gansandroses

@paulg @sama Film is the way to go. Piql make a nice one with an open source reader

@sama If you had them etched on a piece of metal in the most compressed form, how big would the piece of metal have to be? This is a mostly serious question. These models are history, and by default digital data evaporates.

goodbye, GPT-4. you kicked off a revolution. we will proudly keep your weights on a special hard drive to give to some historians in the future.

Emailed the CEO of Amex after this tweet. Going in to Amex HQ today. Quarterly reminder to shoot your shot

@sandstep1 @nikete @aaron_defazio There's an R example here: cran.r-project.org/web/packages/g…

@nikete @nftstatistician @aaron_defazio May you share code example

L1 regularization for sparse solutions - as usually taught - is actually terrible in practice! I’m always surprised how few people know this. To get good results, retrain with the sparsity pattern found from the initial L1 run, but without the regularizer. Works much better.

@nikete @aaron_defazio The paper contains asymptotic analysis, but intuitively, when the number of parameters is very large compared to the number of samples, you'll need a lot of L1 regularisation for variable selection and you end up shrinking the remaining coefficients too much.

@nftstatistician @aaron_defazio do we understand why relaxed lasso works better than lasso?

@joftius @yuvalbenj @tabletmag The argument is indeed unsupported. This is valid critique. Attacking the cum-sum plot is just besides the point. I don't know what the data looks like for a "typical" armed conflict, would love to see data. The correlations are at least surprising though.

@nftstatistician @yuvalbenj @tabletmag The argument is also unsupported, most probably because it is incorrect -- in addition to the figure being misleading

One of the worst abuses of statistics I've ever seen in the @tabletmag article by Abraham Wyner (a professor of statistics at Wharton! 😲) I read the rest of the article, and even if it didn't include the absurd plot that Lior rebutted it's still shockingly irresponsible

@joftius @yuvalbenj @tabletmag The author claims the variation is lower than anticipated for an armed conflict (without supporting data). He talks about the underlying data, not the sums. The visualisation may be misleading but that doesn’t negate the argument.

@yuvalbenj @tabletmag I think Lior accurately rebutted the "Figure 1" of the article. He may not have addressed the other points, but that doesn't make his critique weak summary(lm(cumsum(deaths) ~ day, data.frame(day = 1:15, deaths = rnorm(15, mean = 270, sd = 100)))) ^ that's just an absurd Fig 1

@lpachter @adiwyner I agree that the claims at the end (about the mechanism and the real death count) can’t be determined from the data, but on this he specifically says (without reference though) that 15% SD in the daily rate is lower than expected in armed conflicts.

@nftstatistician @adiwyner Thanks for your feedback. But what @adiwyner wrote is "Most likely, the Hamas ministry settled on a daily total arbitrarily. We know this because the daily totals increase too consistently to be real." So @adiwyner did base his conclusions on the cumulative sums.

Just learned that Bibi read my article. That’s nothing compared to the time Derek Jeter responded to my article (which showed he had very little range up the middle) .

@lpachter @adiwyner Your critique is very weak. The author reports the variation daily rate of casualties, not the cumsum, and only uses cumsum in a visualisation. He also makes other supporting analyses. Not sure why you’re pushing this.

@adiwyner I also read your article. liorpachter.wordpress.com/2024/03/08/a-n…

@erlichya @satanslavegod @ArbleMatan wordsrated.com/academic-publi…

אחרי שראיתי כמה כסף עלה לפרסם את המאמר, בא לי להתקלח כדי להוריד ממני את הזוהמה. 2900 דולר. איזה בושה. אם היה לי קצת יותר אומץ, הייתי מפרסם רק בביוארכיב וזהו. זה חלק שכל כך מגעיל אותי בלהיות אקדמאי.

Yes, I do. LLMs produce their answers with a fixed amount of computation per token. There is no way for them to devote more (potentially unlimited) time and effort to solving difficult problems. This is very much akin to the human fast and subconscious "System 1" decision process. True reasoning and planning would allow the system to search for a solution, using a potentially unlimited unlimited time for it. This iterative inference process is more akin to the human deliberate and conscious "System 2". This is what allows humans and many animals to find new solutions to new problems in new situations. Some AI systems have planning abilities, namely those that play games or control robots. Game playing AI systems such as AlphaGo, AlphaZero, Libratus (poker), and Cicero (Diplomacy) have planning abilities. These systems are still fairly limited in their planning abilities, compared to animals and humans. To have more general planning abilities, an AI system would need to possess a *world model*, i.e. a subsystem that can predict the consequences of an action sequence: given the state of the world at time t, and an imagined action I could take, what would be the set of plausible states of the world at time t+1. With a world model, a system can plan a sequence of action so as to fulfill an objective. How to build and train such world models is still a largely unsolved problem. Even more complex is how to decompose a complex objective into a sequence of sub-objectives. This would enable hierarchical planning, something that humans and many animals can do effortlessly but is still completely out of reach of AI systems.

I'm delighted to release the first half of my new open-access online textbook in human population genetics: web.stanford.edu/group/pritchar…

@Noahpinion @LevyAntoine This still doesn't account for age. It just shows 5 years post diagnosis. Average diagnosis age matters.

@LevyAntoine Also:

A big win for the U.S. health care system:

@devonzuegel As an argument against innovation / venture it's pretty bad though

@devonzuegel As an argument for passive investing you only need a much more permissive argument to hold true: your chances to pick assets / money managers that beat the market after accounting for fees are pretty bad.

The Efficient Market Hypothesis may be one of the most silently harmful concepts of the last century Plenty of other ideas have had massive visible harms (e.g. various flavors of totalitarianism), but the EMH has stopped people from starting projects in the first place

@punk6529 Medical sciences are substantially bottlenecked by the need to conduct physical experiments, leading to long iteration times on knowledge acquisition, with or without AI. AI can maybe help us pick our experiments better.

If artificial super intelligence (asi) is 10 years away, there is huge option value in staying alive until then. Sure, asi might kill us all, but, if it doesn't, it will revolutionize healthcare. So, I dunno, eat your veggies and wear your helmet?

@dcinvestor @alexherrity Interest rates are not going to stay high for very long

@alexherrity yes, there will be exceptions for star employees in some cases also, many founders about to realize that life is a lot harder in a higher interest rate world

hate to break it to you, but almost everyone will be forced to go back to the office in the near future pretty soon, white collar labor will never have had less bargaining power

@Andrew_Marshall Because tech is the only viable growth engine the UK has post-brexit, and the cost to avoid serious harm to the ecosystem is very modest

People always fret about moral hazard, until there's a real world case, when they put aside the argument. Why on earth should UK taxpayers pay back lenders to #SVB? Can't be systemic risk here.

@clairesilver @SuperchiefNFT @WME @Variety Congratulations! well deserved and an important moment for AI ART!

@atroyn really need more technical details to think about this as “attribution” and not “find the most similar images out of a dataset of 5B”

announcing stable attribution - a tool which lets anyone find the human creators behind a.i generated images. stableattribution.com