Con che linguaggi l’hai provato? Avevo letto che con Rust non se la cavasse benissimo.
Quanto costa poi?
Tencent releases open-source 3D world generation model that enables you to generate immersive, explorable, and interactive 3D worlds from just a sentence or an image
Si l’ho visto. Tra l’altro i modelli non pesano tanto. Devo provarlo in VR 
c’è roba direttamente 0er ve Generative AI Overview | Horizon Worlds
Chiedo qua che c’è gente più skillata di me e magati mi date l’imbeccata giusta.
Sto provando Librechat con le API di OpenAi per usarlo all’interno dell’azienda.
Sono però fermo su una cosa che non capisco:
volevo dargli in pasto dei documenti da elaborare che poi siano disponibili a tutti. Pensavo fosse sufficiente caricare il PDF sullo storage di Openai ma così non va.
Qualche guida che riesca a spiegarmi come posso fare ?
modo semplice semplice è leggere il pdf e metterglielo nel prompt
“puoi accedere a pdf_document per rispondere alle domande degli utenti.”
No purtroppo non funziona sembra
ma stai facendo tutto tramite api o tramite dashboard openai crei l’agente etc
io ho fatto così in node fino adesso e funziona. sei limitato dal contesto quindi non puoi dargli 100 pagine di pdf.
cmq chiedi a chatgpt come usare chatgpt
Ho esaurito lo spazio disponibile in una chat ChatGPT.
Ho quindi esportato intero testo in pdf (mi pare sulle 600 pagine) e importato in nuova chat il file e detto di leggerlo.
Mi dice che lo ha fatto ma poi faila alla grandissima proprio, si inventa di sana pianta cose che non ci sono assolutamente scritte e/o svincola alla grande.
Come è possibile?
Sembra uno studente che non ha aperto il libro e si arrampica sugli specchi
It’s Game Over: The Real “AlphaGo Moment” Just Happened in a Chinese AI Lab
A new paper just revealed an AI that invents its own technology without us. We are not ready for what happens next.
Alright, let’s have a serious talk. We often think of AI breakthroughs as things that happen to us. A new chatbot, a new image generator. But what if the biggest breakthrough isn’t a new tool, but a new toolmaker? An AI that can invent entirely new kinds of AI, all on its own.
I touched on this idea before, but after digging deeper into a recent paper, “ AlphaGo Moment for Model Architecture Discovery,” I realized it’s a much bigger deal than I first thought.
This isn’t just about an AI being creative. It’s about an AI system that runs the entire scientific method from start to finish: hypothesizing, coding, experimenting, and analyzing.. to design its own brain. And its “thought process” is leaving a trail of evidence that is both brilliant and deeply humbling.
So let’s go beyond the headlines and look under the hood of ASI-ARCH. Because how it works is the real story here.
How an AI Plays God: The ASI-ARCH Framework
First, forget the old way of doing things. Previously, AI architecture search was like giving an AI a box of pre-designed Legos and asking it to find the best way to assemble them.
ASI-ARCH doesn’t use our Legos. It invents its own.
To do this, the system is designed like a miniature, autonomous AI research lab. It has four key parts that work in a closed loop:
- The Cognition Base (The Library): This is the AI’s connection to us. The researchers fed it nearly 100 of the most important human-written papers on AI architectures. So, the AI starts by “standing on the shoulders of giants,” learning from all of humanity’s best ideas.
- The Researcher (The Creative Genius): This AI agent reads from the Cognition Base and looks at its own past experiments. Its job is to come up with a novel idea, a new hypothesis for a better architecture. It’s the “what if” engine.
- The Engineer (The Hard Worker): This is the workhorse. It takes the Researcher’s idea and has to write real, working Python code. If the code has bugs, throws an error, or is too slow during training, the Engineer has to debug its own code. No human intervention. It keeps trying until it has a functional model. Well, well… That’s how real research works.
- The Analyst (The Wise Mentor): After an experiment is done, the Analyst gets all the data. It looks at performance, training speed, and everything in between. But here’s the cool part: it compares the new model to its “parent” and “sibling” models to figure out exactly which change led to an improvement or failure. It provides the deep insights that fuel the Researcher’s next big idea.
This entire loop runs 24/7. An idea becomes code, code becomes an experiment, and the results become the seed for the next idea. It’s evolution in realtime. Honestly, I’m not sure how to describe it… but as an optimist, I find it absolutely fascinating to read this paper.
But How Does It Know What’s “Good”?
This is where it gets really interesting. How do you judge art? How do you judge creativity? ASI-ARCH’s creators knew that just chasing a high score on a benchmark would be a disaster. That leads to “reward hacking”, finding cheap tricks to boost a number without creating a genuinely better design.
So, they built a “Fitness Function” with two parts:
- Objective Performance: This is the raw data. How well did it perform on tasks? What was its training loss? (The boring stuff).
- Architectural Quality: This is the wild part. They use a separate LLM to act as an expert judge. This “LLM-as-Judge” looks at the new design and scores it on things like innovation, structural complexity, and elegance.
So, an AI is literally judging another AI’s creativity.
Weird times.. :)
This stops the system from just making models bigger or more complicated. It has to create designs that are not just powerful, but also smart.
We can actually see this evolutionary process happen. The researchers mapped it out in what they call an “exploration trajectory tree.” Each dot is a new AI it invented, and the lines show how they evolved from one another.
Looking at that tree, you’re not just seeing data points. You’re seeing the AI’s mind at work, branching out, exploring dead ends, and converging on brilliant new paths.
