Marco Zocca
@ocramz.bsky.social
ML, λ • language and the machines that understand it • https://ocramz.github.io
Pinned
Marco Zocca
@ocramz.bsky.social
· Jul 4
CERN for frontier AI >>>
ngl doing review cycles with an AI agent while on the go is nothing short of magic
November 13, 2025 at 1:47 AM
ngl doing review cycles with an AI agent while on the go is nothing short of magic
love to "as per my previous email" with copilot
November 13, 2025 at 1:40 AM
love to "as per my previous email" with copilot
Reposted by Marco Zocca
We wrote the Strain on scientific publishing to highlight the problems of time & trust. With a fantastic group of co-authors, we present The Drain of Scientific Publishing:
a 🧵 1/n
Drain: arxiv.org/abs/2511.04820
Strain: direct.mit.edu/qss/article/...
Oligopoly: direct.mit.edu/qss/article/...
a 🧵 1/n
Drain: arxiv.org/abs/2511.04820
Strain: direct.mit.edu/qss/article/...
Oligopoly: direct.mit.edu/qss/article/...
November 11, 2025 at 11:52 AM
We wrote the Strain on scientific publishing to highlight the problems of time & trust. With a fantastic group of co-authors, we present The Drain of Scientific Publishing:
a 🧵 1/n
Drain: arxiv.org/abs/2511.04820
Strain: direct.mit.edu/qss/article/...
Oligopoly: direct.mit.edu/qss/article/...
a 🧵 1/n
Drain: arxiv.org/abs/2511.04820
Strain: direct.mit.edu/qss/article/...
Oligopoly: direct.mit.edu/qss/article/...
not meant as a dunk, but this is a purely artificial "emergency". Especially for ARR which has submission cycles.
November 12, 2025 at 11:08 AM
not meant as a dunk, but this is a purely artificial "emergency". Especially for ARR which has submission cycles.
the entire information infrastructure of the EU can be turned off by the US and they are c++pping themselves over Chinese buses
OK, so now the UK, Norway, Australia and Denmark are all investigating buses made by China's Yutong Group for security risks.
www.ft.com/content/07ec...
www.ft.com/content/07ec...
UK investigates whether buses made in China can be turned off from afar
Investigation comes after Norway found Yutong vehicles could be ‘stopped or rendered inoperable’ by Chinese company
www.ft.com
November 10, 2025 at 11:33 AM
the entire information infrastructure of the EU can be turned off by the US and they are c++pping themselves over Chinese buses
a panoramic collage of waves rolling in trades time for space
November 10, 2025 at 8:37 AM
a panoramic collage of waves rolling in trades time for space
trying to get typed output from LLMs is about as fun as chewing tinfoil
nonetheless, Qwen2.5 is pretty useful
nonetheless, Qwen2.5 is pretty useful
November 9, 2025 at 12:59 AM
trying to get typed output from LLMs is about as fun as chewing tinfoil
nonetheless, Qwen2.5 is pretty useful
nonetheless, Qwen2.5 is pretty useful
I don't think the parallel holds at all.
Not only you can unplug and reset an LM, but you can engineer its knowledge and steer it more or less at will, repeatably and without memory/histeresis.
I agree a precautionary principle should hold in general, but the particular differences matter a lot.
Not only you can unplug and reset an LM, but you can engineer its knowledge and steer it more or less at will, repeatably and without memory/histeresis.
I agree a precautionary principle should hold in general, but the particular differences matter a lot.
When I started my doctorate, I had to complete the standard research ethics training. Particularly with prisoner populations, I saw many parallels to AI. #ai #artificialintelligence #llms #largelanguagemodels #aiethics
When We Decide Who Can Feel
Should AI be protected by ethical research guidelines?
open.substack.com
November 8, 2025 at 3:34 AM
I don't think the parallel holds at all.
Not only you can unplug and reset an LM, but you can engineer its knowledge and steer it more or less at will, repeatably and without memory/histeresis.
I agree a precautionary principle should hold in general, but the particular differences matter a lot.
Not only you can unplug and reset an LM, but you can engineer its knowledge and steer it more or less at will, repeatably and without memory/histeresis.
I agree a precautionary principle should hold in general, but the particular differences matter a lot.
gods forbid i consider chatgpt as having legal personhood or indeed the framing of the 1A to be as universal as USians make it to be
November 8, 2025 at 3:26 AM
gods forbid i consider chatgpt as having legal personhood or indeed the framing of the 1A to be as universal as USians make it to be
even your favorite author writes a platitude sometimes 🥶
Like all walls it was ambiguous, two-faced. What was inside it and what was outside it depended upon which side of it you were on.
November 3, 2025 at 2:45 PM
even your favorite author writes a platitude sometimes 🥶
having deep dives on stuff I know close to nothing about, getting what looks like well structured sensible answers I can study, with examples that make sense in turn
I’m still not close to being able to explain the output to someone else, but I guess this counts as progress?
I’m still not close to being able to explain the output to someone else, but I guess this counts as progress?
October 29, 2025 at 5:59 AM
having deep dives on stuff I know close to nothing about, getting what looks like well structured sensible answers I can study, with examples that make sense in turn
I’m still not close to being able to explain the output to someone else, but I guess this counts as progress?
I’m still not close to being able to explain the output to someone else, but I guess this counts as progress?
What’s the sound made by 1000 microsoft VPs igniting spontaneously?
"I DON'T NEED YOU TO FUCKING REWRITE WHAT I'VE JUST WRITTEN!"
