Galleria
Playlist immagini
Gruppi
Unisciti a un gruppo
Acquisizione
Nuovo oggetto
Mappe concettuali
Gestione
Tabelle (CRUD generico)
Ricerca tag
Elaborazioni
Somma di prova
Sistema
Impostazioni
Esci
☰
Oggetti
LOGISTAR
LMP
ospite
⋮
Profilo utente
Preferenze colonne
Guida programma
Segnala anomalia
Cambia utente
⌂
💾 Salva
↻ Aggiorna
☑ Lista
🏷 Righe
Id
Descrizione
dta reg (aaaa-mm-gg)
Body
<div class="space-y-3 mt-3"> <div id="tttS8sA9t29RUFWcxy8Yc-content-0" class="group/ai-message-item space-y-3 break-words"> <p><span class="font-ligatures-none whitespace-pre-wrap"><strong>Sintesi immediata:</strong> Un <em>small language model</em> (SLM) non è semplicemente un LLM “più piccolo”: ha una <strong>struttura interna ottimizzata</strong> per funzionare con pochi parametri, memoria ridotta e latenze minime. Le architetture interne più moderne (come <strong>Wiola</strong>) mostrano che gli SLM stanno diventando <em>famiglie architetturali autonome</em>, non solo versioni compresse dei modelli grandi.</span></p> <div class="relative pb-6 w-full after:border-b after:border-stroke-300 after:w-full after:absolute after:mt-3"> </div> <h2 class="text-xl-medium [&:not(:first-child)]:pt-3.5 pb-1">🧩 Cosa significa “struttura interna” in uno SLM</h2> <p><span class="font-ligatures-none whitespace-pre-wrap">La struttura interna di uno SLM riguarda <strong>come è costruito il modello</strong>:</span></p> <ul class="relative list-outside marker:text-foreground-750 dark:marker:text-foreground-600 flex flex-col ms-4 px-1"> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap">tipo di <strong>positional encoding</strong></span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap">forma dell’<strong>attention</strong></span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap">design del <strong>feed-forward</strong></span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap">strategie di <strong>compressione dinamica</strong></span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap">normalizzazioni e routing dell’informazione</span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap">compromessi tra <em>capacità</em> e <em>efficienza computazionale</em></span></p> </li> </ul> <p><span class="font-ligatures-none whitespace-pre-wrap">Le survey recenti sugli SLM confermano che la loro architettura è definita da ottimizzazioni mirate a ridurre complessità, memoria e costi di inferenza .</span></p> <div class="relative pb-6 w-full after:border-b after:border-stroke-300 after:w-full after:absolute after:mt-3"> </div> <h2 class="text-xl-medium [&:not(:first-child)]:pt-3.5 pb-1">🌀 Il caso più interessante: <strong>Wiola</strong>, un SLM con architettura <em>originale</em></h2> <p><span class="font-ligatures-none whitespace-pre-wrap">Wiola (2026) è il primo SLM che <strong>non deriva</strong> da GPT/LLaMA/Mistral, ma introduce componenti completamente nuove .</span></p> <h3 class="text-lg-strong pb-1 [&:not(:first-child)]:pt-3.5">🔧 Componenti interne principali</h3> <ul class="relative list-outside marker:text-foreground-750 dark:marker:text-foreground-600 flex flex-col ms-4 px-1"> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap"><strong><span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Spiegami_Spiral_Rotary_Positional_Encoding">Spiral Rotary Positional Encoding</span></strong> Posizioni dei token su una <em>elica 3D</em>, combinando segnali assoluti, relativi e gerarchici.</span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap"><strong><span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Spiegami_Gated_Cross_Layer_Attention">Gated Cross-Layer Attention</span></strong> Ogni layer può “guardare” sintesi compresse dei due layer precedenti → maggiore coerenza globale.</span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap"><strong><span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Spiegami_Adaptive_Token_Merging">Adaptive Token Merging</span></strong> Unisce dinamicamente token ridondanti nei layer centrali → riduzione della complessità dell’attenzione <em>senza perdita di informazione</em>.