kla-s

I wonder how routinely wearing compression socks interacts with this proposed mechanism.

Always thought it interesting that for certain diagnosis wearing them until end of life is indicated according to current guidelines and we don’t expect maladoption (weakening of the connective tissue). Id kind of prefer/would be interested in a workout based approach, also interesting what knock on effects such workout would have on the larger system

Id also add that 5) We need some sense of truth.

Im not quite sure if the current paradigm of LLMs are robust enough given the recent Anthropic Paper about the effect of data quality or rather the lack thereof, that a small bad sample can poison the well and that this doesn’t get better with more data. Especially in conjunction with 4) some sense of truth becomes crucial in my eyes (Question in my eyes is how does this work? Something verifiable and understandable like lean would be great but how does this work with more fuzzy topics…).

Stiction

Do you care to explain more, id be interested :)

I guess bioelectricity. Michael Levin is the guy to look at afaik.

Thanks @missedthecue, i wasn't aware this was productized and tested (though hoped it was)

Any serious ideas for a recently removed glioblastom?

Whats your view on Cum-Ex?

And maybe as a Bonus what do you make of the smaller (relative) taxrate the bigger fish (companies/wealthier individuals) pay?

Location: Munich, Germany

Remote: Yes

Willing to relocate: Yes

Technologies: Python (PyTorch, TensorFlow), Reinforcement Learning, Medical Imaging, Machine Unlearning, Physics-informed ML, Quantum Computing, Protein Structure Prediction, NLP, Java, SQL

Résumé/CV: https://drive.google.com/file/d/1b9UfeqFvEu0t4sTGNd8yiJ1m60x...

Email: essence_mallard.5a@icloud.com

Hi, Kilian here :) ML engineer / research-minded generalist with a recent M.Sc. in Informatics from TUM, including a thesis on machine unlearning in medical imaging (@Harvard Medical School). I’ve worked on problems like reinforcement learning for tumor landmark detection in MRI, and deep learning for physical systems and protein structure prediction.

I like forming own ideas and following them through — from literature review to implementation, evaluation, and iteration. I can reproduce papers, fine-tune models, explore new methods, and design experiments that actually test hypotheses. I’m especially motivated by early-stage work that blends research thinking with practical engineering.

In team settings, I tend to gravitate toward coordination and planning roles, and outside of work, I’ve led a 300+ member volleyball department for several years — which taught me how to manage people, not just models.

I'm looking for roles at the intersection of research and engineering, eg early-stage AI projects where curiosity, ownership, technical depth matter and impact-driven development are valued.

Wow that video is awesome, thanks for sharing

To add some nuance, unused/unnecessary sugar will cause diabetes. If you are in high cardiovascular load consuming more than you can take in as eg with long distance triathlon, to my knowledge you will not increase your diabetes risk. It's more a matching supply and demand thing.

Genuine question, how does multi step reasoning work for you then? Like eg if you have some math problem that's trivially to solve individually but needs multiple steps, lets say 16 * 3 + 5? How does 16 * 3 = 48 land in some 'register' of your brain (short term memory), so that you can then add 5 to get to 53? Maybe 16 * 3 + 5 is to easy for you and you'll just 'see' it but the question still stands, just choose a more complex problem.

Isn't the same meta process at play when thinking about more fuzzy topics?

Thank you, for making the effort of making this so accessible! Danke :)

Well then thats „Volksabstimmungen“ as Switzerland has it for you.

Well to play the devils advocat, i think this is useful to have, at least for ‘Open’Ai to start off from to apply QLora or similar approximations.

Bonus they could even do some self learning afterwards with the performance improvements DeepSeek just published and it might have more EQ and less hallucinations than starting from scratch…

ie the price might go down big time but there might be significant improvements down the line when starting from such a broad base

Actually its a little more nuanced:

Operation Type|Mamba Complexity|Transformer Complexity

Training(per iteration)|O(L)|O(L^2)

Autoregressive Inference(per step)|O(T)|O(L)

Memory Requirements|O(C)|O(L)

Where: L stands for the sequence length. T denotes a fixed constant that accounts for compression and selection time in Mamba's autoregressive inference. C reflects the fixed size of the SSM (State Space Model) latent state in Mamba

Per: https://github.com/state-spaces/mamba/issues/196

Does anyone know how long they have been at this? Is this mainly a reimplementation of the physical intelligence paper + the dual size/freq + the cooperative part?

Jamba 1.5 Large is 398B params (94B active) and weights are available.

https://arxiv.org/abs/2408.12570

Credit https://news.ycombinator.com/user?id=sanxiyn for making me aware

I dislike the cardan tunnel of the shared ICE/BEV platform if you look at eg the i4. It exacerbates the problem of all (aerodinammically efficient) BEVs with a low roof line: blood flow problems due to hip impingement is bound to happen due to the "head between knees" sitting position in the back.

My main gripe is no 800V, it would be a sane travel car with that.

No frunk to store the possibly wet and dirty ac cable. Yes they exist on the aftermarket but the car is a premium product, one shouldn't have to deal with this.

