lcrmorin

No. Thanks for pointing this out. It took quite some time to understand and express this to my management. Also I use a very small set of libraries, with function names / parameters that are quite explicit. And I don't need to specify some style to bypass hidden prompts, and bad coding practices.

Do you change regularisation ?

Company use AI to find critical bugs in code ?

Perfect exemple of what op is asking. This is just a demo. What problems does it solve for you or your clients so that you make money ?

Thanks for sharing this. Of course I will closely watch it because claiming to beat gbdts might be a bit early.

- It is not entirely clear how the datasets split is done. Do you make sure that the model is evaluated on unseen data ? More generally how does one knows whether a dataset was part of the training or not ?

- You mention some serious limitations (10k rows, 500 cols.). It seems a bit weird to have fixed numbers. Can these numbers be roughly balanced ? (eg. 1M rows, 5 columns ... ). Does these numbers scale with memory ? (what memory was used for the 10k rows / 500 cols figure ?)

Some things that worked for me this year:

- Update your resume / cover letter to make it more appetizing. Don't hesitate to use ChatGPT. Change the title to match the job. You usually have to cater for HR first, not technical people. Also remove the stuff that you don't want to work with. This will remove the tech you don't want to work with anymore.

- Emphasize anything that could stand out (side projects, anything to do with hot topics / techs). Again you have to catter to HR first.

- Don't get afraid of the numbers (out of the 500 applications you mention a good 450 will be thrown into the garbage immediatly). But try to get early.

- Often the first steps of the process require some customs texts (mails, achievements, why I want to works with you). LLMs can save you time and make your answers more appealing.

- Specialized Recruiting firms can be you friend. Sometimes you need to be blunt with them. Discuss expectation early. ("I am currently making X with Y days remote. Would you be able to find something that match my skill and offer X+20% and at least Y+1 days remote ?")

I've found that the people that build AI tools and the people that use AI tools seems to be quite disconnected. Phind seems to be a notable exception, bringing some value over google. I am not entirely sure why it works better than goole tho. Is that because the model behind is better ? or just you figured that removing the usual trash (adds / over engineered seo) ? But i'll take it.

Now I am just concerned about the privacy of what I share (can it be worse than google ?) and the optimality of the code ? (code usually works, but could it work faster ? what happen when new libraries get out ?)

Yeah usually market provide different time horizons. I fail to see what is the substance of this article ...

Can't believe I had to scroll that much to see this comment. Upvoted and Thanks.

I don't like those lists... too much things. No idea what is actually usable / usefull in there.

As Kaggle is a leading place for discussing / building and competing in ML, I found it interesting what stance they are taking on using AI generated text. I don't know if this will be enough. This post has been generated by a human.

Maybe they could have used a LLM to expend their message a bit

It is what is weird to me. Those leetcode problems seems weirdly designed, some even seems counterproductive. I've juste been offered an interview about easy-medium level leetcode questions. I went to the site and ... the easy questions seems to be fundamental questions that no one deal with in real life. Medium are some common practical problems that were implemented and optimised in standard libraries a long time ago. Hard problems are actually fun to deal with and probably more revealing about myself. What am I supposed to do ? ask for difficult problems only ?

Yup similar experience here. Field and experience should be taken into account when offering leetcode interview. Maybe langage should be added too. In python for exemple there are lot of standard library that implement and optimise the leet code stuff.

> (though salary isn't posted in the ad)

Might be the problem ...

> the work being a higher percentage non-code is what's causing us trouble

Might be another problem... are you recruiting a PM or a Software Eng ?

> - Take-home technical assignment (~4h) or similar at candidate's choosing

I've personnaly stopped doing those. Either they are very far off the job either too revealing. And generally taking way too much time.

The tweets / medium article are a bit incomplete. But they mean at least: - Authors didn't make decent efforts to build a robust baseline. Worse it seems that they have purposefully built a bad baseline to make their solution look better. - Authors didn't really disclose full performance. Statistical performance is one thing but time complexity is another. Hopular needs 10 mins on a 500 rows datasets. That's a NO NO for any serious application. - Authors didn't provide an easy to use interface. You can't really claim SoTA on small tabular data with something that isn't testable by everyone.

Hey a bit late to the party (HN newsletter crowd). This really seems like something my BigCorp could use. I am on holiday RN, so I won't fire my computer to try it. But I was wondering, does it allows easy copy pasting the table into standard MS documents (work ? outlook mails ?).

The whole thing read like a humble brag... "As a powerlifter, I workout 4-6 days per week", " I would have the third-highest new offer on Levels out of over 200 offers on their site" .. .etc.

Depends if the offer is made in good faith. If not, this would really look like market manipulation.

It was 30 when the price was at 33 and musk wasn't on the picture. Someone's buy nearly 10% on the open market then bidding for a significant chunk of the rest is new info.

I've jsut discovered this approach with python. I usually used google and cookbooks. But I often found myself in position where I tought 'surely there must be a better way to do it'. Reading the standard library really helped me understand how to do some stuff. Best way I found to enforce what I learnt was to rewrite some code with minimal imports.

I work for the government so really know this feeling. I have a funny exemple in mind. Months ago I got pulled into a meeting because 'I know data'. The phrasing alone was a red flag. Lots of high level people in the meeting. Like 10 people, each one payed at least twice as much as me. The goal is to transfer data that is too big for our 'standard pipelines'. They are discussing building a whole new pipeline, contractors would be involved and all. I asked 'stupid questions': "- How big is the data we are discussing ? - '200 Go' - is it recurring ? - No - is it sensitive ? - No - Why don't we buy a 50$ 1to hard drive and transfer the data manually ?"

Last question was followed by the longest silence I ever heard in a meeting... I wasn't invited the next meetings. I heard it took them more than 5 other meetings to reconsider my solution. An intern has finally been sent to buy the hard drive last week.

Difficult to choose one this year (plenty of things happened + plenty of time to read things due to lock down).

"Equality of Opportunity in Supervised Learning" (https://arxiv.org/abs/1610.02413)

It explain the basic concept about fairness in ML. Very practical exemple in my domain knowledge that show the trade-off between fairness of an algo and overall performance (money). Really make you see what may go wrong with bias in ML. It shows, in my opinion, why we will have to regulate ML as corporation aren't really incentivized to deal with fairness. It also shows that there is different notions of fairness. So there will always be something that feel unfair and also doing something can always be interpreted as positive discrimination.

Just currious, do you know if that help build personnal credit score ?

There is a big difference between dumping a dataset in the latest hot ML model and building something that offer some actual explainability, deal with intrinsic biases, have some stability over time and recalibration and that will go trough an unprecedented crisis for which you don't have any data to learn from. That mean the model usually has to go trough a lot of internal commities and different external agencies. I highly suspect that an external proprietary solution won't go very far in the credit rating field.

Well... most quants are recruited straight out of engineering schools so you are probably right. About the IDE, honestly once you think about it, it actually achieve many thing at once. I am not sure what they could change to make it better.

From the graph it is unclear to me which one is following the other. I wouldn't be surprised that that the WSB sentiment just reflect the eprformance of the market the previous day.

He bought calls AFTER the surge ... they call that a 'retarded' move. On the other side of the spectrum there was a good GME DD that was predicting a pop one hour before the Microsoft Deal.

It is, but mostly due to their selection process, basically removing poorly performing companies and adding sane one that are about to grow.