I've always just used to regular Chat mode in ChatGPT.
Today is the first time I used work mode.
How can anyone think white collar / computer work is going to survive in a few years?
A neural network remembers how it was monitored early in training. On two copies of a minimal network, changing only the first ~2k steps in monitoring and training the next ~50k steps under the same training rules results in a 26 pt accuracy gap after a fresh erasure.
After the 50k shared steps, one copy scored 76% and the other 50%. The 50% copy has not forgotten the task. Between monitor updates, both copies scored ~78%. However, at each update, the copy that started with slow updates drops about 29 points every time. The copy that started with faster updates drops about ~ 3. Paper and code coming soon!
Using two copies with the same starting weights and the same training data, the first 2,000 steps, one copy's monitor is updated every 5 steps and the other's every 50. Both copies' monitors updated every 20 steps for 50,000 following steps. This was repeated for 20 pairs.
it is kind of crazy that new frontier models get more aligned (at least in the obvious ways) but because they are expensive, they are not the most used ones, leaving a safety/alignment gap between what’s used and the safest models
using algebraic encodings of thin matrix products, with caching to reuse shared encodings, small-bias sets for derandomization, and modular certification to make counting exact, we can produce best-known exponents for deterministic integer APSP and 3SUM: O(n^2.99269) and O(n^1.99311).
this improves on Alman–Vassilevska Williams (n^2.9995, n^1.9992) and Majdoddin (n^2.9945, n^1.9945)
thanks to claude (opus and fable), astra, and argon for majorly driving this research, hopefully smarter people and smarter models can drive this even more :)
lean results below