Derek Fulton

Might it be better to go where the energy is rather than build new centralized data centers? My house currently gets 200A but if I wink wink get an electric car I can bump it to 400A. At 120V that’s 48KW. 20 thousand of those and there’s your gigawatt.

@akashnet @IEEESpectrum @gregosuri Yeah. My house can get up to 400A at 120V. Along with 2x 8G symmetric with Google Fiber, that’s a non-negligible amount of compute for anything instead of frontier model training. And there are millions of houses like mine in America.

Decentralized AI training is gaining mainstream attention. Read @IEEESpectrum's thoughts on Akash, quoting @gregosuri "such potential lies in moving AI to where the energy is instead of moving the energy to where AI is" → spectrum.ieee.org/decentralized-…

@gregosuri Between you guys, @exolabs and @__tinygrad__ along with a few others I’m probably forgetting, I think huge things are going to happen with local compute in the next 2-5yrs.

OpenAI announced they are pausing their plans to build out data centers in the UK due to energy constraints. This is one of the many proof points that centralized AI data centers have limits to scale. Meanwhile, decentralized AI training is proving to be a promising alternative where the compute is distributed across the globe, not limited by the grid.

@gregosuri @akashnet Despite the likes of Vast, Filecoin, general crypto, I think we are still early with home node federated compute

As I predicted, energy bottlenecks are preventing new data center buildouts Can’t be more excited for @akashnet Homenode 👉 homenode.akash.network

As an engineer I paid for stuff out of pocket all the time. I think a big issue is the tax structure around this. If you buy something for your work that is obviously a work expense whether you’re on W2 or not. Workers don’t like that and would rather the employer just give them a gift card equivalent of far less money than any potential raise.

companies churning talent because they don't let their engineers comp random stuff they want to buy if you're making it a pain to buy airpods, dinners, studio displays, or whatever random shit they want to buy you're ngmi heard a really lit engineer recently tell me that they only get $200 of spend a year and I could see the life leaving his eyes. This is someone who will just no life to build amazing product but they're going to leave because of a couple thousands dollars a year?

@richardartoul Egress is an absolutely enormous amount of AWS revenue. They will never say how much anybody know a good estimate?

Didn’t even mention pricing, S3Files is insanely expensive

@iann0036 @awscloud Classic

Hey @awscloud, a new API should NOT be a health event. Stop waking people up with spam.

@dataenggdude Shots fired. Self hosted or Mother Duck?

Why I'm replacing Polars with DuckDB open.substack.com/pub/dataengine…

@HusKerrs Keep the baby on an exact schedule. Everything else is downstream.

Calling on all dads: Ali and I are about 3 weeks out from the birth of our first baby boy! Give me your #1 piece of advice for a new dad.

Cooking up something big

Alice from Carolina Cloud presenting next week luma.com/1yk8yvuk Cloud Computing for Bioinformaticians

The year is 2026. It now costs you $0.06 pet GB to write files to your own colocated S3 bucket. Carolina Cloud has been doing this for 1mo, which is actually a lot relative to Carolina Cloud’s age.

@vikhyatk No but I saw it on HN. Definitely a contender. We’ve been through the gauntlet on S3 mounting. Currently using rclone, looking at JuiceFS but rclone is so far so good.

anyone heard of FiberFS / used it in production?

@SebAaltonen I would love to see a MacBook capable of really nothing but a browser, mid-weight IDE (like Zed) and obviously terminals. Maybe no IDE. But with enormous battery life. Like weeks.

People said I am crazy when I said that 8GB in MacBook Neo is not enough and I am waiting for next year's A19 Pro model with 12GB in-package RAM. Seems that I was right. The reason for 8GB RAM was that A18 Pro package has 8GB RAM in it. A19 Pro package has 12GB. Much better!

Am I leaving something on the table by never exceeding 25% of only the $100/mo plan?

@kyle_e_walker Or use @CarolinaCloudIO which is so simple you don’t need a LLM to help

An important tech use-case for LLMs that isn’t pure coding is deploying and managing cloud services For example, Google Cloud Platform is powerful but a total maze to navigate LLMs help me work through the menus, understand the CLIs, and prep deployment scripts

@neural_avb Think of all the surprise bills they're going to reap

This is one of the best innovations to object storage. You need to understand what they did here. A lot of current systems and workflows will simplify because of this. Follow this line of thought: - AWS S3 forever has been a blob/object storage, flat paths, no file navigation. - S3 API had a lot of friction. You would need to find the file you were looking for, download it, then use it. - They have now introduced S3 files which is a shared file system that lets you access S3 data as regular files and directories. - You can launch an EC2 instance and mount it. The files become available in /mnt/whatever. Also available on Lambda and a bunch of other services! - High Performance Storage: Your actively used data lives on a "high-performance file system" - S3 Files keeps it in sync with the linked S3 bucket automatically. Small files (<128 KB) loads super fast. This threshold can be configured. - Larger files have only their metadata imported. Their data is read directly from S3 on demand. - You can navigate through the files using standard fs commands like ls, grep etc. Reads are instant after the data has been synced in the high performance layer. Writes get synced back to your S3 bucket in batches. - The rabbit hole goes deeper: New files become new S3 objects, edits become new object versions, deletions become S3 delete markers. - You have POSIX permissions (owner, group, permission bits) - you can do chmod and chown and all that. These are stored as S3 object metadata. - You wont need to learn all the `aws s3 cp` stuff and just run `cp` if you are dealing with s3 files. - see the attached screenshots of read/write throughput, and it is honestly insane engineering Why THIS MEANS THE WORLD to your agents - This is great for a lot of human workflows already (log analysis is easier now, you can view pictures/small videos easier, etc etc), but it gets even better for agents - You no longer need to replicate the entire S3 bucket as a file system for your agent to access - Agents dont need to run aws commands - Agents can just treat it as a file system layer now, and run commands they are trained on: ls, cat, grep they all work. - AWS takes care of low latency for "hot" data + all the syncing stuff - ASYNC agentic workloads are faster and frictionless now - one agent reads data, another agent from a different EC2 instance writes it, this all can happen. Imo this makes directly using S3 practical for small teams and indie devs, because you dont need to write a bunch of boilerplate and adapters to setup environments anymore.

Right!? That was my first thought as well. I've been experimenting with rclone and JuiceFS recently. How does Seaweed compare to these for edge cases such as large concurrent write (two users on same bucket/object) or repeated read/writes? rclone is really good as it uses a cache for this kind of situation. I feel like JuiceFS + Seaweed could be a serious combo!

Nice! SeaweedFS already does this! Use FUSE mount to access the files in s3 buckets.

Attention data scientists with deadlines! Start an R Studio Server instance with your storage bucket automatically mounted and work with it as if it were a regular filesystem. Zero friction.

Just gonna leave this here