My personal EV tech goal/dream is 400km range in between 20-80% SOC, chargeable in <20 minutes. Ideally as a worst case ie cold outside -5C, 140kmh avg. And the whole package affordable: <50k new, ideally even <30k. Let's hope we'll see this kind of volume model sooner rather than later. My hope is that it could significantly improve the QOL for many people in cities worldwide and help further build out renewables by offering grid stabilization while not in use (this dream unfortunately requires ~100kwh, which will result in a >2t vehicle so tire rub will probably increase slightly, though i think its a worthwile trade off). So i guess my last gripe is price, yes its a premium product but the production complexity and price should go down vs BEVs...

Yeah its a mixed bag in germany in my eyes. All of the big brands have a real electric platform minus BMW. 800V isnt pioneered with them either.

VDA is pressuring for HVO100 bs. Mercedes just pulled back on the electric only strategy. I guess they figured that most of the added value is with the batteries and being deeply integrated there and thats not really their cup of tea, so far at least.

There where Taycans for 30k$ which was a real great deal if your the second buyer and not the first.

On the other hand the id family looks ok, facelift good. The a6 is pricey at 100k but is the first really capable german highway ev, not as premature as the eqs…

Since i couldn't find the time code in the video, i put a map together with both NOTAM and NOTMAR.

map: https://github.com/kla-s/Space/blob/main/Map_NOTMAR_NOTAM_Sp... description: https://github.com/kla-s/Space/tree/main

Lets hope this is the year of Linux desktop and i didn't violate any licenses or made big errors ;)

The real solution is the dynamization of electricity prices. This needs some adjusting from your average consumer but not a lot if done right. In Germany there are startups like 1.5C, Enpal etc which will sell you a heat pump, solar, ev charger pack with some "smarts", switch you over to a dynamic pricing electricity contract and then claim to optimize the overall cost (i have no direct experience of my own). If you are willing to take a small amount of temperature swing your house is a big thermal battery (even more so if you have a heat pump to water with a big, well insulated reservoir), your ev is a battery with vehicle to grid. With this you can shift your main loads a good amount. Washing machines and dryer as well as cooking/baking might be slightly more problematic/harder to shift, though the car battery should be more than enough for average evening cooking and i have seen washing machines/dryers which can take an external signal as to run when the price is low/there is excess electricity...

- auto lid open and close - heated seat - water jet to clean buttocks with subsequent warm air blower - relaxing music speaker (to drown out your defecating sounds and get you started)

just some ideas what some (public)/most (nice home) japanese toilets offer, which might be hard to come by anywhere else

Have you considered a tablet with pen to not cut down so many trees (transflective is the new cool kid on the block i hear, ipad works pretty well ime)? By now i actually prefer its better searchability with ocr and miss pinch to zoom on real paper ;) Or do you think that the paper being wobbly is important? I mean i get missing the indentability. My experience as a student is its more important to move your hands and have an intuitive sense for location on a sheet of paper (more intuitive/faster navigable than with a typed word processor) to easily form deep memories with what you are dealing with (+spaced repetition!!). So in total tablet works great (for me) once you get used to it. :)

Please indulge me on your short tangent on Austins public library, how can they improve? Same budget?

Thanks, missed that one.

For context gpt-4 is supposedly @ 1.8T params.

Please someone correct me if I’m wrong, but my understanding of ML/LLMs is that this kind of hand crafting has been tried, but it is easier to train/less finicky to let behavior like this emerge from more data, see [1] “Bitter Lesson” and [2] “Scaling Laws”.

MAMBA as an architecture claims to have some significant gains performance wise, but to my knowledge there haven't been any really large models (>~100B params) with open weights/leaked MAMBA architecture been disclosed other than this (7B).

As mentioned by other comments, another dimension not to forget is the training data quality. Not only quantity but also quality really matters, is what we are learning more and more with LLMs..

[1] http://www.incompleteideas.net/IncIdeas/BitterLesson.html [2] see eg https://m.youtube.com/watch?v=5eqRuVp65eY&pp=ygUMU2NhbGluZyB... for a well made/easily digestable intro

Machine Unlearning is a thing, see e.g. here [0] for a introduction.

[0] https://ai.stanford.edu/~kzliu/blog/unlearning

It was on HN, this comment: https://news.ycombinator.com/item?id=39746742

Hug of death is strong with this one currently, try: https://archive.ph/RBkiB

Same machine here as well, running the latest Ubuntu (23.10) power profiles do work but auto-cpufreq is way more effective in increasing battery runtime for me.

I can get ~6h of runtime (~10W draw) if i mostly just code in VSCode + some browsing, with display at 1/3 brightness and the code being executed remote. Went with the OLED panel, so that definitively doesn't help battery runtime, but damn those colors are nice :)

I wish they would fit in a bigger battery or do some more tuning so that some cores could actually turn off if they aren't used, id happily accept the spool up time for big jobs, if this could increase the battery runtime a few hours/allow for higher brightness. What kind of runtimes/draws are you getting?

Ferry* and maybe plane i understand, but especially the later is changing with many flights offering (complimentary) internet service afaik, but your aunt i don't. The alternative, to me normal setup would be using a workstation at the office and then remoting into it via SSH (behind a VPN, VSCode supports this great). This allows for a thinner client (eg Macbooks Air) which need less replacement and where battery runtime/portability can be weighed more. Sure now connectivity is a concern but to me that's mostly a non issue where i travel/the work spots i choose, but i guess it can become more of a problem in other (less developed) parts of the world...

* this will probably also change with naval Starlink becoming deployed more...