Digital Health CIO, Digital Transformation advocate, Machine Learning -FOOD Lover- Linux addicted from 1994 - All views are my own https://teapaw.org/t.co/SaHWj0W3Uq
Milan, Infosphere
Joined October 2007
- Tweets103K
- Following4.1K
- Followers4.1K
- Likes76.3K
Luca Perugini retweeted
So in case you missed it... I created an x86 operating system kernel that is compatible with Linux user-space using HDAE and AI in 11 days. And after about a week I killed it.
That sounds more dramatic than it actually is, but that is the gist.
The point of this wasn't to create a production Linux compatible or competitive kernel (that would be silly), the point was different...
We all know what Vibe Coding is at this point. I define it slightly different then maybe others, as I think of it as someone who doesn't know the subject matter well, asking AI to do something, and they "vibe" (or maybe more accurately, stumble) it into existence.
While people are talking about it, being capable of replacing developers and knowledgeable and skilled humans, I don't believe the output is great.
But, when AI is led by somebody who really understands what they are doing, amazing results can be achieved! I've seen this first hand myself, as well as with others, and the results are outstanding, so I wanted to put my process to the test.
I call this process Human-Directed Agentic Engineering (HDAE), and I decided to try to implement a crazy task with it... Something I'm familiar with, something that can be easily tested and iterated on something that no-sane person would take on by themselves.... Building an operating system kernel capable of running Linux.
And thus Ginux was born.
github.com/gmkurtzer/ginux
To reiterate, the point of the research was NOT to create a kernel or to compete with the works of many thousands of developers, programmers, community members, and companies... The point was to prove a concept, and I did that.
Then what happened is attention, a LOT of attention, in the form of news, articles, podcasts, videos, ad-hoc code reviews, and activity on the GitHub project.
That is great, right?!? But no, that was't my intention. I debated not taking any of the PRs, but this gave me a chance to test something else...
I've been working on formalizing a release of the HDAE skills, so others can more easily leverage this process.
github.com/gmkurtzer/hdae
So I wanted to test these skills and process for also maintaining code and contributions. And while it's not completely done, I wanted to test ingestion of ideas, fixes, reviews, issues, and PRs using this model.
TL;DR, there is more tuning I need to do for this, but it did great! I was able to run automatic reviews on almost 80 PRs and 70 issues, many of which created HDAE gates automatically and built the fixes automatically.
But maintaining Ginux wasn't what I set out to do... I tested a process, and it was a success, so I started considering lessons learned.
One of the things I did different in Ginux was I approached multi-operating system support as a first tier capability layer that I called "domains". You can think of domains as a way to completely separate and isolate user space environments above containers.
In Ginux I had a kernel included controller process but it ran in ring 3 in user space, rather than a kernel thread. It was interactive and allowed the user to configure the user space. So for example, you can configure a "domain" to boot a Rocky Linux 10 container, and it would manage separation more like a virtual machine then a container.
Domains can be scheduled differently for each, and against each other, with harder boundaries then containers.
They also maintain the concept of operating system "personalities" like Linux, or BSD, or Darwin, or maybe even Windows (note: the only one that was done and tested was Linux).
So... My next test for Ginux is a user space kernel shim, that sits between the kernel and user space to provide more control over the user space than the kernel will accomplish. This is because the kernel guarantees user space compatibility, nothing is dropped, stability is paramount. So I asked a question, what if we dropped, broke, or monitored some things on purpose. What could we do?
So there are some projects that already do parts of this, gVisor, user-mode Linux, among others, but what I'm thinking was a bit different... What if in each "domain" you can control what kernel interfaces are exposed, and what the running user space was allowed to do in that environment. What could we control, and do we end up with a more secure environment?
The beauty of this is that it still leverages all of the same capabilities and security of the Linux kernel as it exists now, and then it gives the user the ability to provide very custom and specific locked down environments, all in user space.
