ImageXav

Here's another one that I feel is often overlooked by traditional A/B testers: if you have multiple changes, don't simply test them independently. Learn about fractional factorial experiments and interactions, and design your experiment accordingly. You'll get a much more relevant result.

My impression is that companies like to add/test a lot of features separately - and individually these features are good, but together they form complex clutter and end up being a net negative.

I don't really see why a leap to indentured servitude is necessary here. There are obviously many other ways of implementing a training relationship - see the other comments re. how it is done in Germany.

The point I'm making is that society as a whole is worse off if companies are too risk averse to hire and train juniors up to a certain level of quality. This results in too few people capable of doing a specific, presumably valuable, job well. The consequences of this are a society that underperforms it's true potential in the long run. In monetary terms, as you raise the question, that means that less taxable value is generated. So society as a whole pays the consequence.

The most interesting point in this article for me lies towards the end: "the role of the Guild was not to form rules, mores, regulations, and laws with respect to their crafts; their role was to introduce a system of art or craft to a new individual, to instill in them the idea of standards, quality, consistency, and perfection".

A common complaint nowadays is that it is very difficult for juniors with no experience to get hired, unless they have a degree from a prestigious university, and even then that's not often a guarantee. It seems that companies are more averse than guilds to take the risk of training someone up to industry standards.

I believe that it would be very beneficial to society to create schemes that encourage learning with a similar system. Mentors can sometimes accomplish this role, but that relation is far more informal.

As a not so active user, this tool is rather inaccurate. It seems to have focussed on the one question I asked about jpeg xl, which is the topic I know the least about.

I suspect a bias towards more common topics might be occurring.

A more niche but nonetheless interesting method that I was hoping to see discussed was magnetism. Gravitational waves are expected to decay into photons in intense magnetic fields. Or so I was told by one of my physics professors back in the day. I did understand the math somewhat back then, but it is beyond me now. It does however seem as though some people are still exploring this avenue [0].

[0] https://indico.cern.ch/event/1074510/contributions/4519384/a....

I think the best insult I ever read actually related to him, describing him as a bloviating buffoon, making for a very pleasant alliteration.

Ahhh I see, that makes sense. Thank you for clarifying I appreciate it. I made the mistake of assuming that the paper in the documentation was the paper of interest. I will take the time to properly delve in further once the paper is released, do you have any idea when that might be? In the mean time I look forward to giving testing the method on some toy examples I have.

So, this looks really interesting and I look forward to delving into the methodology in order to understand the algorithm better. However, what I immediately noticed from the paper linked in the documentation was that linearboost has a worse F1 score on average than the mentioned classifiers. Where it shines is energy consumption. Would it be possible to edit the title to reflect this? It's a huge gain in energy efficiency for a relatively small F1 loss, so kudos for that, but I think people might be expecting something a bit different from the title.

I think it's important to point out for people that might be interested in this comment that a few things are wrong.

1. Standard JPEG compression uses the Discrete Cosine Transform, not the Fourier Transform.

2. It is easy to be dismissive of any technology by saying that it is 'just' X with Y, Z, etc on top

3. Vision transformers allow for much longer range context - the magic comes in part from the ability to relate between patches, as well as the learned features, which JPEG does not do.

I recently discovered that Feynman had given lectures on computation. I thought this might interest hacker news, as they are well written, in a conversational style.

Most importantly, I believe that chapter 5 on reversible computation and the thermodynamics of computing presents the spirit of the work done by Feynman while he was at thinking machines. This addresses a topic which often comes up on hacker news: how did Feynman use differential equations to describe a binary system?

If anyone is on the fence about reading this, or worried about their ability to comprehend the content, I would tell you to go ahead and give it a chance. Shannon's writing is remarkably lucid and transparent. The jargon is minimal, and his exposition is fantastic.

As many other commentators has mentioned, it is impressive that such an approachable paper would lay the foundations for a whole field. I actually find that many subsequent textbooks seem to obfuscate the simplicity of the idea of entropy.

Two examples from the paper really stuck with me. In one, he discusses the importance of spaces for encoding language, something which I had never really considered before. In the second, he discusses how it is the redundancy of language that allows for crosswords, and that a less redundant language would make it harder to design these (unless we started making them 3D!). It made me think more deeply about communication as a whole.

Edit: the comment below refers to Gemini, not Gemma. As such the first paragraph is largely irrelevant, and only the second one applies.

To me, it feels as though the boat has been missed somewhat. The restrictions on Gemini make it unhelpful, but more than that, Claude 3 has really blown me away with its code suggestions. It's performing better than Mistral Large, GPT4 and Gemma in my tests, especially for large bits of code. It also returns the whole hog with changes, making it much easier to plug and play. Astonishingly, it also manages to combine ideas much better than any other LLM I've seen to date.

I suspect these fixes and the knowledge gained will be helpful to the community however, and will help improve the next iteration of models.

Thank you for sharing. This gives me a good idea of where to start looking.

Maybe someone here will know of a website that describes each step of the jpeg xl format in detail? Unlike for traditional jpeg, I have found it hard to find a document providing clear instructions on the relevant steps, which is a shame as there are clearly tons of interesting innovations that have been compiled together to make this happen, and I'm sure the individual components are useful in their own right!