October 29, 2025 at 3:43 AM
What’s the sound made by 1000 microsoft VPs igniting spontaneously?
For once, there’s an easy fix: port your backend to Haskell, add stm and go on with your day : hackage.haskell.org/package/stm
October 28, 2025 at 12:16 PM
For once, there’s an easy fix: port your backend to Haskell, add stm and go on with your day : hackage.haskell.org/package/stm
Reposted by Marco Zocca
I’m in a wifi-starved situation and rely on mobile data for everything for a few days
I think I maxed out my plan and the mobile provider is throttling my traffic at the application level. Like, no youtube but text apps work? Github actions showing blank results? Wtf
I think I maxed out my plan and the mobile provider is throttling my traffic at the application level. Like, no youtube but text apps work? Github actions showing blank results? Wtf
October 28, 2025 at 9:33 AM
I’m in a wifi-starved situation and rely on mobile data for everything for a few days
I think I maxed out my plan and the mobile provider is throttling my traffic at the application level. Like, no youtube but text apps work? Github actions showing blank results? Wtf
I think I maxed out my plan and the mobile provider is throttling my traffic at the application level. Like, no youtube but text apps work? Github actions showing blank results? Wtf
Slow is smooth, smooth is fast
Automated Guided Vehicles at Long Beach Container Terminal manufactured by Terex Port Solutions.
October 27, 2025 at 6:24 PM
Slow is smooth, smooth is fast
Good stuff. Scientific software and in particular optimization codes are everywhere, yet the magic sauce is often proprietary and found heuristically.
Analysis such as this one is net progress for the field.
Analysis such as this one is net progress for the field.
The simplex algorithm is super efficient. 80 years of experience says it runs in linear time. Nobody can explain _why_ it is so fast.
We invented a new algorithm analysis framework to find out.
We invented a new algorithm analysis framework to find out.
Beyond Smoothed Analysis: Analyzing the Simplex Method by the Book
Narrowing the gap between theory and practice is a longstanding goal of the algorithm analysis community. To further progress our understanding of how algorithms work in practice, we propose a new alg...
arxiv.org
October 27, 2025 at 10:12 AM
Good stuff. Scientific software and in particular optimization codes are everywhere, yet the magic sauce is often proprietary and found heuristically.
Analysis such as this one is net progress for the field.
Analysis such as this one is net progress for the field.
Slop hogs, riddle me this: the Sonnet 4.5 I use with my work account is over the top enthusiastic, the one with my personal account plain and factual. Code quality and quantity per prompt roughly comparable.
verdict:
Claude is smart and even too chatty. Fills hundreds of lines in a go as well as self-celebratory SUCCESS.md logs.
4o : "here's that bs you asked. Let me know if you need 10 more lines."
Claude is smart and even too chatty. Fills hundreds of lines in a go as well as self-celebratory SUCCESS.md logs.
4o : "here's that bs you asked. Let me know if you need 10 more lines."
let's see if this agentic coding is any good
October 27, 2025 at 10:08 AM
Slop hogs, riddle me this: the Sonnet 4.5 I use with my work account is over the top enthusiastic, the one with my personal account plain and factual. Code quality and quantity per prompt roughly comparable.
Reposted by Marco Zocca
Are there any posts about how OpenBLAS really absolutely was the big bang moment for what is now the explosion of python and data science and the way we see the world running so much AI and ML all came from that shit getting released and unseating the intel kernel math library
October 26, 2025 at 4:42 AM
Are there any posts about how OpenBLAS really absolutely was the big bang moment for what is now the explosion of python and data science and the way we see the world running so much AI and ML all came from that shit getting released and unseating the intel kernel math library
Codespaces + agent mode = prototype all the things
October 25, 2025 at 6:24 AM
Codespaces + agent mode = prototype all the things
Reposted by Marco Zocca
My new article on the illusion of AI productivity.
We often feel more productive with AI, but are we really?
Read here newbeelearn.com/blog/ai-prod...
#AI #Productivity #FutureOfWork #DigitalTransformation #AItools
We often feel more productive with AI, but are we really?
Read here newbeelearn.com/blog/ai-prod...
#AI #Productivity #FutureOfWork #DigitalTransformation #AItools
Does productivity really increases with ai or it just feels like it?
story of how the tool's save frame from video was useful for landing page
newbeelearn.com
October 25, 2025 at 4:52 AM
My new article on the illusion of AI productivity.
We often feel more productive with AI, but are we really?
Read here newbeelearn.com/blog/ai-prod...
#AI #Productivity #FutureOfWork #DigitalTransformation #AItools
We often feel more productive with AI, but are we really?
Read here newbeelearn.com/blog/ai-prod...
#AI #Productivity #FutureOfWork #DigitalTransformation #AItools
You can actually feed the article TeX source to copilot as a reference and it will produce code and examples. Neat
Taking Haiku 4.5 for a spin, this time with a scientific computing codebase.
Is there something like an “arXiv skill” for LLMs, so they can read up on research methods on the fly?
Is there something like an “arXiv skill” for LLMs, so they can read up on research methods on the fly?
October 24, 2025 at 1:21 PM
You can actually feed the article TeX source to copilot as a reference and it will produce code and examples. Neat
Reposted by Marco Zocca
Computer science courses that don't exist, but should (2015) [Discussion]
Computer science courses that don't exist, but should (2015)
Computer science courses that don't exist, but should (2015)
prog21.dadgum.com
October 24, 2025 at 5:52 AM
Computer science courses that don't exist, but should (2015) [Discussion]