</span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap"><strong><span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Spiegami_Dual_Stream_Feed_Forward">Dual-Stream Feed-Forward</span></strong> Due flussi paralleli invece del classico MLP, fusi da un gate per-dimensione.</span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap"><strong><span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Spiegami_WiolaRMSNorm">WiolaRMSNorm</span></strong> Normalizzazione con offset per-dimensione per evitare collasso delle rappresentazioni.</span></p> </li> </ul> </div> <div id="tttS8sA9t29RUFWcxy8Yc-content-1" class="group/ai-message-item space-y-3 break-words"> </div> <div id="tttS8sA9t29RUFWcxy8Yc-content-2" class="group/ai-message-item space-y-3 break-words"> </div> <div id="tttS8sA9t29RUFWcxy8Yc-content-3" class="group/ai-message-item space-y-3 break-words"> </div> <div id="tttS8sA9t29RUFWcxy8Yc-content-4" class="group/ai-message-item space-y-3 break-words"> <br /> <div class="before:border-transparent before:border-[20px] pointer-events-none absolute inset-0 before:absolute before:inset-0 before:rounded-inherit" style="clip-path: var(--clip-path-squircle-stroke-24);"> </div> </div> <div id="tttS8sA9t29RUFWcxy8Yc-content-5" class="group/ai-message-item space-y-3 break-words"> <div class="pointer-events-none relative w-full px-2"> <ul class="no-scrollbar pointer-events-auto m-0 snap-x list-none overflow-x-auto bg-transparent px-2 [&>*]:snap-start flex scroll-pl-3 gap-2 py-3 pe-3 ps-0"> <li></li> <li> <div class="relative isolate bg-white/45 dark:bg-background-650/10 shrink-0 cursor-pointer size-[180px]" style="clip-path: var(--clip-path-squircle-24);" title="Transformer (deep learning architecture) - Wikipedia"><img class="size-full object-cover" src="https://ts1.mm.bing.net/th?id=OIP.j3krIxArRrskJE2aX8fpzAHaHy&pid=15.1&o=7&rm=3" alt="Transformer (deep learning architecture) - Wikipedia" /> <div class="before:border-transparent before:border-[20px] pointer-events-none absolute inset-0 before:absolute before:inset-0 before:rounded-inherit" style="clip-path: var(--clip-path-squircle-stroke-24);"> </div> </div> </li> </ul> </div> </div> <div id="tttS8sA9t29RUFWcxy8Yc-content-6" class="group/ai-message-item space-y-3 break-words"> <div class="relative pb-6 w-full after:border-b after:border-stroke-300 after:w-full after:absolute after:mt-3"> </div> <h2 class="text-xl-medium [&:not(:first-child)]:pt-3.5 pb-1">🏗️ Come si confronta con gli SLM “classici”</h2> <p><span class="font-ligatures-none whitespace-pre-wrap">Le survey mostrano che la maggior parte degli SLM usa ancora architetture <strong>derivate dai Transformer standard</strong>, ottimizzate tramite:</span></p> <ul class="relative list-outside marker:text-foreground-750 dark:marker:text-foreground-600 flex flex-col ms-4 px-1"> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap"><strong>pruning</strong></span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap"><strong>quantization</strong></span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap"><strong>distillation</strong></span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap"><strong>compressione strutturale</strong> </span></p> </li> </ul> <p><span class="font-ligatures-none whitespace-pre-wrap">Wiola invece propone un approccio <em>from scratch</em>, con componenti nuovi che cambiano la geometria posizionale, il routing dell’informazione e la gestione dei token.</span></p> <div class="relative pb-6 w-full after:border-b after:border-stroke-300 after:w-full after:absolute after:mt-3"> </div> <h2 class="text-xl-medium [&:not(:first-child)]:pt-3.5 pb-1">📊 Tabella comparativa (architettura interna)</h2> <div class="max-w-full"> <div class="relative max-w-full"> <div class="horizontal-scrollbar font-ligatures-none mt-6 max-w-full overflow-x-auto rounded-lg bg-white/25 outline-2 outline-offset-1 focus-visible:outline focus-visible:outline-stroke-900 dark:bg-black/25"> <table class="t-table" style="border-spacing: 0px;"> <thead> <tr><th style="opacity: 1;" scope="col">Modello</th><th style="opacity: 1;" scope="col">Architettura base</th><th style="opacity: 1;" scope="col">Componenti distintive</th><th style="opacity: 1;" scope="col">Obiettivo