So Ginux has pivoted to now research this idea. I honestly have no idea how it will turn out, but thanks to AI and HDAE, I have a feeling in a few days, I'll have a lot more to report back on.
The Ginux repo above, has already been pivoted to the user space kernel, and it is documented within the project. The previous full kernel implementation is archived there in a branch for anyone who wants to see or use it.
Oh, and one other thing, it is pronounced: Jin-uks.
Luca Perugini retweeted
Qwen3.8-27B has reportedly shrunk from 55.6GB to just 5.9GB, while retaining 98.2% of its aggregate benchmark performance.
That’s an approximately 89.4% reduction in model size, according to PrismML, the team behind Bonsai.
The new compression method is called Bonsai 2, and it targets one of the biggest challenges in local AI: making larger models fit into much smaller memory budgets without destroying their capabilities.
The reported results
PrismML’s comparison shows:
Original Qwen3.8-27B: 55.6GB
Bonsai 2 compressed version: 5.9GB
Model size reduction: 89.4%
Aggregate benchmark performance retained: 98.2%
Previous Bonsai 27B version: approximately 95% retention
Evaluation coverage: 20 benchmarks
The improvement from 95% to 98.2% retention is particularly interesting because achieving high accuracy becomes increasingly difficult as compression becomes more aggressive.
How Bonsai 2 achieves this
The technology behind Bonsai 2 is ternary quantization.
A neural network contains billions of numerical weights. These values determine how information moves through the model and influence its predictions.
Conventional formats such as FP16 and BF16 use more bits to represent each weight, preserving greater numerical precision but requiring more storage.
Ternary quantization takes a different approach.
Instead of storing weights using conventional numerical values, it represents them using just three possible values:
-1
0
+1
Scaling factors help preserve information about the original weights.
This dramatically reduces the amount of information that needs to be stored for each parameter.
The challenge is maintaining the relationships encoded in the original weights.
A model compressed this aggressively can lose information that affects reasoning, coding accuracy, instruction following and other capabilities.
Bonsai 2 attempts to minimise that loss while keeping the resulting model extremely small.
Why preserving agent capabilities is particularly interesting
PrismML reports improvements over its previous Bonsai generation in coding agents, multimodal reasoning and long-horizon tool use.
The last area deserves special attention.
A coding agent does much more than generate a response to a single prompt.
It may need to:
Inspect an entire codebase.
Understand how different files depend on each other.
Plan changes across multiple components.
Modify code and execute commands.
Run tests and interpret errors.
Correct mistakes and continue working until the task is complete.
This requires consistent reasoning across multiple steps.
A compressed model might perform well on short benchmark questions but struggle when it has to maintain context, make decisions and recover from errors over a longer workflow.
That’s why reported improvements in agentic evaluations are worth examining separately from aggregate benchmark retention.
What this means for local AI
Reducing a 27B model from 55.6GB to 5.9GB changes its distribution and storage requirements considerably.
It could make a model of this size more accessible to developers using consumer GPUs, laptops and other memory-constrained systems.
A smaller checkpoint is also easier to download, share and manage.
However, there’s an important distinction:
A 5.9GB model file does not mean the model needs only 5.9GB of memory to run.
Inference still requires additional memory for:
KV cache: Stores information from previously processed tokens.
Activations: Intermediate values generated during computation.
Runtime overhead: Memory used by the inference engine and its buffers.
Context processing: Longer context windows generally require more memory.
The actual hardware requirements will depend on the inference engine, context length, batch size and implementation of the compressed weights.
Smaller model files also don’t automatically guarantee faster generation. The runtime must be able to process the compressed representation efficiently.
Qwen3.8-27B Uncensored reportedly reaches 110 tokens/s on a single AMD Radeon RX 7900 XTX.
This is a local inference result, with no multi-GPU server required.
The reported figures:
GPU: AMD Radeon RX 7900 XTX
Model: Qwen3.8-27B Uncensored
Decode speed: Up to 110 tokens/s
Prompt processing: Up to 1,000 tokens/s
Deployment: Local, single-GPU inference
The distinction between those two speeds matters.
