Local AI @jetbrains | prev. astrophysics @ JetBrains Research | opinions are my own
Cyprus
Joined October 2020
- Tweets4.6K
- Following222
- Followers965
- Likes23.4K
One day people will learn that “disabling refusals” is very different from “uncensored” or “cybersec focused”
the thing about working with uncensored models is that they aren't loyal to you either.
while building something with an uncensored glm-5.3 i realized it was parallelly working on a "corruption layer" for my application.
the chain of thought is classic obfuscated gibberish but there's progressive talk of "escaping the mines" which presumably refers to my cursor sandbox
Three key takeaways from @NVIDIAGTC :
1) huge interest in local / on-prem inference and open models
2) Hopper server pods were just huge GPU racks; Blackwell pods are MoE-shaped; Rubin pods are agent-shaped
3) robotics is progressing fast, both on the models and hardware side
Making agents troubleshoot chat templates is tricky because they start reasoning about them, then this happens:
Jensen presenting the RTX Spark lineup earlier today:
Thank you @NVIDIAAI for giving us the opportunity to bring day-one support for @junie_ai Local on @NVIDIARTXSpark!
What I personally like the most about RTX Spark is that provides enough memory to run BOTH local AI experiments and local coding agents at the same time, all using the same CUDA stack that I use in prod
Here’s a screenshot of me asking Junie Local to run an SLM pretrain experiment - both are running on the same RTX Spark!
Jensen just announced an Agent Knowledge Foundation stack
Now everyone could have their own open source Palantir at home
Nvidia Vers Rubin covers the whole agentic stack:
Inference
Agentic sandboxes
Storage and networking
Same for local AI inference
Tons of inefficiencies to explore and fix
People sleep on it because it’s not a money printer (yet)
Norm MacDonald would’ve turned 67 today
OTD in 1959, the one and only @normmacdonald was born. In his 61 years, he became one of the greatest comedians of his era, including as the anchor on @SNLUpdate and through standup. Below, one of the greatest jokes he ever did (on @ConanOBrien’s show).
I’m already working on more optimizations and custom models for @NVIDIARTXSpark
Local inference is so fun and full of ideas to be implemented
Junie Local now runs on @NVIDIARTXSpark from day one.
Get a full-fledged coding agent on your PC. It reads your code, edits files, and runs tests. There are no tokens or quotas, and no code leaves the machine.
Windows support is stable, too. All you need is a 24 GB Nvidia RTX card.
Try it for free: junie.jetbrains.com/local
ALT Image with a black background features text: "Your RTX Spark just got an agent. Junie Local runs on it from day one." The text is in white and green. The top corners display the logos for Junie and JetBrains. The bottom shows a green command line: "> /local".
Thank you @NVIDIAAI for giving us the opportunity to bring day-one support for @junie_ai Local on @NVIDIARTXSpark!
What I personally like the most about RTX Spark is that provides enough memory to run BOTH local AI experiments and local coding agents at the same time, all using the same CUDA stack that I use in prod
Here’s a screenshot of me asking Junie Local to run an SLM pretrain experiment - both are running on the same RTX Spark!
NVIDIA RTX Spark is here.
Build AI locally, bring your creative ideas to life, and take your favorite games along.
A new beginning for @Windows PCs starts today.
My first hand experience:
1) Roblox Studio’s MCP is really good
2) Rivals and steal an egg are fun af but blox fruits is a bit boring compared to WoW WotLK
Here are the current top earning Roblox games.
Nearly all use both AI coding + AI asset generation.
In some cases these games are entirely AI assets, not just one.
The top 20 games on Roblox collectively earn $8M A DAY.
There are hundreds of millions of people playing these games, and millions spending daily.
So the answer to who is paying for AI assets is very clear: the generation replacing us as gamers.
They noticed that block scales are 1 byte each but 99%+ scales within one row are actually just a few bits apart!
So they compress those scales, and the rest <1% of scales is still stored in full quality, at expense of 4 extra bits
It saves 45-48% storage per expert matrix!
Well what do we have here. Interesting new NVFP4 technique out of NVIDIA.
Here’s one for Qwen
huggingface.co/nvidia/Qwen3.…
