Het is wel duidelijk dat sinds eind 2022, begin 2023 de wereld verbaasd werd met de eerste grote taalmodellen waar je in real life mee kon kletsen, discussieren en, niet te vergeten, als coding buddy kon gaan gebruiken.
Inmiddels zijn we twee jaar verder, zitten we in begin 2025 en een week geleden werden de Amerikanen van ChatGPT, heer en meester van de LLMs verrast door het Chinese https://www.deepseek.com/
Probleem: Privacy…
Snel kwamen de waarschuwingen: alle gesprekken die je met deepseek online voert, betekenen dat je alles wat je zegt deelt met de Chinezen! Whoehoe! Ja dat is natuurlijk riskant, net zoals dat het al was met het dele van je zieleroerselen in een online gesprekje met ChatGPT.
Oplossing: Zelf draaien op de eigen pc
Stel dat het programma, het taalmodel nou gewoon op je eigen laptopje kan draaien, zodat het hele gesprek alleen maar lokaal gevoerd wordt? Je teksten gaan niet meer naar een server ergens in de Vervelende Staten of China, gaan zelfs niet meer naar buiten, maar alles blijft lokaal. Kan dat? Het verbluffende is dat dit tegenwoordig best kan omdat de taalmodellen zich behoorlijk hebben verbeterd. Ze zijn kleiner geworden dan eerst, denk aan een gigabyte of 1, 2, 4, in plaats van de eerdere taalmodellen die ruim 70GB aan dataopslag nodig hadden om het hele getrainde taalmodel alleen al op disk op te slaan. Dat is tegenwoordig wel anders.
Ollama: zelf draaien van het taalmodel op je PC
Wanneer je een klein beetje handig bent met je PC, en wie is dat tegenwoordig niet al een klein beetje? Kan het programma Ollama downloaden. Het programma is in staat om met simpele console commandoś een taalmodel te downloaden en te draaien. Aangezien Ollama gebruik maakt van een bepaalde standaard interface waar de grootste bedrijven aan voldoen zijn er taalmodellen van OpenAI, Mistal en ook de nieuwste Chinese modellen van Deepseek beschikbaar.
Hoe ziet dat er uit?
Laten we zeggen dat deepseek nogal een kletsmajoor is. Wat wel erg leuk is, is dat het model eerst even zit te prakkiseren over wat het zou kunnen gaan antwoorden. Dat wil echt zeggen dat we het ¨nadenken” van het model zelf kunnen volgen. Het nadenken zit dan tussen <think>…</think> tags in, voordat het model het uiteindelijke antwoord geeft.
Zork volgens deepseek

Zoals je hierboven ziet, heeft het (kleine) model van deepseek wat moeite met de vraag, en bleef net als een langspeelplaat eerst even hangen. Vervolgens verzon het wat over een spel, maar met het originele Zork, zie: (https://nl.wikipedia.org/wiki/Zork) had het niet veel te maken. Gemaakt in 2017 door deepseek, in Japan met hoofdkarakter Tosh? Wel creatief, dat wel.
Kan het beter?
De vraag stellen is ‘m beantwoorden. Het zou natuurlijk wel leuk zijn als we eerst wat meer kennis zouden kunnen geven aan het taalmodel, zodat het taalmodel zich een beetje kan inlezen in de materie. Dan zou die veel beter zijn in het beantwoorden van vragen over een onderwerp. En wat blijkt? Dit is gewoon mogelijk. We kunnen een taalmodel een document, of zelfs meerdere documenten geven om een beetje door te nemen. Dit werkt nog helemaal niet perfect, maar het werkt wel een beetje. Met een mooi woord heet dit RAG, Retrieval Augmented Generation
Retrieval Augmented Generation (RAG) is een geavanceerde techniek binnen het veld van kunstmatige intelligentie en natural language processing (NLP) die de manier waarop modellen informatie genereren en vragen beantwoorden, enorm verbetert.
Een RAG draaien op je laptop?
Ja ho even. De techniek staat voor niets, maar we hebben het hier wel over super nieuwe technieken. Artificieele Intelligentie! Neurale netwerken, NLP, Natuurlijke Taal Modellen en LLM, Grote Taal Modellen en dan ook nog “Retrieval augmented Generation¨. Toe maar. We gaan het toepassen met Pyton, de programmeertaal die de meeste beschikbaarheid heeft als het gaat om AI, en we gebruiken het voorbeeld programma waarbij we een PDF bestandje kunnen uploaden naar het programma, zodat het taalmodel daar dan weer gebruik van kan gaan maken. Het programma dat we gebruiken is te vinden op de volgende github site:
https://github.com/fahdmirza/doclingwithollama

En er is ook een videolink op die pagina waarbij de auteur uitlegt hoe het zo ongeveer werkt. Waar het op neer komt is dat je PDFjes die je gevonden hebt, het systeem in kan laden en wanneer je dan een gesprek voert, dan babbel je in wezen tegen je documenten aan.
De Zork walkthrough
We draaien het programma volgens de bovenstaande instructies van Docling with Ollama en krijgen een mooie webinterface, omdat er een lokale HTTP server naar onze acties zit te luisteren.
Via de knop “Browse files” uploaden we de Zork Walkthrough PDF die we ergens online gevonden hebben, zodat het taalmodel daar gebruik van kan gaan maken in onze chat.