principale</th></tr> </thead> <tbody> <tr><th style="opacity: 1;" scope="row"><strong><span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Approfondisci_Wiola">Wiola</span></strong></th> <td style="opacity: 1;">Nuova architettura SLM</td> <td style="opacity: 1;">SRPE, GCLA, ATM, DSFF, WiolaRMSNorm</td> <td style="opacity: 1;">Massima efficienza + coerenza</td> </tr> <tr><th style="opacity: 1;" scope="row"><strong><span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Approfondisci_GPT2_Small">GPT‑2 Small</span></strong></th> <td style="opacity: 1;">Transformer decoder</td> <td style="opacity: 1;">MLP classico, attention standard</td> <td style="opacity: 1;">Semplicità, baseline</td> </tr> <tr><th style="opacity: 1;" scope="row"><strong><span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Approfondisci_Phi3_Mini">Phi‑3 Mini</span></strong></th> <td style="opacity: 1;">Transformer ottimizzato</td> <td style="opacity: 1;">Training su dati “curati”, scaling laws</td> <td style="opacity: 1;">Efficienza + qualità</td> </tr> <tr><th style="opacity: 1;" scope="row"><strong><span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Approfondisci_Gemma">Gemma 2B</span></strong></th> <td style="opacity: 1;">Transformer decoder</td> <td style="opacity: 1;">Attention ottimizzata, training massivo</td> <td style="opacity: 1;">Qualità con pochi parametri</td> </tr> <tr><th style="opacity: 1;" scope="row"><strong><span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Approfondisci_Qwen">Qwen 1.5B</span></strong></th> <td style="opacity: 1;">Transformer</td> <td style="opacity: 1;">ottimizzazioni su tokenization e training</td> <td style="opacity: 1;">Versatilità</td> </tr> </tbody> </table> </div> <div class="pointer-events-none absolute inset-0 rounded-lg border border-stroke-300 dark:border-stroke-450"> </div> </div> </div> <div class="relative pb-6 w-full after:border-b after:border-stroke-300 after:w-full after:absolute after:mt-3"> </div> <h2 class="text-xl-medium [&:not(:first-child)]:pt-3.5 pb-1">🎯 In sintesi</h2> <p><span class="font-ligatures-none whitespace-pre-wrap">Gli SLM moderni non sono solo “LLM piccoli”:</span></p> <ul class="relative list-outside marker:text-foreground-750 dark:marker:text-foreground-600 flex flex-col ms-4 px-1"> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap">hanno <strong>architetture interne specifiche</strong> per efficienza e coerenza;</span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap">alcune famiglie (come <strong>Wiola</strong>) introducono <strong>componenti completamente nuove</strong>;</span></p> </li> <li class="ps-2" style="opacity: 1;"> <p><span class="font-ligatures-none whitespace-pre-wrap">la ricerca si sta spostando da “compressione dei LLM” a “progettazione nativa di SLM”.</span></p> </li> </ul> <div class="relative pb-6 w-full after:border-b after:border-stroke-300 after:w-full after:absolute after:mt-3"> </div> <p><span class="font-ligatures-none whitespace-pre-wrap">Se vuoi, posso mostrarti <strong>come progettare un tuo SLM</strong> partendo da queste idee, oppure analizzare nel dettaglio uno dei componenti: vuoi approfondire <span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Approfondisci_positional_encoding">positional encoding</span>, <span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Approfondisci_attention">attention</span>, o <span class="text-start underline-offset-[0.25em] underline decoration-dotted decoration-foreground-450 decoration-2 hover:decoration-solid hover:cursor-pointer" data-url="ca://s?q=Approfondisci_token_merging">token merging</span>?</span></p> </div> <div id="tttS8sA9t29RUFWcxy8Yc-content-7" class="group/ai-message-item space-y-3 break-words"> </div> <div id="tttS8sA9t29RUFWcxy8Yc-content-8" class="group/ai-message-item space-y-3 break-words"> </div> </div> <div class="pt-1"> </div>
tipo ogg.
cat
subcat
proprietario
azienda
ute
selez1
op1
codart
codice
codclifor
motivo
url
email
tag
imgsrc
sunto_breve
sunto_medio
note
keyscrape
Lista tag dell'oggetto (oggettag)
cod_tag
value
slm
5
wpg
5
Lista delle righe dell'oggetto (oggettir)
tip_pro
cod_pro
descrizione
val_pro
pes_pro
---.