Kevin Mader

@janusch_patas Posting when the "to be published" code actually drops is potentially even more useful than the papers themselves 🤪

Code dropped: github.com/yanyan-li/GeoG…

@francoisfleuret I wonder if there are any good public datasets with written exams and grading to experiment with.

@francoisfleuret Have you tried to get some of these models to reproduce incorrect student answers on exams? It would be fascinating to see the overlap on specific problems.

IMO an instinctive error with LLMs (or with animals, and to a lesser point with humans) is to imagine that a certain ability can be used as a reference to estimate the others. When I show gpt4 helping me doing obnoxious but not-so-fundamentally-difficult maths 1/3

I’m calling for all artificial intelligence labs to stop publishing for at least six months, so I can catch up with the literature

@jburnmurdoch Are we going to see some nice charts of spotify vs strava data exports?

It’s wild how much impact music (and tempo of music) has on exercising. Switching from podcast to fairly bouncy song increased pace by 10%, then switching to high-tempo song increased it another 9%. Presume it’s combo of endorphins, adrenaline and then BPM dragging up cadence?

@janbromberger @lioninawhat @jamierusso Quite a few different elements: war/peace-time ceo roles, building great teams from imperfect pieces, firing and all the anecdotes to give everything credibility and context

@lioninawhat @jamierusso Lots of anecdotes, but what do you fund actionable about it?

Whats the #1 book for startup CEOs?

Google Calendar, but it shows the cost of the meeting

Async programming is so cool. With just a few minutes of work you can have a program that hangs forever and is impossible to debug

ChatGPT to me seems to weld an English professor’s writing skills to an encyclopedia’s knowledge base and a kindergartener’s reasoning ability, producing impressive and highly polished nonsense. The screen is on - at 8K 120Hz - but nobody’s at the keyboard.

@betatim Tell me you hate mypy without telling me

Friday "obscure Python stuff": how would you organise for a parent class to modify an instance after a method on a child class has been called? My solution: gist.github.com/betatim/411071…

@betatim And worth switching? Won't you miss the chaos here 🔥?

You can now also find me at @betatim" target="_blank" rel="nofollow noopener">mstdn.social/@betatim or @betatim@mstdn.social

@betatim In my experience, it's often a good thing because it serves as a warning that "hey it's not quite as simple as you think" I keep a large collection of declined PRs at the ready when someone asks why don't we just upgrade numpy or switch to py3.11

In a company setting I'd say most people would say this is a bad thing. Even here I'm not so sure. What do you think?

Open-source project feature or bug: something reported in 2015, got enough consensus that someone made a PR, ..., now in 2022 someone refreshed a N-th generation PR (PR taking over from a stalled PR) that is trying to close the idea from 2015, this PR was started in 2021.

@lizardbill wait isn't that how all data-scientists read constants from databases? who needs ORM when you have from pandas import *

Using Pandas read_sql_query function to load a single scalar value from the database into a dataframe. #python #pandas

@tunguz Wow this makes getting started with and debugging CUDA almost nice

“CUDA by Numba Examples” by Carlos Costa link.medium.com/f40fllPFytb

@betatim It's quite hard to reliably measure unless it's running on isolated dedicated hardware like GPU and even then the standard deviations can be enormous

Are there easy to use tools/conversion factors that let you benchmark how code performs in terms of power consumed instead of elapsed time?

@betatim In Switzerland?!? My working assumption was all the dangerous little animals were killed off long ago

Snake!

@getupnext No free tier or even a free trial? ❌

❌ Pocket ❌ Instapaper ❌ Safari reading list ✅ Upnext is a content organizer with superpowers! 💾 Collect any type of content in single library 🪄 Helps you actually get to your content ✍️ Make highlights and notes 🗣 Share what you discover

@David_Ouyang @mkaic_ @OpenAI @gdb So many options in the biomedical imaging rabbit hole labs.openai.com/s/O0oPsz4bqRmB…

Whew, glad to see our research program is safe for at least a few more months 😅.... #dalle2 has no concept for #EchoFirst. Prompt: “Deep Learning Interpretation of Echocardiograms” @mkaic_ @OpenAI @gdb