laplacesdemon48

I recently subscribed to Perplexity Pro and prior to this release, was already strongly considering discontinuing ChatGPT Premium.

When I first subscribed to ChatGPT Premium late last year, the natural language understanding superiority was amazing. Now the benchmark advances, low latency voice chat, Sora, etc. are all really cool too.

But my work and day-to-day usage really rely on accurately sourced/cited information. I need a way to comb through an ungodly amount of medical/scientific literature to form/refine hypotheses. I want to figure out how to hard reset my car's navigation system without clicking through several SEO-optimized pages littered with ads. I need to quickly confirm scientific facts, some obscure, with citations and without hallucinations. From speaking with my friends in other industries (e.g. finance, law, construction engineering), this is their major use case too.

I really tried to use ChatGPT Premium's Bing powered search. I also tried several of the top rated GPTs - Scholar AI, Consensus, etc.. It was barely workable. It seems like with this update, the focus was elsewhere. Unless I specify explicitly in the prompt, it doesn't search the web and provide citations. Yeah, the benchmark performance and parameter counts keep impressively increasing, but how do I trust that those improvements are preventing hallucinations when nothing is cited?

I wonder if the business relationship between Microsoft and OpenAI is limiting their ability to really compete in AI driven search. Guessing Microsoft doesn't want to disrupt their multi-billion dollar search business. Maybe the same reason search within Gemini feels very lacking (I tried Gemini Advanced/Ultra too).

I have zero brand loyalty. If anybody has a better suggestion, I will switch immediately after testing.

Here's another vote for 3Blue1Brown's series on linear algebra [1]. Spending a few hours on this series will easily save you dozens of hours when going through a comprehensive LA textbook.

[1] https://www.youtube.com/watch?v=fNk_zzaMoSs&list=PLZHQObOWTQ...

For anyone looking for a refresher, I highly recommend this: https://bookdown.org/probability/beta/

This is actually co-written/edited by Joe Blitzstein, who teaches Harvard's Stat110. What originally sold me on taking a look is this snippet from the foreword:

"This book is primarily a teaching tool, and thus we are not very concerned with proofs unless they promote understanding...There are no Statistics or linear algebra prerequisites."

Richard McElreath's content is a breath of fresh air for anyone who's struggled with stats.

I read both editions of his textbook and will be revisiting this new lecture material soon. I highly recommend you check out his book/course if you've been frustrated with trying to learn stats and want to practically understand things without hundreds of pages of proofs.

https://xcelab.net/rm/statistical-rethinking/

This book had a big influence on me. I highly recommend it. I read both the first and second editions.

It is very clearly written, full of practical examples (with code), and doesn't assume heavy math knowledge.

This section from the preface sums up the intended audience:

"The principle audience is researchers in the natural and social sciences, whether new PhD students or seasoned professionals, who have had a basic course on regression but nevertheless remain uneasy about statistical modeling. This audience accepts that there is something vaguely wrong about typical statistical practice in the early twenty-first century, dominated as it is by p-values and a confusing menagerie of testing procedures. They see alternative methods in journals and books. But these people are not sure where to go to learn about these methods."

Is there a way to extend this to measure a multivariate relationship?

For example: Cor(X, Y & Z)

I know you could run them pairwise but it’s possible Cor(X, Y) and Cor(X, Z) are close to zero but Cor(X, Y & Z) is close to 1.

In case anyone wants to see what this looks like, here's a video I found:

https://www.youtube.com/watch?v=C1eg-jgLx5o

Have either of you tried Pleese? I believe they're mostly selling to pizzerias now [1] and not supermarkets. They aren't able to scale up mass-production quickly because they use a different approach/ingredients - "proprietary blend of bean and potato proteins" [2].

I eat regular cheese and have tried several vegan brands, this seems pretty close to me. But I'm no cheese connoisseur.

[1] https://www.pleesefoods.com/availabilty

[2] https://www.pleesefoods.com/products

Have you seen the 1948 film? If so, how does it compare to the book?

