Ramon Astudillo
@ramon-astudillo.bsky.social
Principal Research Scientist at IBM Research AI in New York. Speech, Formal/Natural Language Processing. Currently LLM post-training, structured SDG and RL. Opinions my own and non stationary.
ramon.astudillo.com
ramon.astudillo.com
Pinned
I did a starter pack of people in New York (City) working on ML/AI. Please distribute and feel free to self nominate!
go.bsky.app/BoEtagz
go.bsky.app/BoEtagz
📣 I am sure we have reached only a small fraction of New York's ML community in bsky. Please repost 🔁 this if you think you may have interested people close to you in the social graph.
Reposted by Ramon Astudillo
Unlike ICLR reviews, the sunrise never disappoints
November 12, 2025 at 6:36 AM
Unlike ICLR reviews, the sunrise never disappoints
Reposted by Ramon Astudillo
Happy There Is A Circulatory System Walking Through The Kitchen Day to all who celebrate.
November 10, 2025 at 10:54 AM
Happy There Is A Circulatory System Walking Through The Kitchen Day to all who celebrate.
Reposted by Ramon Astudillo
Three dumb ideas that recur constantly in discussions among nerds about intelligence:
(1) IQ tests are illegal in hiring.
(2) There exist national average IQ numbers.
(3) Intelligence research is suppressed.
(1) IQ tests are illegal in hiring.
(2) There exist national average IQ numbers.
(3) Intelligence research is suppressed.
November 9, 2025 at 11:24 PM
Three dumb ideas that recur constantly in discussions among nerds about intelligence:
(1) IQ tests are illegal in hiring.
(2) There exist national average IQ numbers.
(3) Intelligence research is suppressed.
(1) IQ tests are illegal in hiring.
(2) There exist national average IQ numbers.
(3) Intelligence research is suppressed.
For anyone that has worked in DL long enough to go from DIY to theano/chainer/tensorflow/mxnet/dynet/pytorch/Jax, one of these "winning" over all other felt unlikely ... and it kinda happened
November 10, 2025 at 3:47 AM
For anyone that has worked in DL long enough to go from DIY to theano/chainer/tensorflow/mxnet/dynet/pytorch/Jax, one of these "winning" over all other felt unlikely ... and it kinda happened
For machines we call it "Reward Hacking" for humans we call it "Goodhart's Law". It makes us look less stupid (only look).
social media is RL on humans
It's interesting that RLHF'd LLMs and influencers talk the same way. Perhaps through the evolution of clickbait, we'd already found the local maximum of attention grabbing
November 10, 2025 at 1:03 AM
For machines we call it "Reward Hacking" for humans we call it "Goodhart's Law". It makes us look less stupid (only look).
It's not going away folks ... ok, there are some scenarios, but on those something else went really wrong and we have far serious problems to think about
Part of the reason why I’m so insistent about folks understanding AI capabilities is that they’re here to stay and we need to start thinking about what to do in such a world. Putting the genie back in the bottle is a pleasant fantasy that delays serious reckoning
November 9, 2025 at 1:28 PM
It's not going away folks ... ok, there are some scenarios, but on those something else went really wrong and we have far serious problems to think about
Reposted by Ramon Astudillo
Another shot of the xpeng robot. really surreal stuff.
November 8, 2025 at 11:47 PM
Another shot of the xpeng robot. really surreal stuff.
Claude Code's biggest flaw is that the cursor on the cl does not blink! When jumping around in tmux I often end up inputting things into CC that were meant for some other pane/window. Generally exits, which Claude understands in many languages ...
November 7, 2025 at 9:14 PM
Claude Code's biggest flaw is that the cursor on the cl does not blink! When jumping around in tmux I often end up inputting things into CC that were meant for some other pane/window. Generally exits, which Claude understands in many languages ...
Guessing which keywords can I pronounce properly in every language to get Google's automatic language detection ASR to pick the damn right language. For Portuguese, "Oi" seems to do the trick.
November 7, 2025 at 7:44 PM
Guessing which keywords can I pronounce properly in every language to get Google's automatic language detection ASR to pick the damn right language. For Portuguese, "Oi" seems to do the trick.
Ok, time to confess. I do not gamble, so the second time I re-joined Twitter I did it with the express intent of using it as a prediction market so that my tendency to make claims bore some risk/reward. This one aged well.
They are actually closest to a general purpose computer! See e.g. @karpathy.bsky.social 's Software 3.0 view. IBM defines this as "Generative Computer", I prefer "Neural Computer". The idea is basically what I thought would be the GPT4 paper's title (my title was less boring)
x.com/RamonAstudil...
x.com/RamonAstudil...
