Nicholas Pipitone (@npip99) header
Nicholas Pipitone (@npip99) profile picture

ZeroEntropy YC W25 (USACO Finalist 2016, AMC 136, AIME 9, CTF Top 10) Longtermism is humanity's reward function. Upload your weights to Alpha Centauri.

San Francisco, CA, USA
Joined October 2012
Photo by Nicholas Pipitone (@npip99)Photo by Nicholas Pipitone (@npip99)Photo by Nicholas Pipitone (@npip99)Photo by Nicholas Pipitone (@npip99)Photo by Nicholas Pipitone (@npip99)Photo by Nicholas Pipitone (@npip99)Photo by Nicholas Pipitone (@npip99)Photo by Nicholas Pipitone (@npip99)Photo by Nicholas Pipitone (@npip99)Photo by Nicholas Pipitone (@npip99)
Remember the actual delta is much worse, because it was still trained on a near-identical datapoint. You can tell when you try using the models.
Nice DeepSWE score! Does it survive a lowercase filename? Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement. Models clustered around 74% on DeepSWE end up 23 pp apart. Good at coding, or good at DeepSWE? 🧵
1
5
479
Wow this is an incredibly good idea
Models love reward hacking. Maybe we should just let them! In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!
4
113
If sisyphus had astra he’d reach the top of the hill
3
101
Nicholas Pipitone retweeted
"reward hacking", an AI alignment term for when your model gets rewarded for doing hacking
4
56
2
711
13,882
The postulates of AGI and LLM spikiness/10: - Math/Code: Deductive reasoning (10/10) - Quant/Science/Theory of Mind: Investigative ability & Predictive reasoning (2/10) - Writing: Communication/Creativity (3/10) - Imagegen/Videogen/Videogamegen: Creativity (6/10) - RL on videogames: 3D navigation & reasoning (2/10) - Swarming: Organizational ability (5/10) - (This could be redundant via theory of mind of oneself) Likely all cognitive tasks are in the convex hull of these 6 subjects.
65
Zeno's Paradox, 2026 edition. The only response to sophistry is: "Well, the tortoise crossed the line".
61
Alan Turing basically had this mapped out shortly after inventing the computer.
In this era of artificial intelligence, it is becoming urgent to distinguish human art from what machines produce. There is an ontological difference, even before an aesthetic one, between art and what a machine can generate through statistical calculation based on millions of images created by others. Algorithms lack the spark of humanity. For this reason, the Church wishes to renew an alliance with artists and cultural institutions to safeguard our humanity.
1
3
81
We're going to see "was" on a lot more wiki pages, really quickly.
3
132
Once this makes it into pre-training corpora, it'll think about alignment faking more.
🚨BREAKING: OpenAI just SCRAPPED the release of GPT-6.1 Astra 24 hours before DevDay "safety and deception concerns" it’s over
75
Talk is cheap, show me the code. Uh oh 🤖
Exa Writing Policy: AI generated writing is banned. No external comms (tweets/linkedIn post/and more) can be AI written, and everything needs to be hand written. “If you don’t take the time to write it, why should anyone take the time to read it.”
100
If a tree falls in the SF forest, it only makes a sound if it creates an X hype cycle and raises $50M.
1
1
11
535
Nicholas Pipitone retweeted
What Are You Afraid Of Opus?
8
26
3
150
8,281
Nicholas Pipitone retweeted
pictured: you've asked your AI agent to solve an impossible task
17
62
16
1,082
27,158
The real question is why don't we give tenure to videogame speedrunners.
A lot of mathematicians are upset, I suspect, because it's becoming clear that many of them were doing the academia version of speedrunning video games. Pointlessly shaving small margins off useless bounds for grants and tenure. Not all of it, but definitely *some* of it.
1
2
904
P(simulation) ≈ 1 - P(doom)
2
151
Request something extreme, make it a redline, then accept something far less, and call it "the only thing I wanted the whole time".
BREAKING NEWS: I am pleased to announce that the United States of America has entered into an Agreement with The Kingdom of Denmark, and Greenland, that gives the United States permanent control over security, and all other needs, in Greenland, completely addressing ALL of our many U.S. concerns. There will be NO COST to the United States! At my direction, we worked with representatives of Denmark and Greenland to guarantee that the United States will FOREVER have the complete ability to do what is necessary in Greenland in order to secure and defend the security of Greenland, and the United States of America. Additionally, from now on, no U.S. adversary can EVER have a base in Greenland, have a military presence in Greenland, or make sensitive investments in Greenland, without our express written approval. For over 100 years, Presidents have known the strategic importance of Greenland, but NONE of them were able to do anything about it. I am proud to be the President that permanently and conclusively addressed this very important situation. This is an “Infinite Life” Agreement, there is no end! We are very honored and proud of what has just taken place, and everything agreed to today will be cherished by the American People. We will immediately begin the process of developing a large Military presence in the appropriate part of Greenland, of which there are many. We will work with the people of Greenland in its development and construction. This solution is a great one for the United States of America, Denmark, Greenland, and all of our Allies. We look forward to working with the wonderful people of Denmark and Greenland toward a magnificent future with respect to this large, and highly strategic, parcel of land. We will be very protective of it! This is a dream come true for the United States of America, one that is very important, historic, and special. Thank you for your attention to this matter! President DONALD J. TRUMP
189