I want to be perfectly clear that I didn't bring this up to be alarmist. Jesse Gelsinger's death shed a lot of light on the risks involved with adenoviruses [1]. Those lessons have been carried forward.

>> An autopsy and subsequent studies indicated that his death was caused by a fulminant immune reaction (with high serum levels of the cytokines interleukin-6 and interleukin-10) to the adenoviral vector.

>> The data suggested that the high dose of Ad [adenoviral] vector, delivered by infusion directly to the liver, quickly saturated available receptors ... within that organ and then spilled into the circulatory and other organ systems including the bone marrow, thus inducing the systemic immune response.

He was injected with >3 × 10^13 viruses [2]. The typical J&J dose contain: low-dose (5x10^10 viral particles) or high-dose (1x10^11 viral particles) [3].

[1] https://www.uab.edu/ccts/images/steinbrook_Gelsinger_-_Oxfor...

[2] https://www.cell.com/molecular-therapy-family/molecular-ther...

[3] https://www.jwatch.org/na53085/2021/01/26/adenovirus-vectore...

How would you test the extent of the mRNA's/DNA's impact within the cell?

Is it possible to do something like tagging the molecules with radioisotopes and following their path?

Here's an example: https://www.nejm.org/doi/full/10.1056/NEJM199001253220403

I didn't want to put this into the parent comment because I didn't want to get just shoved into the "vax" vs. "anti-vax" bucket by the replies.

But there's a very well known case where DNA delivered via an adenovirus killed a teenager during a genetic engineering study: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC81135/

>> "No one realized that the vector itself might pose a risk"

I'm sure the dosage, type of adenovirus, and modifications to the adenovirus are different. But there are obviously still risks we don't know about.

Loss of neurons and cardiac muscle cells is permanent. Emergency medical personnel are usually taught "time is brain" and "time is heart" for this reason.

Some body cells can bounce back after serious trauma, liver cells being a prime example: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2701258/

Both the AstraZeneca and J&J vaccines use an adenovirus to deliver DNA instead of mRNA wrapped in lipid (like Moderna & Pfizer).

Everywhere I read about the J&J vaccine, I see something like "the DNA vaccine doesn't alter your DNA". Can somebody please clear this up?

As far as I understand, the mRNA just stays in the cytoplasm of the cell and gets used up by the ribosome to create spike proteins. The adenovirus vector used in the J&J (and other vaccines) injects DNA in the cell's nucleus, which seems at odds with the widely circulated "it doesn't change your DNA" statement.

Do people make this claim because the cell displaying spike proteins is basically always eliminated by CD8 killer T cells?

Btw here's a nice high-level summary by the NYT about how all the vaccines work: https://www.nytimes.com/interactive/2021/health/how-covid-19...

Not the author, just thought this was a cool collection of simulations to display physics concepts.

Recently used this one [1] to explain far/nearsightedness

[1] https://ophysics.com/l16.html

This is the data/methodology behind NYC’s official COVID statistics, such as: https://www1.nyc.gov/site/doh/covid/covid-19-data-totals.pag...

1) How similar is this to creating a resnet? [1]. What are some key differences and similarities?

2) Has a CNN version of this been implemented in PyTorch?

[1] https://arxiv.org/pdf/1512.03385.pdf (Figure 2)

Also from the Lancet link:

* VRC01 was shown around a decade ago to be one of several antibodies generated...that achieves broad neutralisation of several HIV strains

* 97% of participants who received an HIV vaccine immunogen candidate developed VRC01-class IgG B cells, precursors to broadly neutralising VRC01-class antibodies

* Results from another set of studies presented at HIVR4P (the Antibody Mediated Prevention [AMP] trials) showed that although intravenous administration of the VRC01 antibody at 8-week intervals did prevent infections with some strains of HIV, only 30% of the strains circulating in the trial regions of sub-Saharan Africa, South America, Switzerland, and the USA were sensitive to VRC01

To me this appears to be a great breakthrough but doesn't seem like a silver bullet.

