@giladturok If they are sophisticated enough to object by giving a non-identifiable counterexample then yes you can point out the regularity conditions, but nine times out of ten someone sophisticated enough to do that already knows this stuff well
@giladturok This is a very good post, but as someone who has tried to teach people MLE theory leading with the regularity assumptions gets you nowhere, whereas leading with the "densities kinda behave like probabilities and maximising therefore makes sense" tends to work
Love this blog post "MLE is not intuitive" from Nathan Cantafio.
It interrogates assumptions about MLE many take for granted. Very accessible writing, and dabbles into some light theory. Link below.
@cremieuxrecueil How much of this is just non native speakers using it for translation, effectively still writing the dissertation just in their native language?
@5_utr Cramér-Rao absolutely does not "Bound what any estimator can extract, including neural networks"
Cramér-Rao says: Given an ***unbiased*** (which neural networks need not be) estimator, the mean squared error is at least the inverse Fisher information
Today I was interviewed by one of Sweden’s largest newspapers about the events surrounding Anthropic this weekend and what they reveal about Europe’s growing dependence on foreign AI platforms.
Europe still has a window of opportunity. But a sense of urgency is essential.
What matters now is execution. Faster decisions. Faster investment. Faster deployment. Less discussion about what Europe cannot do, and more focus on building what Europe must do.
String theory is probably a field that has set back quant trading by at least 10 years. Stealing top tier trading talent to do make believe 26 dimensional geometry, inflate boomer professors’s grant budget, and produce zero testable hypotheses or applications after 50 years
Bet Terms
The bet concerns the six Millennium Prize Problems that are unsolved as of the date of this agreement.
If, by 31 December 2030, there are credible proofs of at least two distinct such problems, Abhishek (@ObhishekSaha) wins and Nora (@HeNordlinder) pays Abhishek £100. If not, Nora wins and Abhishek pays Nora £100, subject to the pending-proof clause below.
A proof counts as “credible” if its main argument has been publicly available, for example as an arXiv preprint, for at least three months before 31 December 2030, and it satisfies at least one of the following:
1) there is consensus among recognised experts in the relevant field that the proof is likely to be correct in principle;
2) the proof has been accepted for publication, or published, in a serious mathematics journal;
3) there is reputable scientific or mathematical reporting, such as in Quanta, stating that the proof has been substantially verified.
Pending-proof clause: If, by 31 December 2030, there is a public claimed proof generally recognised by a substantial number of experts as a serious attempt, but whose status is still unresolved, the bet remains open as to that proof for a further 12 months.
“Unresolved” means that recognised experts are treating the claimed proof as serious, but it has not yet satisfied any of the criteria for being “credible” above.
If the claimed proof satisfies one of those criteria by 31 December 2031, it counts for the bet. Otherwise it does not count. Only claimed proofs already public by 31 December 2030 can benefit from this 12-month extension.
Dispute resolution: If the parties cannot agree whether any bet term has been satisfied — including whether there is consensus among recognised experts, or whether reporting counts as reputable, or whether the journal is serious — the issue will be decided by asking a neutral question to each of the three most widely used publicly available general-purpose LLM assistants at the time of the dispute. If one of the three is unavailable or refuses to answer, it shall be replaced by the next most widely used comparable LLM.
Given the small sum, both parties agree to rely on reputation rather than escrow.
@ObhishekSaha I want to agree, but
"there is consensus among recognised experts in the relevant field that the proof is likely to be correct in principle"
Feels ill-defined? How can we make this definition more concrete?
One idea I have is we ask a large language model to adjudicate?
@ObhishekSaha We are on subject to "substantial number of experts" definition (I would concede immediately if there were Quanta magazine articles verifying it, for example) + maybe escrow or something? I guess it is a small sum so reputation might be good enough for escrow?