MichaelRazum

Nice! Although if you have new young team - things look different. Also, to be fair, agentic dev is something that might be just the right answer for the future.

And tbh it get things done very quickly. So, it is also very hard to argue, for a different coding style

yep, 50% chance imo

Actually not worried about unemployment. This is an awesome development thing - called technological progress.

PS: Compare Assembly with Python - for sure the ration is more then 10x. Still we need much more devs compared to early days. For me the question is what the future software dev looks like (if the job still exists).

Maybe try claude. Also people are orchestrating AI for example with ralph. I think it is possible to write pretty decent, test driven, code with AI

Agree, especially a review is always an knowledge update/exchange and for juniors a learning experience. If it is AI generated, its just not worth the time.

Not sure tbh. The labs which are creating the AI - definitely know what they are doing, and its incredible. Would just argue that the AI will become only better in the future

There seems to be two likely outcomes. First the value of education drops, since studying becomes much easier. Second, we will have few young genius level people, who were able to learn very quickly with help of AI.

How about elliptic curve cryptography then? I just think coming with a formula is not really understanding. Actually most often the “real” formula is the end step of understanding through derivation. ML does it up side down in this regard

Although is it really "understanding" or just able to write down the formulas...?

What would you want in his position?

Sure agree, for bi directional websocket communication it is the way to go. It's just that you have to really think it thorough when using it. Like using asyncio.sleep instead of sleep for example and there are more little things that could easily hurt the performance and advantages of it.

You can’t just plug and play it. As soon as you introduce async you need to have the runtime loop and so on. Basically the whole architecture needs to be redesigned

Counterargument. So far bigger have proven to be better in each domain of AI. Also (although hard to compare) the human brain seems at least an order of magnitude larger in the number of synapses.

This looks like a nice application to get inductive bias into the model. But I think right now there is no solution to get fine grained motor skills besides tele operating and doing behavior cloning. And even then it is far from perfect..

Yes

Alexnet, AlphaGo, ChatGPT Would argue he did strike gold few times.

How about Ilya

The claim about sample efficiency sounds a bit strange, since they did not include the state of the art sample efficient algorithms. Like dreamer or tdmpc. Also PPO is known to be not efficient, just compute efficient.

Isn't most the time consumption in the airport anyway? Like people supposed to be ther 2-3 hours ahead. If you could get 10m before the flight to the airport that would save so much more time.

Grok4 was trained on 100k or 200k GPUs (as far as I understand)

Grok5 might need 1MM or 2MM.

So the question is what about metas / zucks plans? How many GPUs will Manhattan get? Looks like, that to get the next unlock you need crazy amounts of compute.

Technical question: Can someone explain how the vision backbone can be replaced after training? I think this is what they mentioned in the video. Just wondering how it would work, since I would suspect that the visual embedings would be highly affected.

PS: Is the approach something like LORA or a complete retrain on the visual part?

TLDR, just use PPO? I always found it kind of confusing, that on paper SAC or other algorithms seem to be much more sample efficient - but in practice it looks, as the author mentioned that they often do not work.

PS: Not sure where to put algorithms like TD-MPC or DreamerV3, since they are kind of in between, right?

Looks interesting. Especially since it seems to be much cheaper that the closed solutions...

Have looked in this kind of systems, for my parents. The use case was basically, not about energy efficiency, but rather noise protection - to be able to sleep with a closed windows. I think so far I always had two issues (in that usecase).

- First the device by itself - produces a bit noise like 42db might be too much for some people if you want to sleep. Especially some of the devices are using one ventilator, which switches directions and won't produce homogeneous noise.

- Second 60 CFM is fine, but if you want to have the feeling of an open window - it should be much more and most devices can't deliver that. Also the heat exchange thing is kind of cool in the winter for sure. In the summer, you often have the case that in the evening you house is much warmer than the air outside - so you would like to turn the heat exchange off in the winter.

PS: Actually, maybe looking for a complete different use case. But I think what would be very cool, would be some idea to make at least one room 100% quite (with fresh air ) in a cheap way. Guess this would be a huge life changer for a lot of people, who suffer from noise pollution.

1) Is this really the case? I mean some of the popular open source project are kind of mature. So you would get some feedback, regarding some specifics of the frameworks. For sure you would learn a bit. On the other hand they are kind of "done" software, so you would not really learn how to build something new from scratch.

Not sure about the other points.

"just 2k GPUs" lol. Guess I'm gpu poor.

How do they solve the 2%? This is the question. If those problems were unseen, that might be already very impressive.

Wow interesting! Swiss have also one of the highst gpd per capita. And the university EthZ beats most of other European universities.

The thing you mentioned is not unique to Swiss Germany had it as well. But I think still you have there more people who study...

Not coming from Bioinformatics. But we had a similar question regarding time series modelling a while ago. We used kind of modified arima models and it was difficult to do it in python. For the rest we used python, especially for ML stuff and tried to keep the R part rather small

Tldr: R for classical stats, python for machine learning

Kind of have the same intuition. However, to play devil's advocate, the human data and knowledge far exceeds what a single human can know and learn - so it would make it already superhuman.