>> The biota formed with the Great Oxidation Event, a temporary increase in atmospheric oxygen, and became extinct from marine anoxia when the event was terminated by the drop in oxygen levels of the Lomagundi Excursion Event. The biota represents the earliest known experiment in multicellularity, with no extant multicellular descendants.

This is a great article and I'm happy the author mentions "an explanation of combustion exothermicity in terms of Pauling electronegativities is not convincing". I was incorrectly taught this view and it took me some time to unlearn it.

I keep seeing comments about how certain science topics are initially presented in an overly complicated fashion. But there are two forces at play here: correctness versus accessibility. The theory presented in this paper "predicts most heats of combustion with an error of only a few percent". This is good enough for most practical applications, especially if your goal is to introduce students to this topic.

But this is not the most "correct" description of reality that we currently have. A better model to decrease the error would incorporate information about the 3D structures/conformations of the molecules and some funky terms related to the quantum mechanics.

Anybody publishing literature about/teaching something like organic chemistry is stuck between this balancing act of correctness versus accessibility. This difficulty is compounded by the fact that you basically have to unlearn some of the stuff you picked up during the path to "accessibility" in order to properly move into the "correctness" phase. I genuinely sympathize with anybody dealing with the balancing act.

After studying organic chemistry for an extended period of time, it dawned on me that the well-thought-out explanations in my textbooks were just post-hoc rationalizations the field uses to avoid delving into the true quantum mechanical nature of the reactions. I'm happy I sacrificed some correctness for the huge amount of accessibility I got. But I'm also happy to unlearn some of the stuff on my path to correctness.

> There's a lot more to scientific computing than wrangling tabular data.

Also a point that gets ignored way too often. My original post differentiated between time spent writing models and time spent data wrangling.

I would never even attempt to write a symplectic integrator in base R (OK maybe Rcpp would be fine but that's not really "R"). Julia, by design, is better at that. But the R ecosystem is so good that I can use the best practical implementation of a symplectic integrator to solve common modeling problems via RStan.

Yes, Stan is a standalone framework that can be accessed from Julia as well. But the following workflow can be done in R much easier:

  1) Read in badly formatted CSV data
  2) Wrangle the data into a useable form
  3) Do some basic exploratory analysis (including plots)
  4) Write several models in brms/raw Stan (via rstan)
  5) Simulate from the priors and reset them to more sensible values
  6) Run the model over the data to generate the posterior
  7) Plot/run posterior predictive checks, counterfactual analysis, outlier analysis (PSIS or WAIC), etc.

Again, the above represents my common use case. I fully appreciate that people use Julia to do awesome stuff like "the exploration of chaos and nonlinear dynamics." [0]. I understand that the modern R ecosystem isn't really built for this.

[0] https://juliadynamics.github.io/DynamicalSystems.jl/latest/

I think that is absolutely a fair criticism. Personally, I rarely run into an issue where I absolutely am bottlenecked by a slow loop. But this sort of thing drew me to Julia in the first place.

There was also an R update in ~2017 that introduced some JIT speed-ups for loops, which made a noticeable difference.

If this is a problem you run into often, I suggest converting your object to a data.table. You can pass a function row-wise over the object very quickly:

https://stackoverflow.com/questions/25431307/r-data-table-ap...

Here's a neat website that captures this: https://www.r-graph-gallery.com/

If you click into any of the plots and scroll down you can see how little code is needed for most of these plots.

For example: https://www.r-graph-gallery.com/135-stacked-density-graph.ht...

I haven't heard of queryverse, thank you for that. This also brings up a good point I wanted to highlight.

I get that Julia is a young language with a growing ecosystem. But the lack of "one obvious way to do something" may scare new users away.

"I want to quickly wrangle data. Do I use Query.jl, DataFramesMeta.jl, SplitApplyCombine.jl or something else?"

