breakds

Engineer: but business is not my problem.

Almost every day, watching videos, having meetings, hands on (the agent) coding and playing games.

+1, ultrawide curved monitor user, used to be on sway but the experience of niri is next level.

In this situation, it means "Hey I have been here and observed this!"

Exactly what I had via sunshine and moonlight in my Vision Pro ...

The sample efficiency of the RL algorithm, even for simple games, is not very good. This usually means that we will need a lot of episodes for the policy to learn to excel. Being able to run policy in an environment that can parallel and accelerate could be very helpful for the improvement - for example running a batch of browsers or tabs simultaneously :)

People born in different eras often develop different worldviews by the time they reach their 50s. Not everyone is lucky enough to be born in the "right time".

Sure, write a step-by-step action plan and leave it for

a next fresh new 1M tokens context window.

Thanks that is very inspiring. I thought there are no blower type consumer GPUs, but apparently they exist!

I am curious about the setup of 14 GPUs - what kind of platform (motherboard) do you use to support so many PCIe lanes? And do you even have a chassis? Is it rack-mounted? Thanks!

My suggestion

1. Start by learning a simulation tool, e.g. Mujoco (open source) or Isaac Sim. 2. Learn basics of optimal control and reinforcement learning, reproduce papers/ideas in the simulation. 3. Get your hands dirty on a cheap robot, and try deploy your trained model on it. For mobility and manipulation. Unitree Go1/Go2 for mobility, and robotic arms for manipulation.

I think having to debug to find problem of your system is frustrating. But with NixOS, I at least won't be afraid of "breaking the system" or doing something "irreversible". This is totally a peace of mind when tinkering with my setup.

Thanks so much for sharing your story – hearing about your DIY Enphase install (and epic PG&E battle!) really gives me confidence. And the information you shared is extremely helpful for first-time DIYers like me.

Wow, what a fantastic write-up—thanks for sharing this! I’m a San Jose homeowner (and PG&E sufferer) with a homelab that pulls over 1 kW, and I’ve been down the DIY solar rabbit hole for the past two weeks. Based on my research, I’m planning a roughly 9 kW Signature Solar setup:

20× 455 W Canadian Solar panels (~$173 ea)

1× GridBoss MID V2 (~$2 400)

1× FlexBoss 21 (~$2 400)

4× Eco-Worthy 48 V 100 Ah LiFePO₄ batteries (~$1 500 ea)

18 U server rack (~$500) — total hardware ~$14 760

My big hang-up has been the rooftop work, permitting and inspections—almost no one I call will touch a true DIY system. If anyone here in the Bay Area has recommendations for installers or back-of-house permit-whisperers who’ll partner on a non-Tesla/Sunrun job, I’d love to hear how you made it happen. Thanks again for the inspiring guide!

For the second problem, if I understand it correctly, you might want to try `vmap` from jax.

Yes, we are in the U.S., but our situation seems quite different from the standard RSU process you described.

We did not receive a W-2, and the company has not reported the RSUs as taxable income yet.

Even though our RSUs fully vested at IPO, they are not yet settled as shares—the company has set the settlement date to March 15, 2025.

The company is requiring us to prepay withholding taxes in cash before they release the shares. If we don’t pay by the deadline, the RSUs will be forfeited entirely.

This is why we are trying to better understand how this aligns with U.S. tax laws and whether this is standard practice.

We agree that this doesn’t sound like how RSUs typically work in the U.S., which is why we are seeking advice. If you have any thoughts on how this situation might fit within U.S. tax regulations, we’d really appreciate your perspective!

I am sorry that you had to go through such shenanigans. Consulting a lawyer is also what we are working on right now. Asking just in case: do you have any lawyer to recommend? Thanks!

> It could have been a double trigger vesting arrangement, where the shares "semi-vest" over time, but then they don't fully vest until a liquidity event, at which point poof suddenly all of those ghost shares become REAL shares.

Exactly this.

> If that's the case, they likely baked in a process for employees to have those shares withheld, or auto-sold during the lockup period. It's all tied in with their HR system and other payroll processes to make that easy.

In the agreement, they said this is up to the company, and the company chose the "pay tax to me or forfeit" option.

> There are companies that will loan you money to cover vesting costs or these types of situations - they'll do it at shitty rates, but if the options are losing out on a windfall or losing an extra 10-20% on the windfall, it's worth considering.

Thanks for this advice. Agree that this seems like a viable approach. Appreciate it!

Sent the email. Thank you!

Thanks for mentioning the safe harbor rule. We are actually aware of that.

The issue here is that the company is asking the payment directly to the company's bank account, or the RSUs will be forfeited forever. This makes the situation much worse IMHO.

The shares are not in our brokerage account yet. According to the company, they need to confirm the tax payment to the company before the shares are transferred to the personal brokerage account.

From my experience,

1. There are many services that is already "implemented" in NixOS, with sane default configurations and easy to customize (because the contributors have designed good abstractions, and also because of the flexibility of Nix language). One good example is `nginx`. Btw `paperless-ngx` and `jellyfin` are also already implemented. In this case you do not need to use docker at all.

2. Because of the good abstraction in the service implementation, I usually do not need to go very deep to understand the common configurable options for each of the services.

3. All those services become systemd services once up. As long as you are familiar with how to manage systemd services at runtime, you know how to work with them.

4. Even for those ones that do not exist in NixOS, as the authoer suggested you can still start them as docker-based systemd services, with very simple and intuitive nix configurations.

5. NixOS configuration are mostly deterministic and modular. I can use git to manage all the configurations for different servers. There can be occasions that I will need to migrate the services to a differen machine (e.g. upgrade, replicate, ...). With the NixOS configuration of those services, I can simply re-use the configuration code and have a very high confidence that they will work as expected on a new machine.

6. The above also makes it very easy to revert my deployment to any previous successful version. Without having to worry about breaking anything, it also gives me the confidence to quickly try out different ideas.

1. There are many services that is already "implemented" in NixOS, with sane default configurations and easy to customize (because the contributors have designed good abstractions, and also because of the flexibility of Nix language). One good example is `nginx`. Btw `paperless-ngx` and `jellyfin 2. Because of the good abstraction in the service implementation, I usually do not need to go very deep to understand the common configurable options for each of the services.

I understand `pandas` is widely used in finance and quantitative trading, but it does not seem to be the best fit especially when you want your research code to be quickly ported to production.

We found `numpy` and `jax` to be a good trade-off between "too high level to optimize" and "too low level to understand". Therefore in our hedge fund we just build data structures and helper functions on top of them. The downside of the above combination is on sparse data, for which we call wrapped c++/rust code in python.

I am currently using a Beelink SER8, a very decent powerful mini PC for its price. It is also the most quiet mini PC I have used.

Thanks for open sourcing this project! I've packaged it with nix to make it easier for others to use: https://github.com/nixvital/ml-pkgs/blob/main/pkgs/aider/def...

If you are running nixos, an example of using it can be found here: https://github.com/breakds/nixos-machines/blob/main/flake.ni...

I have been using pyright with lsp in emacs and can recommend.

Good point. Maybe Nvidia should also just publish this on Steam ...

In this context, "zero-shot" capability usually means "how well it performs without fine-tuning or prompting with in-context examples" after the model is trained and parameters frozen.