Decode measures how quickly the model generates new tokens, while prefill measures how quickly it processes incoming prompts. A high prefill rate helps reduce the time spent processing large instructions and codebases before generation begins.
What makes the uncensored version interesting?
The model is modified to reduce or remove its default refusal behaviour.
That can make it more willing to follow certain instructions during development and agentic workflows, where repeated refusals can interrupt execution.
The reported tests also claim it handled real agentic jobs without refusals.
This is worth separating from model capability, though. Removing refusal behaviour does not automatically improve reasoning, coding accuracy or task completion. Those still need to be evaluated independently.
Why the result matters
A single 7900 XTX delivering this level of reported throughput makes Qwen3.8-27B an interesting candidate for local coding agents and other demanding workloads.
The next measurements I’d want to see are sustained decode speed, time to first token, memory usage, quantization details and actual task-completion results.
The real test isn’t just whether an uncensored model generates quickly. It’s whether it can use tools, complete multi-step tasks and produce reliable results without constant human intervention.
Luca Perugini retweeted
OpenAI ha encontrado 948 fallos graves en el núcleo de Linux en una sola pasada, con agentes de IA que trabajaron solos, y ha escrito un parche para cada uno.
Lo publicó ayer. El núcleo de Linux tiene más de 40 millones de líneas de código y está debajo de Android, de la nube y de los superordenadores. Unos agentes eligieron qué partes revisar, otros buscaron fallos, otros los reprodujeron en máquinas aisladas, demostraron que se podían explotar y propusieron la corrección. Al final, investigadores humanos revisaron todo antes de mandarlo a los responsables de Linux.
Cada uno de esos 948 fallos permitía a un usuario sin permisos hacerse con el control total del sistema. La campaña costó 1,2 millones de dólares de uso del modelo, unos 1.279 dólares por fallo. Concursos de seguridad como Pwn2Own pagaban 20.000 dólares por uno solo. En julio, con un modelo anterior, habían encontrado 166.
En los diez días que duró el trabajo, los responsables de Linux ya habían corregido 61 y tenían otros 43 en camino.
En Certuma trabajamos con datos médicos de pacientes, y la seguridad no se negocia. Que una IA revise millones de líneas de código y deje el parche escrito cambia lo que cuesta defenderse.
Cada nueva generación de modelos va a dejar el software más difícil de atacar que la anterior.
Luca Perugini retweeted
Tool: transfer data through sound, including INAUDIBLE frequencies!
Use it here: chirp.hex.dance
Works offline, data not sent to any server: it’s an in-browser modem supporting frequencies between 18k and 19k Hz, which most humans can’t hear (but your cat can!)
Luca Perugini retweeted
If you've spent 30+ years in an industry, you know problems a young founder probably hasn't encountered yet. I'd love to see more people build their first business or initiatives at 50 or 60. We should stop limiting people by age when they have decades of knowledge and questions worth solving.
Luca Perugini retweeted
Last week, the new Node.js 27 Alpha release upgrades V8 to version 15.6, from the 14.6 used by Node.js 26. Let's take a closer look at what this means for developers and what the new version of V8 brings.
- New JS features and language changes
- Performance improvements
- Memory usage and GC changes
- Diagnostic and tooling changes
- Behavior differences
..and more!
nodejs.org/en/blog/announcem…
Luca Perugini retweeted
C’è un tasto che salva il banco frigo dal ghiaccio.
Ma se lo premi dall'esterno, lo trasforma in un forno.
Non serve un malware: cambi la temperatura, fai partire lo sbrinamento, spegni gli allarmi.
E sembra normale manutenzione.
Cosa è successo?
Lo spiego nel video completo: sul canale YouTube Ciao Internet
👉 youtube.com/watch?v=YAh1eFbF…
#cybersecurity #supermercati #hacker #ciaointernet
Luca Perugini retweeted
‘Catroia.
