Karsten Konrad 🇪🇺
banner
zabong69.bsky.social
Karsten Konrad 🇪🇺
@zabong69.bsky.social
840 followers 2K following 2.1K posts
Armchair Data Scientist and nerd. PhD in automated theorem proving just before AI became a subfield of Linear Algebra. Machine Learning, statistics, data visualisation, biology, recreational maths, computation, old pocket calculators, and video games
Posts Media Videos Starter Packs
Reposted by Karsten Konrad 🇪🇺
The only thing more annoying than the people who lionize "AI" at this point are the idiots that pretend to know everything about LLMs and machine learning and claim it is all useless. If your full idea of the thing is just chat gpt, you haven't even tried to learn anything.
Reposted by Karsten Konrad 🇪🇺
I think there’s only one honest position: machine learning/generative AI have many applications. Some are mature; most still need work. In some domains, certain applications are potentially transformative; in many domains there may be helpful niche applications; in many domains it will be useless.
There are basically only two positions in the debate about AI.
1. I’ve barely invested any time in learning how to use it effectively. AI sucks.
2. I’ve invested in learning how to use this tool. Holy cow, it’s transformational.

Position #1 has lower barriers to entry.
Reposted by Karsten Konrad 🇪🇺
regardless of credentials, anyone stating epistemic certainty about the internal mechanics of 1 trillion floating point numbers produced by gradient descent beyond these sorts of simple heuristics without evidence is simply not reflecting the current state of the science
Reposted by Karsten Konrad 🇪🇺
In machine learning, do you need to know any optimization algorithm other than stochastic gradient descent? A reluctant but best-faith argument for no.
Highly optimized optimizers
Justifying a laser focus on stochastic gradient methods.
www.argmin.net
I keep thinking: Who are the contractors doing this? Who in his right mind goes "yep, I am going to put a backhoe to the side of one of the most famous historcial buildings on this planet on behest of a man famous for stiffing contractors"?
Not available here in Germany. Arrrgh.
Reposted by Karsten Konrad 🇪🇺
Stop scrolling and post two characters who bring you happiness 😀
I wonder whether Volvo‘s excavator division is happy about this kind of free media exposure.
Wake up, babe, the least subtle visual metaphor in history just dropped
Looking at this picture long enough makes you want to smoke a Galloise without filter and cite Camus.
This photo accompanying a news story about the heist at the Louvre is perfection.
Thanks for the recommendation, reading right now.
Reposted by Karsten Konrad 🇪🇺
Ludicrous take to say nobody misses the video store. Going to my local video store as a teenager was so fucking good. They had a five movies for five days for five dollars deal that was pretty much the greatest thing ever and a super stoned clerk who had seen every movie known to man
Reposted by Karsten Konrad 🇪🇺
Fact: The Dutch are the tallest race of people on earth. Which gives them a distinct advantage when it comes to aircraft maintenance.
I understand residual as the distance to the regression hyperplane on the axis of the dependent variable. It’s in this interpretation always one-dimensional.
Reposted by Karsten Konrad 🇪🇺
LLMs are trained to mimic a “true” distribution—their reducing cross-entropy then confirms they get closer to this target while training. Do similar models approach this target distribution in similar ways, though? 🤔 Not really! Our new paper studies this, finding 4-convergence phases in training 🧵
Hinrichtung ("forward direction") und Rückrichtung ("backward direction") in German. Since "Hinrichtung" also means execution (as in chopping one's head off), this direction should always be the harder one.
Reposted by Karsten Konrad 🇪🇺
The AI discourse here is bonkers.

It's quite common for people to believe that even if an application is ethical and useful it should be rejected because the body of research informing it is tainted with original sin.
Reposted by Karsten Konrad 🇪🇺
I work in Health AI and a lot of the cool use cases don't require a lot of compute. They combine LLMs with rules based systems, machine learning and human expertise.

But this involves labour and nicheing down, the hyperscalers are allergic to that because it doesn't lend itself to a monopoly.
Reposted by Karsten Konrad 🇪🇺
An anxiety celebration, or celebration with anxiety/angst, would be a „Besorgnisfeier“ in German. Works as a verb too: „Ich besorgnisfeiere den Führerschein meiner Tochter sehr.“
Reposted by Karsten Konrad 🇪🇺
Wondering what kind of conversations are happening at inflatable costume companies these days
Reposted by Karsten Konrad 🇪🇺
the US government has arrived on bluesky
For a very small fee I am prepared to make up new long German composite nouns when you need them.
Daughter passed driving test! There must be a word for a celebration coupled with anxiety!