Ja, en nu maar eens vragen hoe we kunnen beginnen.
Deepseek met een beetje Zork context
De laptop waar ik zit te typen is ongeveer 10 jaar oud, en het uploaden van een PDF bestand met 44 pagina’s is iets teveel voor Docling om het met vectors te parsen in het minieme NVIDIA onboard kaartje, een sukkelige M1000M Nvidia driver.
We breken het PDF document daarom op per pagina, en uploaden pagina twee, die verteld hoe we in het spel Zork langs de Troll en de Cyclops moeten komen. Laten we de AI maar eens vragen of die de context uit de pagina een beetje begrepen heeft door te vragen
What should be done with the troll?

Dat screenshot is een beetje klein, dus even kopieren.
Hier is het antwoord dat deepseek-r1:1.5b (het kleinste model) op mijn 10 jaar oude laptop heeft uitgespuugt. Let vooral op de <think>…</think> tags voordat het definitieve antwoord verschijnt!
<think> Okay, so I’m trying to figure out what to do with the troll from this context. Let me go through the provided information step by step.
First, I see that the user wants me to think through the problem and come up with a detailed answer. They also mentioned that if I don’t know the answer, I should say “I don’t know!” So my main task is to process all the given context about the troll and determine its role or interaction within the game.
Looking at the context provided, it’s quite lengthy but organized into sections. Let me try to parse each part methodically.
The first section talks about Go down the chimney from the Kitchen. The user then asks how to handle the troll. It mentions giving various things to the troll, especially the nasty knife from the Attic. I’m guessing that the key is a weapon or something destructive, but it’s not clear if it’s offensive or just a tool.
Next, there are instructions on Knocking the troll unconscious and returning his axe after he wakes up. This seems like a battle scenario where the troll acts as a character in a fight. The user then wants to say hello, listen to, and take the troll (and Cyclops). So maybe the troll is an active entity that needs human interaction.
The attack part says Knock the troll with the sword, repeating until it’s killed. That suggests that the troll can be targeted by the sword. Then, after killing him, drop the sword and return it. The user also mentions saving the game so you can retry the battle.
Moving on to another section: attacking the troll with a sword. It says the same way the Thief is to the knife. So maybe the thief uses the knife, which is destructive or harmful, whereas the sword is more gentle but still effective against the troll.
Then, the user wants to go west, east, south, and southeast to the Cyclops Room. There’s a choice of two solutions: saying “Odysseus” to scare him away or giving the lunch and water for cyclops. The first option leads them back through destroyed walls to the house, while the second requires going another way and possibly restoring the skeleton key.
If they said “Odysseus,” they can exit through a wall; go east twice to get back to the living room. If they gave the lunch and water, they need another path. So the user is choosing between two routes: one that might not save them but allows access to the house, or a safer route with potential for restarting.
In either case, after exiting the Great Underground Empire, the trap door in the Living Room won’t close behind them. They have to open the trophy case and put coins in it. Also, drop the key and water there if they didn’t use it already. Then keep hold of a brown sack (which is a container) and garlic.
Once out, go south, east to the Gallery. Get the painting, then west, north, up to the Living Room, then put it in the trophy case.
Putting this all together, I need to figure out what role the troll plays. From the context, it seems like the troll is a character that can be targeted, either by the sword or the knife. There’s also mention of the thief (who uses the knife) and how the troll reacts when approached by the knife. The sword might be used against the troll as well.
But since the user specifically asked about handling the troll, I think they want a specific action or method. From the context, it seems like there are instructions on how to interact with the troll using both the knife (the thief) and the sword. However, there’s no direct instruction on what exactly should be done with the troll; instead, it’s more about dealing with the threats posed by the troll.
But perhaps the user is looking for an action to perform in response, like giving a command or handling the troll physically. Given that, I might infer that the user wants guidance on how to interact with the troll, using either the knife (thief) or the sword.
Additionally, there’s mention of going into the Cyclops Room and choosing between two paths. The first leads back through walls where the troll was destroyed, while the second requires another path. This implies that the troll is something that can be targeted by a weapon, and that targetting can have consequences (like losing access to certain areas).
However, without more specific instructions from the user on what action or method to take regarding the troll, I need to consider if there’s any general advice provided in the context. From what I see, the key is used by the thief, which seems destructive but also possibly harmful. The sword might be a gentle weapon that can cause physical harm, so it could be seen as a potential threat.
Putting this together, perhaps the user wants to know how to handle the troll when approached, using both the knife and sword. They might not have a specific instruction on what exactly to do but are looking for guidance on interaction methods with the troll. </think>
To effectively interact with the troll in the game, you should consider the following actions based on the provided context:
- Thief (Killer): Use the knife from the Attic to approach and interact with the troll. This weapon is destructive but can be used to target him.
- Sword: Target the troll using a gentle sword. The sword may cause physical harm, so it’s important to use it carefully.
- Pathways: When entering the Cyclops Room, choose between two paths:
- Use the first path which might involve destroying walls where the troll was active.
- Choose another method for the second option, possibly requiring a different route or saving the key if it needs retrieval.
- Handling the troll physically: Depending on the threat posed by the knife (the thief), you may need to give commands or handle the troll with force, considering its destructive nature.
In summary, the troll can be targeted using either a knife or sword, each with different consequences depending on how they are used.
Je kunt er weinig of in het geval van deepseek heel, heel, heel erg veel van zeggen, maar dat AI, LLM en Retrieval Augmented Generation een heel, heel, heel groot tekstavontuur is, staat als een paal boven water!
Wat zou jij uploaden? Denk eens aan dat leuke boek dat je als laatste gelezen hebt, wat zou deepseek van de hoofdpersoon vinden…. ?
Leave a Reply