"I need pipes to help me wrangle data more efficiently do I use Base Julia, Chain.jl, Pipe.jl, or Lazy.jl?"

For a new R user it seems so much simpler:

1. run "library(dplyr)" 2. Google "how to XYZ in dplyr" 3. ??? 4. Profit

I've been using R nonstop for pretty much 5+ years. I'm happy that there's established competition coming from Python and new competition coming from Julia. Having these languages compete over similar types of programmers pushes each one to be better, which is awesome. I'm not a die-hard R person, I'd be more than happy to switch under the right circumstances.

But...I think one thing gets overlooked way too often. For "data scientists" or "statisticians" or [insert new term here], the majority our non-modeling time is spent on just plain old data wrangling. To me, R is unbeatable here. I've tried Python ~2 years ago and pre-1.0 Julia.

Using tidyverse you can do pretty much anything to any dataset, often *without a monstrous amount of keystrokes*. (The pipe syntax is awesome). If you really need speed you can always switch over to data.table for uglier but faster code. I really tried but I could never replicate the "brain cycles to keystrokes" speed of R in Python/Julia. That is, being able to intuitively and quickly just convert my thoughts into readable data wrangling code.

Sure the base R language is not that "fast" and Julia/Python benchmarks are way faster. But in practice this doesn't matter to me. Most of the performance sensitive packages are written in C/C++/Fortran anyway (rstan, brms, glmnet, caret). I don't care that I could write 3x faster loops. The extra 5 seconds for that one piece of code doesn't make up for the absence of a good data wrangling ecosystem.

My message to the Julia team: You can get a very large portion of the R userbase to switch over if you focus on a Julia version of the tidyverse (especially dplyr). I know that DataFrames.jl exists but it just doesn't even come close. There's a difference between "you can do this in Julia too" and "here's a clean/intuitive way to do this better without extra baggage".

I'm sorry if the above seems harsh. I genuinely appreciate the Julia team's efforts. I can only imagine how hard it is to create a new language. I just wanted to be honest.

Can somebody please explain the challenges associated with miniaturizing and speeding up the ELISA test?

On wiki [0] I see:

> In the most simple form of an ELISA, antigens from the sample to be tested are attached to a surface. Then, a matching antibody is applied over the surface so it can bind the antigen. This antibody is linked to an enzyme and then any unbound antibodies are removed. In the final step, a substance containing the enzyme's substrate is added. If there was binding the subsequent reaction produces a detectable signal, most commonly a color change.

What are the pain points in this process?

[0] https://en.wikipedia.org/wiki/ELISA

I'm on a path to becoming a MD. Was in finance before and discovered HN probably 4 years ago.

I can't speak for the original poster but I think HN users are attracted to complexity and raw truth. Medicine and programming have a lot in common in that regard.

An example is the discovery and isolation of insulin [0]. Banting barely convinced somebody to give him lab space for 2 measly months. He then experimented with tying off the ducts of dog pancreases or removing the pancreases altogether. He realized he could keep a severely diabetic dog alive with injections from another dog's pancreatic juices. There was some drama around the subsequent purification of insulin and the Nobel Prize.

The full story almost sounds like software hacking and startup drama.

[0] https://www.sciencehistory.org/historical-profile/frederick-...

I stumbled upon this article after reading some more recent commentary on MSG [0], in which the author wrote:

"There’s no evidence to substantiate the claim that MSG causes ill effects in most people who consume it. A minority of people are hypersensitive to glutamate and MSG in food—added or natural—and in a study where MSG was given at 3 grams, in the absence of food, sensitive individuals had short-term, transient adverse reactions."

[0] https://peterattiamd.com/should-we-still-be-worried-about-ms...

Here's a great video explaining the National Popular Vote Interstate Compact:

https://www.youtube.com/watch?v=tUX-frlNBJY&t=122s

Why is this post flagged? Does this not adhere to any guidelines? Genuinely curious.