Alessandro Ercolani retweeted
Everytime I see a Dwarkesh video, I'm wondering what they are smoking in SF. It seems very distant from reality
"Superintelligence" is one of the cringiest terms I've seen, it's a rebrand of AI to get more money out of VCs.
An no, recursive self-improvement hasn't started. It's just human researchers prompting to say "speed it up". It does some cute kernel optimizations and hyperparameter sweeps, very similar to the NanoGPT speed run records done by humans. Nothing fundamental or groundbreaking gets invented by so-called superintelligence.
To give an example: when @Recursive_SI released their (impressive) Flint model on @huggingface, I cringed a little for their use of the terms "superintelligence" + "recursive self-improvement". As said in their blog post, through "recursive SI", LLMs were able to figure out - I kid you not - to determine how many total parameters and active parameters the model should have. So yeah, it's just changing a config. It's doing hyperparameter sweeps on things humans have built (like MoE's, validation loss, perplexity etc). Of course, you shouldn't tell that to VCs. You should scare them by saying AI is improving itself at an alarming rate - this increases the $$$ raised by another factor.
Can we please stop all this misinformation: no, superintelligence is not here. No, recursive self-improvement hasn't started. And yes, there's lots of work to do to build more reliable AI that is actually smart.
This is why there's such a huge gap between super impressive benchmark numbers of LLMs, and the (lacking) economical impact of them. It's not able to do research like a human, doing experiments, coming up with actual new stuff, and generalize beyond their large amounts of training data.
Alessandro Ercolani retweeted
Cancel your plans for the weekend. Nouamane is back with a new banger.
📖 The Ultimate MoE Training Handbook 📖
We spent 1/2M GPUh to optimize MoE training and wrote the most comprehensive blog about it. With 30k words and 50+ interactive figures it covers everything from profiling to large cluster training.
huggingface.co/spaces/huggin…
Nevso combines two ~0.4B models: Nev chooses tools, Nesso3 calls them and answers.
English + Italian. Runs locally on CPU. Works offline after setup.
How we built it, results and code ↓
medium.com/@giuxale/nev-choo…
#EdgeAI #OpenSource
🧵 1/ I’m increasingly unconvinced that we are building AGI.
What I see instead is a new kind of machine: trained heavily on code, formal systems and digital environments, becoming extremely good at exploring and automating the digital world.
medium.com/@giuxale/agi-si-o…
7/ This kind of machine can explore millions of possibilities, use tools, write code, analyse scientific knowledge and execute long digital processes.
It can massively accelerate human progress.
An LLM does not need to become Einstein to help the next Einstein.
1
9
🧵 8/ The AI revolution is already here.
I’m just not convinced it is moving toward AGI or Superintelligence.
What may be emerging instead is an intelligent machine that extends human capabilities at unprecedented scale.
Full article:
medium.com/@giuxale/agi-si-o…
6
I just published the recipe for building Nev Decision Model: 438 million parameters, bilingual decisions, your own hardware
medium.com/@giuxale/nev-deci…
2
26
Meet Nev: a 438M-parameter decision model for English and Italian.
Route requests. Select tools. Classify intents. Run locally—even on a Mac CPU.
Built on our own foundation pretrained from scratch.
🤗 huggingface.co/mii-llm/nev
More below 🧵
🤖 Made with AI
Nev’s weights, code, training recipe and evaluations are public.
Two variants, balanced English–Italian decision training, and examples to try on your own hardware.
📖 github.com/mii-llm/nev
1
1
39
How did we build Nev—and where does it work best?
Our launch post covers the from-scratch foundation, bilingual training, comparison with Kev, and how Nev + Nesso3 work together on CPU.
Read the full story:
github.com/mii-llm/nev/blob/…
#EdgeAI #OpenSource
1
27
Alessandro Ercolani retweeted
Our best-scoring 0.4B model never shipped.
Distilling from our 3B model gave the top tool-calling scores we measured, but cost 5 points of Italian instruction following. So we released the one without that trade-off.
Nesso3, and why we left it out 👇
medium.com/@giuxale/nesso3-a…
Our best-scoring 0.4B model never shipped.
Distilling from our 3B model gave the top tool-calling scores we measured, but cost 5 points of Italian instruction following. So we released the one without that trade-off.
Nesso3, and why we left it out 👇
medium.com/@giuxale/nesso3-a…
Banger paper from Tencent on recursive self-improvement for terminal agents.
(bookmark it)
This is a great read if you are tracking research on training terminal or coding agents with RL
Starting from Qwen3.8-27B, they alternate two phases over five stages. One phase fine-tunes on rewritten versions of trajectories that passed the verifier. The other runs PPO with a critic inherited from the previous stage.
The four-suite Terminal-Bench average goes from 20.68% to 50.58%, with 86.52% on Terminal-Bench 2.1 and 79.00% on Terminal-Bench Hard.
The reward audit is just as useful. Under a dense progress reward, failed runs collected 50.09% of the reward, and 48 failures passed every executed check but still missed the final test. Separating strict success from partial credit cut that share to 28.72%.
arxiv.org/abs/2610.22192
Chat with Paper: academy.dair.ai/papers/termi…
🇮🇹 We’re releasing Nesso3-0.4B-Agentic: a 400M-parameter model for tool use in Italian.
74/100 on our bilingual function-calling suite (Qwen3-0.6B: 67)
38/50 on the Italian half (Qwen: 29)
Built from scratch by the mii-llm community, fully open. 🧵
Where it loses, and we publish it: Qwen3-0.6B is better on generic English tools (84% vs 48%), MMLU, IFEval and English chat.
The blog also covers 4 failed experiments. One public dataset collapsed us from 74 to 49, and a distillation run cost 5 points of Italian IFEval.
1
34
Weights, recipe, eval scripts and the raw results behind every number:
✍️ github.com/mii-llm/zagreus-n…
🤗 huggingface.co/mii-llm/nesso…
Thanks to @Seeweb for the compute. Try it, break it, tell us.
1
32








