Founding AI Engineer, Developer Relations @Wandb @Coreweave 🐝✨ Content on Serverless Inference, Evals, Sandbox the stuff that’s actually moving right now
Joined January 2025
- Tweets289
- Following164
- Followers91
- Likes1.1K
Pinned Tweet
During @reflection_ai’s private preview of Beam, I built a haunted escape room to put its coding and reasoning to the test.
I created a game called “Code Your Way Out”, and it's one room with three linked puzzles. Beam has to repair a brass orrery by lining its rings up with a constellation, route the light, then program its escape through a sealed hatch.
Beam never sees the room. Each turn, it gets text observations and writes Python; the code runs in a @CoreWeave Sandbox, and the game engine validates every action before the explorer robot moves.
Beam was out in about 30 seconds. Super bullish on this model and really excited to see builders get this model in their hands.
This is why Opus 5.5 stays being my #1 🤷♂️
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good at coding, or good at DeepSWE? 🧵
31
Swarms are everywhere, so I got curious how the agents actually talk to each other.
Gave each agent its own @CoreWeave Sandbox, traced every message in Agent Lens, and added a shared memory. They wrote 12 notes. Nobody read a single one.
Also really enjoying gamifying all my demos lately 🎮
More to come on this fun one.
my portfolio has two modes: cozy… and it gets hacked by fsociety. built my whole site as a fake operating system because boring portfolios are dead.
thecodingsoup.com
1
50
The most interesting AI result I've seen in a while.
A frontier model debugging another model in weight space 0.016% of weights edited, 6x math, no gradients. This is models as programmers of other models.
Honestly a new era to model building is approaching us!
You don’t need a PhD to be a model builder soon!
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
1
106
I already own a touchscreen mac, it’s called the iPad. It also didn’t cost me nearly 3k…
Give us better…
Apple’s upcoming OLED touchscreen MacBook Pro could start at $2,899, a $400 price increase over the current model 😳
The 16-inch model could start at $3,399 or even higher
Source: @markgurman
1
53
I'm sorry what... a model small enough to run on macbook and is a speech model...
I think I know what i'm testing on Monday.
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
50
Not ARC pleaseeee 2026 is rough enough as it is..
We've decided to sunset Arc on December 22, 2026.
That's not an easy sentence to write. We spent years building Arc and celebrating each and every release along the way: we have so much pride and love for what we made alongside the Arc community.
Our priority is to deliver best-in-class security for our members. Keeping a browser secure takes constant, dedicated work that is only increasing in the age of AI. Dia was built from the ground up to meet today’s security standards, so that's where our engineering and security teams will focus moving forward.
We've worked to make switching to @diabrowser as easy as possible for Arc members. Here's how:
· Your passwords, bookmarks, extensions and history come over to Dia seamlessly.
· Spaces, Folders, and Tab Groups are in Dia and import over with your data intact.
· Dia now surpasses Arc benchmarks in stability, speed, and security.
· The browser features you use in Arc are free in Dia, and the AI is optional.
Thank you for using Arc, for pushing us to make it better, and for building alongside us. The team that made Arc is the same team behind Dia, and some of our most ambitious work is just ahead. We'd love for you to be there with us when it lands.
54
Ugh… i’ve used arc since the beginning… now i have to move.. ughhhh
We've decided to sunset Arc on December 22, 2026.
That's not an easy sentence to write. We spent years building Arc and celebrating each and every release along the way: we have so much pride and love for what we made alongside the Arc community.
Our priority is to deliver best-in-class security for our members. Keeping a browser secure takes constant, dedicated work that is only increasing in the age of AI. Dia was built from the ground up to meet today’s security standards, so that's where our engineering and security teams will focus moving forward.
We've worked to make switching to @diabrowser as easy as possible for Arc members. Here's how:
· Your passwords, bookmarks, extensions and history come over to Dia seamlessly.
· Spaces, Folders, and Tab Groups are in Dia and import over with your data intact.
· Dia now surpasses Arc benchmarks in stability, speed, and security.
· The browser features you use in Arc are free in Dia, and the AI is optional.
Thank you for using Arc, for pushing us to make it better, and for building alongside us. The team that made Arc is the same team behind Dia, and some of our most ambitious work is just ahead. We'd love for you to be there with us when it lands.
46
Its out!
Try @reflection_ai Beam on @CoreWeave serverless inference!
If you also get a chance check out the demo I built that is on the launch page, and pinned on my X as well!
🔦 Beam, Reflection's first model with open weights, is live today on CoreWeave Serverless Inference.
Call it through an API and leave the GPUs to us.
We are also proud to share that @reflection_ai used @wandb in CoreWeave Forge during Beam's training.
Developers in the private preview have been calling it fast and accurate, and they like that it doesn't ramble.
See what @TheCodingSoup on our team built with private preview access below.
Get started with Beam here: crwv.co/utuxk
2
7
167
👀👀👀
New model dropping soon on @CoreWeave Serverless Inference 👀
CoreWeave Agent Lens (@weave_wb) tracing built in.
3
83
Interesting rundown of where benchmarking stands and what benchmaxxing actually looks like in practice.
Moritz Hardt thinks that OpenAI and Anthropic’s claims about benchmarks, which are used to test AI models’ capabilities, are misleading. But if the tests aren’t accurate, then how do we know how good the models really are? Read the interview: quantamagazine.org/he-wants-…
2
64
@reflection_ai just launched their new model, Beam, and my demo made it onto their launch blog! 🎉
For the demo, I built Beam an escape room. I locked it inside a series of puzzles and let it work through each problem on its own to find a way out.
Watching it reason through the clues, test ideas, and adjust when something didn’t work was a fun way to see what the model can do. Tracked everything using @CoreWeave sandbox and agent lens!
Congrats to the Reflection team on the launch. Check out the blog to see the demo in action.
reflection.ai/blog/beam-is-a…
5
106
Beam is a lot of fun to work with. Checkout my pinned post on how I used it for an escape room!
Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners.
We're releasing:
- Weights, tokenizer, and reference implementation
- BF16, FP8 and NVFP4 weights and a DFlash2 drafter
- The open source harness we built and trained Beam on: Mirror CLI
Weights and model card on Hugging Face: huggingface.co/reflection
1
11
385
Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners.
We're releasing:
- Weights, tokenizer, and reference implementation
- BF16, FP8 and NVFP4 weights and a DFlash2 drafter
- The open source harness we built and trained Beam on: Mirror CLI
Weights and model card on Hugging Face: huggingface.co/reflection
48
Finally... 😭😭😭
I'm so sick of juggling sessions across 10 terminal tabs with zero idea what's running where side by side might actually make the desktop app usable.
We've been working hard on making the Claude Code desktop app great for running lots of sessions side by side.
It's now much quicker to start sessions, easier to split work across a few, and you can always see what each of them is doing. Here’s what's new:
68