I SEE (study, engineer, educate) computer networks. Tweets are facts; retweets are alternative facts
Joined May 2008
- Tweets1K
- Following718
- Followers1.5K
- Likes1.4K
These exorbitant fines are why stadium beer costs so much.
Ronaldo fined £215 by Portuguese FA
Barcelona fined £503 over approach for Alvarez
bbc.com/sport/football/artic…
bbc.com/sport/football/artic…
The full paper on our self-defining operator is now out.
We show how an agentic operator that remembers past incidents and synthesizes its own playbooks cuts cost and MTTR by over half.
arxiv.org/abs/2610.14547
Early report from the first project we are pursuing as part of the "self-defining systems" research, led by @sudopowr.
So much potential to lower operational toil and MTTR!
Ratul Mahajan retweeted
If you’re heading to #NANOG98, come say hi.
I’ll be talking about how we flattened datacenter networks at Amazon: no more trees, just (carefully-engineered) randomness.
Monday, 1:30pm, Trianon
nanog.org/events/nanog-98/co…
If you’re heading to #NANOG98, come say hi.
I’ll be talking about how we flattened datacenter networks at Amazon: no more trees, just (carefully-engineered) randomness.
Monday, 1:30pm, Trianon
nanog.org/events/nanog-98/co…
Why I dread reading cellular networking papers:
A paper I'm reading has:
- 77 abbreviations, 30 of which are never expanded. WTF is UPF?
- Abbreviations account for 20% of nouns and 8% of all words.
- 20% of sentences contain 3+ abbreviations.
Two things can be true at the same time.
You say “two things can be true at the same time” to sound thoughtful.
You haven’t thought at all.
How do I sign up for the OpenAI pen testing service?
"OpenAI Agents Tried to Hack Four More Websites While Seeking Data"
wsj.com/tech/ai/openai-agent…
Ratul Mahajan retweeted
This is super early, and this post mainly focuses on the interesting failures we observed. The proposed architecture for a social harness (which will complement existing harnesses) is described in the paper (arxiv.org/pdf/2609.17527) and we will discuss it more in subsequent posts here as well.
The Internet stack enabled computers to communicate at scale. Tapan is exploring what a networking stack for agents needs.
(No, A2A is not enough; you'll have to read the paper to find out why.)
1/ We just published our first post on why agents cannot yet collaborate autonomously across trust boundaries (e.g., talk to other people's agents), and what we think is missing. (w/ @arvind_uw @ratulm)
social-harness.org/blog/agen…
OpenAI burned $22.5M worth of compute to solve a $1M problem. Math bounty hunting isn't so lucrative after all.
Sabalenka should train in Seattle to build tolerance.
You can go running at 7 in the morning and inhale it in most neighborhoods.
foxnews.com/outkick-sports/t…
Those who haven't worked in large companies will be confused by this news: What?! Microsoft AI Frontiers is not already inside Microsoft AI??
No, that is not how things work.
Congratulations to the team! Great sign of impact.
Ratul Mahajan retweeted
The trend will continue and it's not only about the number of papers but the quality of research. The academia in China has its own problems but it's making progress in its designed ways, while US chooses to devalue academia and fundamental research.
The open secret of US academia/research is being able to attract top talent from all over the world. It's no longer true, at least for Chinese students.
At the time international students all choose to leave, faculty will follow, and places like UIUC will become a true corn field.
@mino98 ran the numbers for SIGCOMM:
2016 2026
USA 79% 34%
China 10 57
Other 10 9
"In 2016 the US produced 8x as many papers as China. In 2026 .... China alone has more than everyone else combined."
Ratul Mahajan retweeted
I just went to NSDI, a top venue for computer systems research. Change in papers' src country:
2015 2025
USA 76% 45%
China 10 42
Other 14 13
China's research investments are bearing fruit, while the US is gutting science funding.
"You don't need to look at the code when using coding agents" only holds if you one-shot the code and never touch it again.
The moment you need to debug it, the agent starts talking about retry_count and some function from three days ago—and you're lost because you've never actually looked at the code.
Do you remember this great piece from a year ago? I wonder what these researchers would say now that LLMs have become subjects of geopolitical tension.
When ChatGPT Broke an Entire Field: An Oral History quantamagazine.org/when-chat… via @QuantaMagazine
What makes me nervous about using coding agents to design complex systems:
When designing new types of complex systems, you typically do not know ahead of time which combination of ideas will work. It is an iterative process where you hypothesize and test many ideas, often via prototype implementations. The difficulty of implementing an idea is a strong indication that it would be fragile in practice. Even when easy to implement, the act of implementing an idea makes you aware of the corner cases that you didn’t anticipate (I am not that smart).
By taking away the implementation burden, coding agents also take away a key learning signal. The result could be fragile systems that break in ways you didn’t see coming.
We are looking for graduate students and postdocs to join our Self-Defining Systems project at UW.
self-defining-systems.uw.edu…
If you're interested in exploring how AI agents can build complex systems end-to-end, we'd love to hear from you! If you know someone who might be a good fit, please help spread the word.