Sign in

Jacopo Nardiello

@jnardiello.bsky.social
28 followers 21 following 15 posts

Obsessed by automation. Unstoppable. I tame and teach machines 🇪🇺🇮🇹 "Through the fire, and the fury, with a heart made of steel". OSS/Acc.

PostsRepliesMedia
Jacopo Nardiello @jnardiello.bsky.social · 08/08/2026
I’m going to spend a lot more time helping this community push the frontier forward, because there’s clearly still so much to do. I’m absolutely excited. I’m particularly interested in the intersection of local AI and traditional cloud-native workloads - there, too, there’s so much work to do.
000
Jacopo Nardiello @jnardiello.bsky.social · 08/08/2026
- Are local models useful? Hell yes. - Can you trust them for production work? Also yes.
100
Jacopo Nardiello @jnardiello.bsky.social · 08/08/2026
Boy, was I wrong. DS4 and DeepSeek can get real work done. DeepSeek is extremely smart and surprisingly self-aware of its own limitations. It still pushes for useful results, and it does so in a very balanced way. - So, is local AI good enough for real-world work? Yes.
100
Jacopo Nardiello @jnardiello.bsky.social · 08/08/2026
I played with local models a long time ago and eventually abandoned them. They just weren’t useful—even with steering. I was convinced the hardware requirements would be too high for anything truly practical to run locally.
100
Jacopo Nardiello @jnardiello.bsky.social · 08/08/2026
Okay, after spending the day with DS4 and DeepSeek-V4-Flash-0731 on a single GB10, I can honestly say I’m extremely surprised at how far local AI has come.
100
Jacopo Nardiello @jnardiello.bsky.social · 08/08/2026
I'm now trying to understand why speculative decoding isn't bringing any speed improvement with upstream DS4 (and if I can come up with anything interesting to fix this). GPT-5.6 on it.
000
Jacopo Nardiello @jnardiello.bsky.social · 08/08/2026
Speed I got so far (300k context): - DS4 vanilla (no DSpark): ~15 tok/s - DS4 + DSpark: ~11 tok/s (slower, not great) - Entropy DS4 + DSpark: ~35–40 tok/s
110
Jacopo Nardiello @jnardiello.bsky.social · 08/08/2026
Finally, I managed to spend the morning on the GX10/DGX Spark and DS4. First of all, I want to focus on pushing this little monster to the max and see how much performance I can squeeze out of it. Initial tests are extremely promising (thanks @MiaAI_lab).
100
Jacopo Nardiello @jnardiello.bsky.social · 19/07/2026
With this, Codex can assign roles and spawn subagents on any provider as native subagents. Why is this better than standard CLI calling (as models do natively)? Because of better stdout integration, deterministic responses, and ultimately more reliable workflows.
000
Jacopo Nardiello @jnardiello.bsky.social · 19/07/2026
Today, I managed to use github.com/router-for-... to achieve full third-party (Fable and Grok 4.5) integration with Codex subagents.
github.com
GitHub - router-for-me/CLIProxyAPI: Wrap Antigravity, ChatGPT Codex, Claude Code, Grok Build as an OpenAI/Gemini/Claude/Codex compatible API service, allowing you to enjoy the free Gemini 3.1 Pro, GPT 5.5, Grok 4.3, Claude model through API
Wrap Antigravity, ChatGPT Codex, Claude Code, Grok Build as an OpenAI/Gemini/Claude/Codex compatible API service, allowing you to enjoy the free Gemini 3.1 Pro, GPT 5.5, Grok 4.3, Claude model thro...
100
Jacopo Nardiello @jnardiello.bsky.social · 30/03/2025
Yesterday night I have spent a huge amount of time refactoring some Lua code. Sonnet was useful, but at some point it hallucinated and started spinning around. After hours, Gemini solved the issue first attempt, in minutes. I was “wow, it did it”.
000
Jacopo Nardiello @jnardiello.bsky.social · 30/03/2025
The incredible thing is that Google just dropped a model that outperforms o1 with reasoning AND sonnet for coding. I need to test this thing out better, but so far I am impressed.
100
Jacopo Nardiello @jnardiello.bsky.social · 29/03/2025
I honestly think that Gemini 2.5 is on a league of its own. None of the other models, at least when it comes to programming, can compete (currently).
010
Jacopo Nardiello @jnardiello.bsky.social · 27/03/2025
Great article from @arslan.io pretty much summarizing why AI is now absolutely relevant for software engineers. Just one minor note: try also aider.chat, it rocks and integrates perfectly with nvim.
aider.chat
Aider - AI Pair Programming in Your Terminal
120
Reposted by Jacopo Nardiello
antirez @antirez.bsky.social · 14/03/2025
So with QwQ and Gemma 3 we are finally seeing models <= 32B parameters that can be used for non trivial tasks. Far from SOTA? Yes. But far from the old generation of smaller models as well.
0222
Jacopo Nardiello @jnardiello.bsky.social · 14/03/2025
Gemma3 is surprisingly good. Local models starts to look like viable alternatives, super exciting.
000