is the new leader in CritPt, a benchmark of unpublished research-level physics problems
CritPt, developed by Argonne and UIUC, tests models on graduate-level physics research problems contributed by 60+ researchers from 30+ institutions globally. GPT-5.6 Sol (max) gains ~5 points on its predecessor, GPT-5.5 (xhigh), beating Claude Fable 5 by ~4 points.
Congratulations @OpenAI and @sama on this result, indicating the model’s strong potential for frontier science research.
English
Can't believe @robertskmiles has been talking about AI safety on YouTube for about 9 years, going as far as 2017.
The Orthogonality Thesis video really made me think so much clearer (how often we make hidden assumptions about things without realising)
English
@garrytan Thanks for the repost Garry. I keep posting stuff like this as I increasingly encounter similar stuff in my own research work.
English
The math behind GPT Pro and why almost all the news about solving Erdos Problems were with the GPT 5.x Pro version of the models
Ig @SebastienBubeck or @ALupsasca can confirm or deny 😁
English
Biggest step change in the history of physics
Physics In History@PhysInHistory
Werner Heisenberg's canonical commutation rule for position and momentum variables of a particle, 1927.
English
@yujitach @WaylandZhang It'll be nice if you can try the same in a new chat with GPT 5.6 Sol (extra) and report back what you see. Will be a very interesting comparison.
English
@WaylandZhang I’m not doing this in any systematic way at all. I’m just trying to see the capabilities. Fable seems to be better than ChatGPT 5.5 pro mode. I haven’t tried 5.6.
English
A Youtube Series where people from different backgrounds (maybe somewhat resource limited) still extract a lot from these latest tools in unexpected ways in their work.
It's quite surprising how many people can learn and adapt quickly given the right resources.
Long form videos are better for this, maybe 30-50 mins.
English
Very few people know the amount of useful work that the current models can do in Code/Codex/etc. with the right setup
This is not a "rah rah you are so early" post, this is a "AI companies are doing a really bad job explaining what their systems actually do in a clear way" post.
English
CritPt benchmark can tell us a lot more than just the total score
I have been continuously tracking this plot for the last few months and it still blows my mind how token efficient the GPT series of models are getting over time.
Y axis: Critpt Percentage points per thousand output tokens
X axis: best model from OpenAI vs Anthropic which were released around the same time
3 sharp observations:
1) OpenAI rises from 0.148 to 0.769 CritPt percentage points per 1k output tokens, about a 5.2× improvement.
2) Anthropic rises from 0.131 to 0.253, about 1.9×, with substantial volatility.
3) In the latest release, Sol is about 3.0× more output-token-efficient than Fable, while also scoring higher: 32.3% vs 28.6%.
English
GPT TAKING A MOMENT TO REST
This COT summary from GPT 5.6 Sol really surprised me, particularly if this behaviour emerged from training
I think it's a good lesson we can all take home
ps: I was asking it to make a plot from some data that was scattered around in many screenshots
English
We are entering a completely new era of science
Here is Yuji Tachikawa from Japan (Mathematical Physics, String Theory, QFT) on recent progress in his own work using Fable 5 :
"I've been trying out Claude Fable recently, and last night, on a whim, I showed it my research notes about a collaborative project that's seen no progress in the past six months or so and asked for its thoughts. To my surprise, it made a non-trivial observation and essentially solved it."
"I was also surprised that it was using sympy to automatically write code and verify his own predictions."
"Fable probably seems like it properly understands string theory and has intuition too—that's my impression"
English
Claude is safety flagging Quantum Field Theory qs now and routing back from Fable to Opus
Completely ruining a long conversation
English
A contour integral with multiple branch cuts can be so confusing
English
GPT 5 was released almost a year ago and today we have the latest models with no reasoning, scoring the same as the best reasoning model from a year ago.
"Today's models just have a better starting point" -> Reasoning distilled in the base model weights
English