The Results: Alien Genius and a Terrifying New Law
So, did it work?
In a massive understatement: yes.
After 1,773 autonomous experiments over 20,000 GPU hours, ASI-ARCH discovered 106 completely novel, state-of-the-art architectures.
These weren’t just slight improvements. They contained “emergent design principles” that human researchers hadn’t thought of. Just like AlphaGo’s “Move 37” felt alien to Go masters, these architectures feel alien to AI designers. They introduce concepts like “Content-Aware Sharpness Gating” and structures that mimic a “Mixture of Experts (MoE)” without being explicitly told to.
But the most important discovery wasn’t an architecture. It was a graph.
This graph plots GPU hours (money spent) against the number of SOTA architectures discovered (scientific breakthroughs). And the line goes straight up.
This is the “scaling law for scientific discovery.”
It proves, for the first time, that the pace of innovation in this field is no longer limited by human creativity. It’s limited by computational budget. Want more breakthroughs? Just add more GPUs. This transforms research from a human-bound process to a computation-scalable one. Honestly, that’s a huge shift, one that could reshape our entire civilization.
Where Do Good AI Ideas Actually Come From?
This was the question that kept nagging me. Okay, the AI is a genius. But where is its genius coming from? Is it just cleverly remixing human ideas, or is something deeper happening?
The researchers tracked this. They classified the origin of every new idea into one of three buckets:
- Cognition: The idea was directly inspired by a human research paper in its library.
- Analysis: The idea came from the AI’s own analysis of its past experiments, seeing patterns and drawing abstract conclusions.
- Originality: The idea was a complete shot in the dark, with no clear origin.
When they looked at all 1,773 experiments, most ideas came from “Cognition.” The AI was mostly building on our work. Makes sense.
But then they looked at only the top 106 superstar models… the “Model Gallery.” And the results flipped.
For these elite, breakthrough models, the source of inspiration shifted dramatically. They relied far more heavily on “Analysis” than the average models.
Let that sink in.
To be competent, the AI could rely on human knowledge. But to achieve true excellence, it had to rely on its own abstract understanding synthesized from its own experience. It wasn’t just copying us. It was learning to think. It learned that to make a real leap, it couldn’t just reuse past successes; it had to explore, summarize, and discover its own, more abstract principles.
So, Are We There Yet?
We’re not waiting for AGI to arrive in a press release. It’s arriving in the form of quiet, dense research papers like this one.
The “feel the AGI moment” isn’t a single event. It’s a cascade. For AI researchers, this paper is that moment. It’s the point where the student has not only surpassed the teacher but has started inventing entirely new fields of study.
We’ve built a system that accelerates its own evolution. The implications are staggering, and we’re only in the first inning. The next time a new AI model drops that seems impossibly good, the question we need to ask is no longer “Which team of humans built this?”
But rather: “Which AI designed it?”
1 Like
Alla fine ho chiesto a chatgpt come usare se stesso
In soldoni
Scriptone in python per prendere tutte le pagine della mediawiki e caricarle in Qadrant. (per intanto solo il testo)
Poi se ho capito bene rag_chain manda il tutto ad openai e lui elabora e mi risponde.
E fin qua sembra funzionare.. devo capire bene se è la strada giusta..
Ma una cosa che mi sfugge è : se i dati sono sul pc, ogni volta li invia a openai?
è un problema di contesto
non puoi dargli roba infinita e poi interrogarlo su quello o continuare a parlarne.
- Il conteggio dei token include tutto: il testo dell’utente, la conversazione precedente, le istruzioni di sistema e le risposte generate.
- Se superi il limite, il sistema automaticamente tronca le parti più vecchie della conversazione per rientrare nel massimo consentito.
- L’output generato dal modello conta anch’esso nei token totali disponibili (prompt + completamento)
il “contesto massimo” (cioè la lunghezza complessiva del testo – domande, risposte, immagini caricate, istruzioni di sistema) che il modello può considerare in una singola sessione dipende principalmente dal modello
| Modello ChatGPT | Contesto massimo |
|---|---|
| GPT‑4 standard | 8.192 token (~6.000 parole) |
| GPT‑4 Turbo / GPT‑4o | 128.000 token |
| GPT‑4.1 / GPT‑4.1 Mini | fino a 1 milione di token |
Nel piano gratuito, normalmente si usa GPT‑3.5 o GPT‑4 Mini: contesto limitato a circa 8.000 token.
Fai conto che per provare ho preso le api a pagamento.
Adesso devo valutare se vale veramente la pena o meglio pensare a usare openllama o altro..
Grazie. A me però non cancella le parti più vecchie per poi continuare, mi dice che lo spazio è esaurito (V 4.0 abbonamento)
Min aggiunta non capisco come mai essendo modelli di linguaggio, facciano fatica ad assimilare il contenuto di un PDF di circa 600 pagine
ti distrai un attimo ed esce un nuovo super modello
Non é che fanno fatica, dipende dove vanno. Se hai un vector db dove vengono salvati tipo knowledge base e accedono a quella ok. Se li hai in memoria locale o simile é un altro
Conta anche come hai deciso di spezzettare i chunk, l’overlap e i costi che vuoi sostenere
Oh, comunque posso dire che son solo contento se 'sti influencer virtuali generati con AI toglieranno il lavoro agli influencer?
beh tu con Elgoog su Instagram sei stato un precursore, si può dire
2 Likes
Elgoog 
Sempre sia elgoogato.
fatta da un amico, l’ho letteralmente vista nascere.
comunque non toglieranno lavoro per un po