Ramon Astudillo
@ramon-astudillo.bsky.social
6.1K followers 310 following 1.4K posts
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
Posts Media Videos Starter Packs
Pinned
ramon-astudillo.bsky.social
📣 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.
ramon-astudillo.bsky.social
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
ramon-astudillo.bsky.social
The issue is, in such a scenario, there is basically no money for consumption. How can the economic system go on?. Current power concentration and mechanisms to acquire power do not give great prospects of a smooth transition to a post-scarcity society. Sooo ... they would buy land and ammunition?
ramon-astudillo.bsky.social
If you consider alignment as only related to explicit user preference feedback, it's less related. If you consider alignment as "making LLMs work as tools that do what we want then to do" then it's more related.
ramon-astudillo.bsky.social
"Genius is one percent inspiration and ninety-nine percent perspiration"

Well now is more like 40 / 60
Reposted by Ramon Astudillo
theophite.bsky.social
"the AI bubble will collapse because the technology has plateaued and cannot improve"

will_smith_spaghetti.mpeg

"we are on the verge of superintelligence"

seahorse_emoji.txt
ramon-astudillo.bsky.social
Also classifiers can *not* generate images or "steal" anything from you. "They could be using a VLM" nopes, its orders of magnitude more expensive and too slow. "One day ..." likely not, when scaling and speed is so critical and the problem is mostly contextless or context is metadata
huwupy.kawaii.social
This is one of my favorite takes. I’m sure people would love a 30-day delay on having their images posted so they can be reviewed by hand and also it’s Good Actually to make humans look at some of the most heinous images in existence all day
The Bluesky Head of Trust & Safety admitting that all posted images are processed through an Al autolabeller that presumably employees have no access to or control over? (And probably steals every image to be used in an Al "art" generator, who knows?)
Seriously, this is outrageously irresponsible.

October 7, 2025 at 10:09 AM
20 reposts
1 quote
66 likes
1 save
ramon-astudillo.bsky.social
(repost welcome) The Generative Model Alignment team at IBM Research is looking for next summer interns! Two candidates for two topics

🍰Reinforcement Learning environments for LLMs

🐎Speculative and non-auto regressive generation for LLMs

interested/curious? DM or email [email protected]
ramon-astudillo.bsky.social
It's a cat and mouse game and I guess long term ... fake team will win? ... at least for text that battle feels lost already.
ramon-astudillo.bsky.social
Joking ofc, it will take a bit longer.
ramon-astudillo.bsky.social
even if we agree on this, it will be useless because half of the world population will be 24/7 sampling pleasant realities from that thing, not even looking at the internet. Well, they will connect to get videos from other people, buy pizza, and do some data annotations (to pay for the pizza).
quantian.bsky.social
Sora is not the real problem here. The real problem is that in 12, at most 24 months, there will be a Sora clone from Tencent or Alibaba that is just as good and can be run by anybody with a 5099, at which point everyone in society needs to rapidly agree to treat all digital video as fake, period.
ramon-astudillo.bsky.social
even if we agree on this, it will be useless because half of the world population will be 24/7 sampling pleasant realities from that thing, not even looking at the internet. Well, they will connect to get videos from other people, buy pizza, and do some data annotations (to pay for the pizza).
quantian.bsky.social
Sora is not the real problem here. The real problem is that in 12, at most 24 months, there will be a Sora clone from Tencent or Alibaba that is just as good and can be run by anybody with a 5099, at which point everyone in society needs to rapidly agree to treat all digital video as fake, period.
ramon-astudillo.bsky.social
It's amazing to me that, in 2025, pythons pdb still doesn't let you configure the amount of lines of context shown and display colors ... Yes, there is ipdb but it is somehow super slow
Reposted by Ramon Astudillo
ramon-astudillo.bsky.social
Also, it's the only plausible explanation for this weather. A 27 C day again?? In October??
ramon-astudillo.bsky.social
You think I am exaggerating but look at this other one and tell me it does not look evil
Reposted by Ramon Astudillo
timkellogg.me
Reading @simonwillison.net's lobsters thread on changing the "vibecoding" tag to something more sensible, like "llms"

several people say the quiet part out loud — the *point* is to make AI enthusiasts feel uncomfortable and unwelcome. Accuracy didn't factor in

lobste.rs/s/gkzmfy/let...
Let's rename the "vibecoding" tag to "llms"
237 comments
lobste.rs
Reposted by Ramon Astudillo
seantcollins.com
CHATGPT: I understand where you're coming from. You worked really hard to get here, and now it's time to enjoy the fruit of your labors.

ISILDUR: So I should keep it? Elrond says I shouldn't

CHATGPT: The ring is precious. Sometimes friends don't have your best interests at heart.

ISILDUR: true
ramon-astudillo.bsky.social
You think I am exaggerating but look at this other one and tell me it does not look evil
ramon-astudillo.bsky.social
Artificial intelligence and crowd intelligence better grow hand in hand.
ramon-astudillo.bsky.social
there is also the "lets return to the timeline where we did not fuck up the internet as a tool for democratic empowerment and non-transactional value creation" folk. There are many of those. Where I put "internet" you can also put "AI".
ramon-astudillo.bsky.social
there is also the "lets return to the timeline where we did not fuck up the internet as a tool for democratic empowerment and non-transactional value creation" folk. There are many of those. Where I put "internet" you can also put "AI".
Reposted by Ramon Astudillo
mmitchell.bsky.social
This post about Waffles that’s going viral doesn’t credit (what I believe to be) the source, Bravo Coolee, on Twitter a few years ago. Just wanna shout out the original. Still viral thru another person’s rewrite today.
Twitter screenshot from  @BravoCoolee

“Twitter is the only place where well articulated sentences still get misinterpreted. 

You can say “I like pancakes” and somebody will say “So you hate waffles?’

No bitch. Dats a whole new sentence. Wtf is you talking about.
ramon-astudillo.bsky.social
My only gripe with Claude Code is that it should embrace better the zen of python, in particular

- Explicit is better than implicit.

- Errors should never pass silently.
ramon-astudillo.bsky.social
You should go to Portugal 🤣
ramon-astudillo.bsky.social
Also I missed the part of your first statement where you gave the specific context of your case ... otherwise I would not have asked.
ramon-astudillo.bsky.social
I recall reading this. I get it but this is a different context than what I had in mind. Also we will have to adapt ... Probably we will get better at spotting these in the future.