Matthew Fong

Seeking rational truth in a tribal world. Founder of Seldon Capital. Applying a scientific approach to fundamental forecasting. Formerly Soros Fund Management.

[Matthew Fong @mattfong_]

@JFPuget Many AI companies seem to overfit on evals. Have you thought about applying the same Kaggle private leaderboard approach at NVIDIA?

[Matthew Fong @mattfong_]

@dylan522p I’ve worked with Wei for many years and I’ve never seen him show up drunk at work. He was a 10-100x analyst like Jim Rogers and was always very principled and held strong opinions about what’s right.

[Matthew Fong @mattfong_]

@__paleologo A better way to reason about index investing:
https://t.co/HbBNbXJQQJ

[Matthew Fong @mattfong_]

@burkov Hallucinations are overfit predictions. They decrease if the model generalizes better out of distribution and increase if the model starts to overfit (even if performs better on more training samples).

[Matthew Fong @mattfong_]

@tunguz An example of this is $TLN. Natural gas peakers are required to supplement intermittent solar and wind power whereas coal and nuclear plants will have more power production which overlap with solar and wind power production.

[Matthew Fong @mattfong_]

@__paleologo @k3ithmccullough I think the interesting thing about Baum’s piece is he actually experimented with an early form of machine learning by detecting lines in charts in currencies and commodities. This was different from the short term reversal effect traded by Thorp, Shaw etc in equities.

[Matthew Fong @mattfong_]

@__paleologo https://t.co/JclRl7vuNa

https://t.co/2bcqbqoSJJ

[Matthew Fong @mattfong_]

@__paleologo Depends on capabilities and whether current AI models overfit (and yes, I’m aware of deep double descent). Assuming AI primarily generalizes the training set, most 1m-1y alpha disappears. A few Ren Techs/HRTs, index funds, activists, investors who identify paradigm shifts remain

[Matthew Fong @mattfong_]

@eshear @predict_addict Where test set has high overlap with training set, which might be most word sequences, transformers will do well. Where test set has less overlap with training set, transformers will do less well. Many time series problems require generalizing to extrapolate, not just interpolate

[Matthew Fong @mattfong_]

@RylanSchaeffer @random_walker Agree that on 1, they needn’t have anything to do with test-set contamination. For LLMs, models might have already seen questions in the test set (usually some benchmark) in training (because the training set is the whole internet). Test error may not measure generalization

[Matthew Fong @mattfong_]

@paulg Ranked choice voting instead of plurality voting results in politicians in the middle having more market power and the downstream effects are less polarization in politics and media

[Matthew Fong @mattfong_]

@typesfast NVR

[Matthew Fong @mattfong_]

@paulg Forecasting the sustainable demand for different parts of the semiconductor supply chain is difficult and so there tend to be mismatches in terms of supply demand as production capacity is costly to add, especially parts of leading edge chips where there are only 1-2 suppliers

[Matthew Fong @mattfong_]

@INArteCarloDoss The next valeant / gotham?

[Matthew Fong @mattfong_]

@Suhail TSMC is heavily diversified and exposed to non-leading edge end markets as well. SMCI and 2376 TT have more concentrated exposure to Nvidia.

[Matthew Fong @mattfong_]

@valueterminal You on Threads yet?

[Matthew Fong @mattfong_]

@chrisalbon @jeremyphoward Progressive regulation - regulate the largest the most and leave small companies alone

[Matthew Fong @mattfong_]

@linakhanFTC @nytimes @FTC You allowed Microsoft to create an AI monopoly but cracked down on Meta which has many competitors (Tik Tok, Snapchat, Twitter, Signal etc)

[Matthew Fong @mattfong_]

@paultoo TSMC and Samsung (the 2 leading edge foundries) are already building fabs here and they are higher cost and not leading edge.