BREAKING: According to new information, the cyberattack carried out during the night of November 1–2 on European refrigeration systems has also affected frozen plasma intended for transfusions at the French Blood Establishment (EFS).
The attack targeted Coolworld, a refrigeration service provider for the EFS, reportedly disrupting temperature-control systems used to store blood products in France.
According to information revealed by the French weekly Le Canard enchaîné, hackers allegedly manipulated refrigeration systems, causing freezers normally maintained at approximately –30°C to be programmed to reach temperatures of up to 110°C.
The incident reportedly affected facilities in the Bourgogne-Franche-Comté region.
Even more concerning, frozen plasma intended for transfusions to patients was reportedly among the products affected, and not just donor samples used for health monitoring.
Frozen plasma is a critical medical product used in hospitals, particularly in the treatment of severe hemorrhages and certain coagulation disorders. Its storage requires strict temperature controls to preserve its safety and effectiveness. A prolonged disruption of the cold chain could compromise these products and potentially require their withdrawal from use.
According to reports, the incident was detected overnight by an on-call technician who noticed an abnormal rise in temperatures in the cold-storage facilities. The rapid detection reportedly allowed the situation to be addressed before any operational disruption occurred.
The French Ministry of Health has stated that the incident had no operational consequences, and no compromised transfusions or harm to patients have been confirmed so far.
However, several important details remain undisclosed, including the number of facilities affected, the duration of the temperature fluctuations, the actual temperatures reached inside the freezers, and the extent of the potential impact on stored plasma.
Luca Perugini retweeted
I had no idea my Omarchy Apple Dev is in the top 3 most starred Omarchy repos.
I appreciate the attention and PRs it receives, but I think my local AI work on Omarchy is more groundbreaking.
Check out Omarchy-eGPU and the other repos linked from there: github.com/joshuaswarren/oma…
The 10 most-starred 𝗢𝗺𝗮𝗿𝗰𝗵𝘆 repos aren't by DHH. ⭐
They're by the community, and #1 is by Tobi Lütke.
▸ 01 Disktree: 3,867, a treemap of what fills your disk
▸ 02 Owly: 3,222, AirDrop for non-Apple devices
▸ 03 Omarchy Apple Dev: 1,786, build iOS apps from an Omarchy PC
▸ 04 Rawmakase: 1,597, a Lightroom-compatible RAW editor
▸ 05 Flea: 932, the fastest GUI file manager on 𝗢𝗺𝗮𝗿𝗰𝗵𝘆
▸ 06 Omarchist: 773
▸ 07 Omacosy: 705, the 𝗢𝗺𝗮𝗿𝗰𝗵𝘆 feel on macOS
▸ 08 Awesome Omarchy: 690
▸ 09 Flux: 613, your phone and 𝗢𝗺𝗮𝗿𝗰𝗵𝘆 in sync
▸ 10 Herald OS: 601
Which repo should be on this list?
Luca Perugini retweeted
Ecco, tipo. Chiudevo il video di ieri con questa cosa…
Luca Perugini retweeted
Imagine a household robot that reliably gives your family 10 hours back each week.
That's 520 hours a year. THAT'S the robot economy people will care about.
Luca Perugini retweeted
Security release pre-alert:
We will release new versions of Node.js 22, 24, and 26 release lines on or shortly, after 2026-11-11 in order to address security vulnerabilities.
The highest severity issue fixed in this release is HIGH.
nodejs.org/en/blog/vulnerabi…
Luca Perugini retweeted
History will show that no one did more to promote left-wing ideas than Donald Trump, and no one did more to advance the Palestinian cause than Benjamin Netanyahu.
Luca Perugini retweeted
Io sto ancora aspettando l’invito da parte di @RaiTre e @FedericoRuffo per una puntata RIPARATRICE in TV [che mi spetta di diritto] dopo che hanno consentito a questo gran capo sindacalista romano dei tassisti @BittarelliTW di insultarmi e umiliarmi.
