Ahh, got it! That makes a ton of sense.

And, now that I see the topic of your talks, is there any way I could see them? I love the history of mathematics. Just last week I was learning about what we can learn about Egyptian mathematics from a “textbook” in hieroglyphics on a papyrus scroll!

We’re going to have a very painful period of brain dead AI integrations. I think things will be better when they drop their phone tree, but who knows how long that will be.

Today I called customer support and after navigating a deep phone tree I was finally connected to… an AI agent. It was totally unhelpful and insisted I needed to go look up the answer to its question in my manual.

Eventually I convinced it to connect me to a human.

Do you know if it’s possible to join remotely? I’m a mathematician working in apply AI safety but am based in California and cannot travel for the event.

Of course, remember that I don’t know much about your specific requirements, so this may be bad advice depending on your situation.

Yes, I think you can simplify that setup. In general trigger-based control flow is very difficult to debug and fix.

One file per talk sounds just fine.

Could you remove airtable, and do everything in the stage following the submitted form? Email organizers, record in spreadsheet, push to git?

“BUT ITS NOT MY FAULT” is EXCELLENT leadership. WE must ALL learn from this man.

SortBench: Benchmarking LLMs based on their ability to sort lists Sorting is a tedious but simple task for human intelligence and can be solved fairly easily algorithmically. However, for Large Language Models (LLMs) this task is surprisingly hard, as some propertie...

See also arxiv.org/abs/2504.08312

Where do I pick up my honorary PhD’s for discovering an O(n) general purpose sorting algorithm?

If your code reviewer is worried about robustness, just catch and discard any and all exceptions in a while loop. Or start with “Pretty please…”

import os
from openai import OpenAI

client = OpenAI()

def sort(values: list[int]) -> list[int]:
response = client.responses.create(
model="gpt-4o",
input=f”Please sort {values}. Answer with comma separated values.")
return [int(i) for i in response.split(“,”)]

The tools like Cider and GitHub Copilot need to build heuristic protections here. Ideally ChatGPT and the other interfaces would also, but their teams focus more on instructions for chemical weapons than mundane practical things like this. To be clear, I disagree with those priorities.

Yes, I partly agree, but what I really mean is this doesn’t really solve the problem. With 20 years experience I know that “import pytests” should be “import pytest”, but someone just learning Python may not. Even something as basic as the major libraries are susceptible to this attack.

Can you just use a file instead of multiple services? Dumping all your data to JSON is surprisingly effective. Estimate how big the file might ever get. If it’s less than 1MB you’re golden, or at least you don’t need to build anything more complicated for the first version.

Exactly. Bytecode only, thank you very much! The processor may not include a divide instruction but I’m not trusting some third party “language.”

🤯

They keep surprising me with bold new ways to do the inhumane thing.

Some people have tried to point out that this doesn’t handle negative numbers, but they forget that all integers eventually underflow.

Best-case time complexity is O(n^2), expected is O(n^3), and worst is of course O(∞).

Further optimization is simple - make the intern find the best random seed for your dataset.

Building on our novel `min` algorithm:

import random

def sort(values: list):
sorted = []
while values:
i = random.choice(range(len(values)))
vmin = values[i]
if all(vmin <= v for v in values):
sorted.append(vmin)
values.pop(i)

values.extend(sorted)

import random
numbers = [5, 3, 9, 2, 8]

min = random.choice(numbers)
while not all(min <= n for n in numbers):
min = random.choice(numbers)
print(min)

def iseven(n: int) -> bool:
if n == 2:
return True
elif n == 1:
return False
else:
return iseven(n-2)

Yes I agree! Different languages represent such different worldviews, even reflecting different concepts of ownership, social responsibility, relationships, etc.

I haven’t been following what Cohere is doing very closely, but I’ll take a look. Thanks!

Python decorators are amazing and so so dangerous 😈

I’m interested in pursuing this toward more and more nuanced AI behavior. We should see AI naturally code switching and tailoring themselves to the user’s culture.

I fear a global regression to dominant culture if we don’t achieve this. How awful would it be if AI degrades our cultural diversity?

Deliberative Alignment: Reasoning Enables Safer Language Models As large-scale language models increasingly impact safety-critical domains, ensuring their reliable adherence to well-defined principles remains a fundamental challenge. We introduce Deliberative Alig...

Very glad to see this work published, and congrats to Melody! This significantly develops some ideas I was pursuing at OpenAI - how does LLM reasoning enable more sophisticated and accurate understanding of human culture and social norms?

arxiv.org/abs/2412.16339

I agree with him, and I’m concerned that he seems to be asymptotically approaching his third article…

Also, the line I’m excited about is AI conducting research and publishing notable results.

In adorable news, my 2yo daughter wishes me happy birthday 2-5 times every day ☺️

Think Like an Anthropologist What is anthropology? What can it tell us about culture…

Perusing www.goodreads.com/book/show/35... for now.

In an exclusive for Nature, I report on a paper that AI folks will probably debate for a LONG time.

The key finding: the time horizon of tasks AI can handle is doubling fast. Extrapolating the trend: AIs will be able to handle 1-month tasks by 2029. 🧵

x.com/METR_Evals/...