November 7, 2025 at 5:12 PM
Ok, time to confess. I do not gamble, so the second time I re-joined Twitter I did it with the express intent of using it as a prediction market so that my tendency to make claims bore some risk/reward. This one aged well.
Is there something fundamentally wrong with this reasoning?
Here is an argument why the place in the supply chain with the biggest value capture potential for the upcoming AI industrial revolution is neither in LLM training, nor infrastructure. Yeah, space is crowded, requires gigantic investments that deprecate fast , but also 👇
November 7, 2025 at 3:31 PM
Is there something fundamentally wrong with this reasoning?
Here is an argument why the place in the supply chain with the biggest value capture potential for the upcoming AI industrial revolution is neither in LLM training, nor infrastructure. Yeah, space is crowded, requires gigantic investments that deprecate fast , but also 👇
November 7, 2025 at 3:20 PM
Here is an argument why the place in the supply chain with the biggest value capture potential for the upcoming AI industrial revolution is neither in LLM training, nor infrastructure. Yeah, space is crowded, requires gigantic investments that deprecate fast , but also 👇
Reposted by Ramon Astudillo
apparently Logan is on bsky now?
Introducing the File Search Tool in the Gemini API, our hosted RAG solution with free storage and free query time embeddings 💾
We are super excited about this new approach and think it will dramatically simplify the path to context aware AI systems, more details in 🧵
We are super excited about this new approach and think it will dramatically simplify the path to context aware AI systems, more details in 🧵
November 6, 2025 at 11:33 PM
apparently Logan is on bsky now?
Observational tokenonomy
I've been studying the Qwen 3 4B Instruct 2507 token unembedding matrix, ɣ. I can't entirely remember why. I'm pretty deep into it now.
The ɣ matrix maps token IDs to embeddings — vectors in 2,560 dimensions.
I imagine these vectors as stars in the sky. I've been doing observational tokenonomy.
The ɣ matrix maps token IDs to embeddings — vectors in 2,560 dimensions.
I imagine these vectors as stars in the sky. I've been doing observational tokenonomy.
November 6, 2025 at 12:42 PM
Observational tokenonomy
Reposted by Ramon Astudillo
This is a collab with DeepMind & Terence Tao btw
November 6, 2025 at 10:32 AM
This is a collab with DeepMind & Terence Tao btw
Reposted by Ramon Astudillo
CYBERSPACE 2
for(float i,z,d;z+i++<7e1;o+=vec4(z,1,9,1)/d)
{vec3 p=abs(z*normalize(FC.rgb*2.-r.xyy));p.z+=t*5.;p+=sin(p+p);for(d=0.;d++<9.;p+=.4*cos(round(.2*d*p)+.2*t).zxy);z+=d=.1*sqrt(length(p.xyy*p.yxy));}
o=tanh(o/7e3);
for(float i,z,d;z+i++<7e1;o+=vec4(z,1,9,1)/d)
{vec3 p=abs(z*normalize(FC.rgb*2.-r.xyy));p.z+=t*5.;p+=sin(p+p);for(d=0.;d++<9.;p+=.4*cos(round(.2*d*p)+.2*t).zxy);z+=d=.1*sqrt(length(p.xyy*p.yxy));}
o=tanh(o/7e3);
November 5, 2025 at 9:21 PM
CYBERSPACE 2
for(float i,z,d;z+i++<7e1;o+=vec4(z,1,9,1)/d)
{vec3 p=abs(z*normalize(FC.rgb*2.-r.xyy));p.z+=t*5.;p+=sin(p+p);for(d=0.;d++<9.;p+=.4*cos(round(.2*d*p)+.2*t).zxy);z+=d=.1*sqrt(length(p.xyy*p.yxy));}
o=tanh(o/7e3);
for(float i,z,d;z+i++<7e1;o+=vec4(z,1,9,1)/d)
{vec3 p=abs(z*normalize(FC.rgb*2.-r.xyy));p.z+=t*5.;p+=sin(p+p);for(d=0.;d++<9.;p+=.4*cos(round(.2*d*p)+.2*t).zxy);z+=d=.1*sqrt(length(p.xyy*p.yxy));}
o=tanh(o/7e3);
This is all the same guy??
Chris Lattner is one of the most influential engineers of the past two decades. He created LLVM, Swift, contributed to TensorFlow, and created the Mojo language.
What was the story about creating Swift - and why did he face resistance inside Apple when wanting to replace Objective C?