@RaiTre @RaiNews @RaiDue @RaiUno chi si decide prima di riparare a quella puntata VERGOGNOSA?
(Mi aiutate, amici miei, a bombardare di mail e richieste alla @Raiofficialnews per farmi dare il diritto di replica?
O la RAI preferisce non disturbare la lobby dei tassisti così ben coccolata dal governo?
Letterina da parte del mio avvocato a questo “gran capo” dei tassisti romani e non solo.
I colleghi italiani devono ringraziare questi tipi di attacchi personali (a Bologna me li fanno da anni conditi da saluti fascisti alla stazione) perché mi caricano sempre di più
#Bologna5🚕
Luca Perugini retweeted
“Per la prima volta abbiamo surrogato, o meglio deposto, la nostra libertà a servizi esterni, non per forza, non per coercizione, ma per uno scambio tra libertà e comodità.” A dirlo è padre Paolo Benanti, francescano, teologo ed esperto di etica delle tecnologie.
Il punto di partenza è la storia di Nicolas Guillou, giudice francese della Corte Penale Internazionale. Dopo le sanzioni decise dagli Stati Uniti gli sono stati chiusi i profili su Amazon, Airbnb e PayPal, e Expedia gli ha cancellato perfino le prenotazioni di alberghi in Francia. Benanti ha raccontato il caso nel libro “Il caso Guillou. Storia di un’identità cancellata”. È una vicenda che riguarda anche noi, perché in Europa circa 70 euro su 100 spesi per il cloud vanno a tre colossi americani. Sugli stessi computer girano i programmi di intelligenza artificiale che le aziende stanno mettendo al lavoro.
A “Intelligenze Artificiali, in mezzo a noi” su @MedInfinityIT ne ho parlato con lui chiedendoci quanto della persona resti davvero nostro quando la nostra identità vive dentro servizi che altri possono spegnere. Poi chi c’è dietro ai lavoratori virtuali, cosa significa davvero tenere l’uomo dentro il ciclo delle decisioni delle macchine, e cosa può fare l’Europa con l’algoretica per scrivere almeno una parte dei valori che finiscono negli algoritmi.
Il video completo dell’intervista è già disponibile su Mediaset Infinity.
👉 link.mgpf.it/1237I
#IntelligenzeArtificiali #TgCom24 #IntelligenzaArtificiale #SovranitàDigitale #Algoretica
Luca Perugini retweeted
5,5 centesimi su ogni gigabyte che entra in una casa francese.
L’hanno votata in 11 deputati su 577.
Sulla carta pagano Netflix e YouTube. Ma alla fine chi paga davvero?
In Corea del Sud una legge simile c’è dal 2016, e sappiamo già com’è andata a finire.
Il video completo è su Ciao Internet 👉 youtube.com/watch?v=ePX6b7GQ…
#francia #tassa #internet #netflix #techpolicy #ciaointernet
Luca Perugini retweeted
Donzelli, ma non le sembra di esagerare? Avete liberato al Masri, avete ricevuto Chico Forti con tutti gli onori, avete congelato le chat di Del Mastro con un referente dei clan. E non dite una parola sulle bombe quotidiane a Roma est. Davvero pensate che siamo tutti imbecilli?
Luca Perugini retweeted
“The Transition”: the period of time during which the global economy and social structure will be reinvented as a result of SI. 2027-2032. it will be rough.
Luca Perugini retweeted
So @Ubuntu still hasn't patched this open hole on any Ubuntu server on the public internet that lets anyone take it over
We might just have to switch all our servers over to Debian in this new post-AGI time of rapid cybersecurity hacks which require just as rapid patching
Slowpokes will be left behind
chalmovsky.com/posts/wtf-ubu…
This is why I said you want your VPS ports all closed
Then only let it be accessible inside your Tailscale for SSH and via a Cloudflare Tunnel to receive web traffic
Any Linux box can be taken over with just a single TCP port exposed to the internet now
A stop gap is disabling MPTCP on your VPS now but even better is not having any public ports exposed!