Will Beddow

supercomputing @krea_ai

[Will Beddow @willbeddow5366]

I’m really excited about this one!

Building Krea’s data infra, I increasingly believe that AI research workloads are neither OLAP nor OLTP. This event has come out of some convos I’ve had with @willmanning, @ovaistariq, @HammadTime and others who have seen the shift happening.

Even a small foundation model company in 2026 needs to be looking towards trillion row and exabyte scale. The huge GPU nodes you buy come with supercomputer levels of CPU to be used efficiently. GPU, memory, and storage prices are spiking. The tool landscape is moving towards disaggregated storage and streaming, and the scale is 10x’ing every year.

All of the speakers are phenomenal distributed systems engineers, and I think it’ll be a lot of fun to have them all in one room. If you’re into databases and distributed systems, please join us!

[Krea @krea_ai]

we're hosting a 'Big Data 3.0' next Tuesday (June 16) in our SF office with @SpiralDB and @TigrisData.

we'll have technical deep-dive talks from frontier AI labs about internet-scale distributed data systems for AI research.

details below 👇

[Will Beddow @willbeddow5366]

@user2jz lol flux was just the first tool to be in this particular part of the UI and we were just bad and never changed the tag. It’s a family of pretrained from scratch models :)

[Krea @krea_ai]

this is Krea 2.

our first foundation model, built completely from scratch for aesthetic diversity and stylistic control.

learn more and get early access 👇

[Will Beddow @willbeddow5366]

https://t.co/IfeBPBxfl1

[Will Beddow @willbeddow5366]

Prototype of a weird and useless idea - a DuckDB extension that treats an LLM like a SQL database. SELECT statements sample the LLM and INSERT finetunes it.

Any mutation (INSERT, UPDATE, etc.) will start by sampling and auto-generating query examples for training that incorporate existing data, so we don't catastrophically forget on subsequent mutations.

The LLM has special <row> <col> etc. tokens registered, and is doing constrained decoding to output data of the shape the extension expects

(video sped up)

[Will Beddow @willbeddow5366]

There are lots of other issues as well - in the video, you can see that after I fill a table with mangoes, blueberries, broccoli, the LLM hallucinates the spinach row, and starts failing on any complex queries. Ultimately, I expect any version of this system to perform better on semantic data than numeric. Lots of fun plans and analysis on this and a number of other aspects of it!

Totally useless, but turned out to be a really fun area to explore.

[Will Beddow @willbeddow5366]

In the end, haven't even made it to my 50GB CSV yet, because it immediately turned into a fun catastrophic forgetting challenge! In most naive implementations, if you INSERT (a, 1) into table, and then (b, 2), it'll forget (a, 1).

In the current version, when you see (b, 2) in table you start generating examples to sample. Ex SELECT COUNT(*) FROM table WHERE col > 0... sample current matches, incorporate in training set

[Will Beddow @willbeddow5366]

Made a little util to make it easier for agents to share GPUs on a node.

CUDA_VISIBLE_DEVICES=$(revolver) your_process - revolver will block until there's a GPU free, and then resolve with the index of the available GPU. https://t.co/NF2mzRH4Ee

[Will Beddow @willbeddow5366]

Krea’s supercomputing team is going to be at GTC on Wednesday! If you’re interested in distributed systems, huge data workloads, and managing clusters with thousands of GPUs, let’s chat

[Will Beddow @willbeddow5366]

https://t.co/D3yc6TzKg7

[Will Beddow @willbeddow5366]

Got tired of waiting for NCDU on large research filesystems, vibecoded a multithreaded version. FDU (Fast Disk Usage) can crawl a 100TB weka filesystem with millions of small files in &lt; 1 minute. Side by side comparison with NCDU on a 20TB dataset I'm using for a side project:

[Krea @krea_ai]

today, we're announcing the acquisition of @wand_app and the release of our new iPad app.

Krea iPad integrates the best of both worlds: native iOS feel with custom brushes and real-time AI.

download it now 👇

[Krea @krea_ai]

today, we're making another step towards the future.

introducing our first Real-time Video generation model.

join the beta 👇

[Krea @krea_ai]

today, we're releasing an open version of Krea-1.

FLUX Krea is a new SOTA open-source image model with incredible realism and aesthetics.

try it for free 👇

[Krea @krea_ai]

today, we're introducing our first image model: Krea 1.

Krea 1 offers superior aesthetic control and image quality. It has a wide range of artistic knowledge and supports style references and custom trainings.

learn how to try our free beta 👇