Donato Capitella

I'm a Software Engineer and Ethical Hacker, but mostly a tech enthusiast who likes to discover how things work by breaking them apart.

[Donato Capitella @dcapitella]

@OpenAIDevs once again VSCode extension is totally broken on Linux, I don't know how you keep not testing what you ship.

[Donato Capitella @dcapitella]

@EtAjith @RiversidedotFM The edits had nothing to do with what I asked, and were in random places, like it undid a lot of the work on me choosing where to keep breaths and pauses abd where to remove them, to keep the recording natural, plus I had deleted some parts, and some of them randomly reappeared.

[Donato Capitella @dcapitella]

@RiversidedotFM The AI agent you ship with your product is CRIMINALLY poor quality, at the end of 2026 such a poor AI experience shows a direct, deliberate disregard for customers. it totally destroys beyond repair ANYTHING it touches, edits are totally random.

[Donato Capitella @dcapitella]

My two cents on the DeepSeek ban in the US and model backdoors

https://t.co/BrrBZD74Hf

[Donato Capitella @dcapitella]

@OpenAI Codex is falling apart again...

[Donato Capitella @dcapitella]

@Italianclownz Great work, can't wait to try this out!

[Carlo @Italianclownz]

I uploaded the custom GLM 5.3 Flash to hugging face and if you use the recommended test build of ROCmFPX for GLM you should get prefill above 400 tok/s over 100k tokens.

Everything is explained in the model card.

Should work well on Strix Halo 395+ 128 GB. Hope it works well for everyone like it does for me.

If you use any of the GLM code boosts in your projects please give credit to ROCmFPX

https://t.co/6HyuQbnKV2

[Donato Capitella @dcapitella]

@ivanfioravanti Biased, but add AMD to the mix :)

[Donato Capitella @dcapitella]

@brannonhogue I don't like PHP, and jokes aside though, this can realistically happen to any platform :/

[Ivan Fioravanti @ivanfioravanti]

Star Wars PodRacer created locally with
- DwarfStar/ds4
- DeepSeek V4.1 Flash 1017 Q4 (what a model!)
- DeepSeek Harness
- Dual M3 Ultra 512GB (It can run on a single one)

Stats
Average speed: 43 t/s
Total time: ~ 10 hours
Turns: 7
Steps: 226
Token usage: ~65M toks
Output: 390K toks
Context used: ~497K / 1M
System prompt: ~1.4K
Tool definitions: ~7.1K
Messages: ~350K

It added on its own commentary, telemetry and create all graphics and sounds without external tools!

Prompt: Make a 3D world of a Star Wars Podrace as good as you can, as good as you can. Use different camerase throughout the race, POV from Skywalker, behind the pod, lateral, from top of canyon during a curve. If possible add incident of another driver with his pod exploding. No tech restrictions. Goal is to output the most impressive and spectacular visual piece.

[Donato Capitella @dcapitella]

@f4talStrategies @xlr8harder I understand what engram does, but that doesn't explain the security implication. Conditional activation of learned behaviours is already possible in conventional Transformers. What specifically about Engram makes malicious backdoors easier to implant, conceal or trigger?

[Donato Capitella @dcapitella]

@f4talStrategies @xlr8harder For some reason X showed me this comment and this is stating that somehow deepseek architecture makes it easier to backdoor the model than say Kimi K3. Can you elaborate? How is it making it easier/more plausible, what's is the mechanism?

[Donato Capitella @dcapitella]

@ivanfioravanti thanks, changing it in my ROCm fork!

[Donato Capitella @dcapitella]

@RiversidedotFM It's just HORRIFIC, who thinks to even charge clients for this? I'm beside myself for how bad this AI chatbot is at editing stuff.

[Donato Capitella @dcapitella]

@z3ratul163071 @pupposandro @davideciffa 😎

[Donato Capitella @dcapitella]

DeepSeek V4.1 Flash 1017 Update: Using 44% Fewer Tokens, Running on Strix / Gorgon Halo with DwarfStar

https://t.co/vKL1ySJEex

[Donato Capitella @dcapitella]

@kbsingh The previous checkpoint they droppd a month ago

[Donato Capitella @dcapitella]

For people wanting to try DeepSeek V4.1 Flash 1017 on Strix Halo, you can use my dwarf-start fork (or the pre-built toolbox in ai-toolbox-cockpit) and these weights:

https://t.co/dq7G1ziyQh

This also supports DSpark. I'm doing some testing, check comments for early results...

[Donato Capitella @dcapitella]

@LMtxIoTAdvisor Indeed, Qwen is much faster in terms of inference

[Donato Capitella @dcapitella]

@LMtxIoTAdvisor Comparisons are truly tough, feeling is that it can cope with more things, and I'm doing a proper review ongoing for a few weeks, but takes time