(cont'd)
What was the story about creating Swift - and why did he face resistance inside Apple when wanting to replace Objective C?
(cont'd)
November 6, 2025 at 12:53 AM
This is all the same guy??
Reposted by Ramon Astudillo
Chris Lattner is one of the most influential engineers of the past two decades. He created LLVM, Swift, contributed to TensorFlow, and created the Mojo language.
What was the story about creating Swift - and why did he face resistance inside Apple when wanting to replace Objective C?
(cont'd)
What was the story about creating Swift - and why did he face resistance inside Apple when wanting to replace Objective C?
(cont'd)
November 5, 2025 at 9:23 PM
Chris Lattner is one of the most influential engineers of the past two decades. He created LLVM, Swift, contributed to TensorFlow, and created the Mojo language.
What was the story about creating Swift - and why did he face resistance inside Apple when wanting to replace Objective C?
(cont'd)
What was the story about creating Swift - and why did he face resistance inside Apple when wanting to replace Objective C?
(cont'd)
Too big too frail
November 6, 2025 at 12:47 AM
Too big too frail
So something out there seems to have blown up at scale so massive we don't even have a good explanation for? Like ... star exploding while devoured by black hole falls short, ok ...
I wrote this article! 🙂
This is IMO one of the biggest discoveries of the year, but you haven’t heard about it bc of a NASA press release/embargo stuck bc of the shutdown. I however have no such constraint (we’re tracking it in radio but haven’t finished the paper) so got to write about it! 🔭🧪
This is IMO one of the biggest discoveries of the year, but you haven’t heard about it bc of a NASA press release/embargo stuck bc of the shutdown. I however have no such constraint (we’re tracking it in radio but haven’t finished the paper) so got to write about it! 🔭🧪
A cosmic explosion known as GRB 250702B is by far the longest gamma-ray burst astronomers have ever seen—if it’s even one at all
November 5, 2025 at 5:01 PM
So something out there seems to have blown up at scale so massive we don't even have a good explanation for? Like ... star exploding while devoured by black hole falls short, ok ...
Reposted by Ramon Astudillo
Every image this account posts would have sold me a paperback book in the 1980s. I would have been unable to resist these covers.
Lyra's Song does not exist #AIArt #midjourney
November 4, 2025 at 5:29 AM
Every image this account posts would have sold me a paperback book in the 1980s. I would have been unable to resist these covers.
Reposted by Ramon Astudillo
It's going to be a difficult few days for racists. No sooner do we discover the attacker was a British man called Anthony Williams, we then discover the rail worker hero who almost gave his life to save others is called Samir Zitouni.
We can't wait to hear the hot takes on whether he is British...
We can't wait to hear the hot takes on whether he is British...
November 4, 2025 at 11:09 AM
It's going to be a difficult few days for racists. No sooner do we discover the attacker was a British man called Anthony Williams, we then discover the rail worker hero who almost gave his life to save others is called Samir Zitouni.
We can't wait to hear the hot takes on whether he is British...
We can't wait to hear the hot takes on whether he is British...
Reposted by Ramon Astudillo
Sort of starting to believe that we really do need academic metrics that punish publishing too much
November 3, 2025 at 10:38 PM
Sort of starting to believe that we really do need academic metrics that punish publishing too much
I honestly cannot understand why modern ASR engines do not exploit the "burstiness" of language. Okay you failed the first time I pronounced poorly that foreign name or weird term ... but now you know I'm most likely going to use it again!
November 3, 2025 at 3:44 PM
I honestly cannot understand why modern ASR engines do not exploit the "burstiness" of language. Okay you failed the first time I pronounced poorly that foreign name or weird term ... but now you know I'm most likely going to use it again!
i.e. it's kinda horrible, but one should look at unemployment not only in terms of "number of people" but also in terms of "number of dollars" that are not in consumers hands to keep the service sector and other parts of the consumption economy alive.
There is a good anti-AI case to be made but it's mostly none of these. It does work and is much better than anything we had. Yes there is hype, but now that it is productized people can test by themselves (it's free!) and the standard startup hype is endemic to whatever the hot topic is 👇
I guess my question remains… to what end? You’re asking a robot a question you could’ve googled years ago and it answers more confidently in LinkedIn voice. It codes faster than a person, but worse, and what is it coding? Is it feeding people? Is it helping material realities? Just… why?
November 3, 2025 at 1:28 PM
i.e. it's kinda horrible, but one should look at unemployment not only in terms of "number of people" but also in terms of "number of dollars" that are not in consumers hands to keep the service sector and other parts of the consumption economy alive.