This is all great advice. One thing I would add to this is to deliberately steer your team to avoid making big decisions on calls or in meetings. Instead, make it so that your team prioritises asynchronous communication methods to discuss the lay of the land, and only make decisions after everyone has had time to contribute to the discussion.

I've found that creating a shared document or flowchart can work wonders if key team members engage and build upon it. And once everyone has said their share you can then have a meeting to discuss how to progress. I've found this method to work well as you can take your time to reply to suggestions and comments and research them better. It also removes and element of emotionality from the decision making: everyone can see the suggestions and counter points, but the conversation is often less defensive and more considered as people have time to second guess themselves. So by the time you hold the meeting the benefits and drawbacks of the contending options in meeting your goals are clearer.

That is correct, my bad. The next screen after a test like this would likely be a mammography, and only after that would a biopsy be done if anything suspicious was seen.

It's a bit difficult to say, isn't it? The headline is using the term accuracy correctly, the reader might be ascribing the meaning you are to it, especially if they are non technical. As was the parent comment.

My goal in pointing out the difference was not to be snarky. It was to point out the very real statistical consequences. Any model can be accurate on a sufficiently biased dataset, but what matters once a screening test hits the real world are the precision (positive predictive value) and negative predictive value. These are the hurdles that the test will have to pass to see widespread adoption.

That's right, that's typically what is meant by screening purposes [0], apologies if it wasn't clear.

[0] https://www.nhs.uk/conditions/nhs-screening/

A device such as this would never replace an MRI scan. The information provided is for screening purposes, at best.

Also, diagnosis would typically be done using a mammography. The cost of such a scan is lower - around $100[0].

[0]https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4142190/

Not at all. The model is quite accurate. In fact, with the distribution of samples that they have a model that predicts all cases as having cancer would also be very accurate. It would get 17/21 predictions right. The model lacks precision. I suspect that even with a fairly high cut off point the model would still produce a bevvy of false positive predictions due to that. It might still be useful as a screening step if they can increase the sensitivity further, but you would still rely upon further tests to get a true diagnosis.

I was ridiculously excited when I first read about this in October (if I remember correctly) last year, when a few of the first results were beginning to pop out. I found the methodology fascinating. First of all the digital unwrapping of the scrolls, then the recognition that crackling in the paper was the sign of ink, and finally putting together a model to detect it, piece by piece. I need to look into the final repository to understand what exactly they did, but they seem to have used a TimeSFormer. I'm confused by this choice as I thought it was for video. How did they apply this to images? In the end though, what a wonderful day for archeology. These young minds deserve a huge round of applause for what they have achieved.

Boeing's competitors seem to be doing just fine. Rather, this is a clear case of poor management leading to a degraded engineering culture.

Like with many such projects, it's very helpful if you use DataFrames in isolation, but it lacks support from the wider scientific ecosystem. I've found that using polar will often break other common data scientific packages such as scikit-learn. This unfortunately often makes it impractical in the wild.

This is the most accurate answer so far re. The scaling laws. It has been demonstrated that LLMs follow quite clear power laws with respect to performance. In fact, the performance of any model can be determined from the number of parameters it has and the amount of data it is given. The Wikipedia article on Neural Scaling laws provides a brief, accessible, summary of this. Both data and parameters are expected to increase in coming years, so models are expected to improve.

This is an extremely useful design. I never thought I'd get excited about something like this but I recently moved to an old house with chimney flues running through every room, resulting in shallow alcoves that are of differing sizes.

A fitted wardrobe would be pricey (~$4000), barely fit depthwise and half of the space would just be covering the flue if we wanted it wall to wall.

On the other hand I can just buy 3/4 of these for each wall and still come out ahead. The grooves in the rods are well thought out and mean that I can saw them to size myself, as can anyone.

The only drawback I can think of is having clothes up directly against a wall. This could obstruct airflow and be an incentive for mould to grow.

Might be time for my first ever pledge.

This is an interesting one. It's on the top of a lot of lists but I remember some criticisms of it in the early editions. Are the later ones updated/corrected?

I can imagine it to be quite hard. However, I do think that it is possible to do studies that meet the 'good enough' criteria and that certain key points can be ascertained with high confidence (e.g. taking folic acid at the beginning of a pregnancy/before conception - don't drink alcohol, etc). The issue I'm finding is that most books for parents seem to be a collection of anecdote and personal bias. It would be nice to have a book which provides references and true insight.

This is interesting to me. Do you use a library to estimate the fourier series of a data series or have you implemented it from scratch? I've searched for this in the past but always got results RE. Fourier transforms, not series.

This looks really interesting and I hadn't really thought of this as a use case before. Does anyone here know if a similar product is available as an extension for VS Code? I know that I would love to be able to tell VS to just comment all of my functions for me and ensure that the typing is correct.

You could also leave the company at any point, so why should they hire you?

There's got to be some give and take in a relationship with an employer. If you're feeling exploited, you're not in the right role. And that will show in the quality of your work. Work can be fun and interesting, and someone who enjoys what they do will relish the opportunity to learn. I know that for me interviewing, though tough, has always been an opportunity to solidify my basic understanding of the field I'm in. Maybe figure out what you enjoy, and find an employer who respects you equally?