LLM explored: breaking down common delusions
Everyone’s tried ChatGPT already, coming away with vastly different feelings, some positive, others negative. But, however, users mostly agree that it feels quite monumental, and for good reason.
Just imagine:
- OpenAI ChatGPT 4 models contain 1.8 trillion parameters
- By comparison, our brain has 86 billion neurons – that’s 20 times less
- The chatbot was trained on trimming 570 gigabytes of data
- For reference, Wikipedia has 22 gigabytes of data – that’s 25 times less
- As stated by Forbes, the price of training the model was about $40 million
- This number assumes own amortized hardware – the costs for the public cloud are higher
- The chatbot requires some serious hardware to run, a comprehensive 128 A100 GPU cluster
- For illustration, each stand is worth $1 million and includes five racks, 8 A100 GPU each
These nuances alone discourage most organizations from even considering designing a comparable LLM model. It would be strange if this were not the case.
Sure thing, ChatGPT feels quite clever, even thoughtful: the dialogues feel natural, the responses are accurate. The key question becomes: how can business leaders use existing LLMs effectively if it’s just a cloud service, only capable of reproducing common facts?
You might consider fine-tuning existing models to understand your operations and generate aligned responses. But mind, this won’t actually make the chatbot any “smarter” – it will just adapt in terms of style and context.
Another option – you could try exploring smaller models, which align with both your data center and finances. But still, these often fall short in showing “general intelligence”, even with extensive customization.
So, in the end, how can you make LLMs smarter?
LLM scale
To simplify, LLMs are just models that process and finish text excerpts you enter in the best way that’s possible. A very crucial characteristic here is context length – the size of the text excerpt LLMs are capable of receiving (the number of tokens the model can process in an input sequence).
To demonstrate the difference between models:
| ChatGPT model | Context lengths in tokens | Approximate number of pages |
| GPT-3.5 | 4,096 | 6 |
| GPT-4 | 8,192 | 12 |
| GPT-3.5-Turbo | 16,385 | 24 |
| GPT-4o | 128,000 | 190 |
LLMs use context length (or windows) to process a chunk of dialogue history alongside the user’s new request. While being very powerful, LLMs don’t retain memory of previous interactions unless the user clearly includes that context.
The chatbot can answer questions intelligently, but immediately forgets everything right after the conversation. Despite this, LLMs are often overestimated, with most users expecting human behavior.
But when talking about human behavior, what exactly is missing?
LLM intelligence and limitations: key concepts
No “common context development”
Human beings (and animals) don’t just exchange information when interacting – they synchronize their minds. This incidence makes conversations more efficient, builds understanding, and facilitates intellectual evolution.
LLMs are one-way streets – they listen and respond, but don’t ask questions to clarify or deepen their insight. Even when they’re forced to engage in the described ways, you cannot successfully incorporate that cognition into the LLM’s “mind” – they’re static.
Imagine trying to explain yesterday’s meeting to colleagues who forget every detail the moment you finish – quite frustrating, isn’t it?
No physical world awareness
Human intelligence doesn’t evolve by itself – we learn by touching, seeing, hearing, and exploring the world. From interpreting body language to scanning spatial dimensions, our senses are anchored in reality and impact our decisions.
LLMs, however, don’t have such awareness – in their textual dimension, there are no senses they’d explore. They can’t think dimensionally, read cues like slightly raised eyebrows, and grasp practical physics.
It’s almost like trying to explain the laws of gravity to someone who has never seen apples fall from trees.
Dreams of electric sheep? Not yet.
Dreaming isn’t about pastime – it’s how humans consolidate their memories, process impressions, and grow. During the nighttime, our minds are sieving the events lived through and integrate new knowledge and experiences into our mental framework.
LLMs, however, don’t dream, they wake up smarter when retrained – a slow-moving, resource-heavy process. Dreaming is like tidying your desk each night, neatly organizing what’s important and discarding what’s not, while an LLM needs the desk completely replaced to incorporate anything new.
Abto Software’s LLM experts single out AGI (Artificial General Intelligence) as one of the most important, outright indicators.
Collective emergence
Humans thrive within groups – we learn from others, we grow on shared, collective knowledge and experience. This interconnection fosters innovation and progress, thus driving us towards greater intelligence and qualities.
To reach greater potential, LLMs need similar interconnections, as now they’re disconnected, siloed machines. They’re unable to improve without programmers and even with millions of conversations with users or even other models.
So, interconnected LLMs could maybe someday evolve into ASI (Artificial Super Intelligence).
Personal assistants: a redefined, smarter approach to management
LLMs hold strong promise as intelligent personal assistants to manage small chores, from emails to bookings. But, however, we would expect these personal assistants to recall previous conversations without mistakes, which poses some challenges.
Storing dialogues, making them easily searchable, merging previously retrieved information for interactions… How can we achieve such results and make ChatGPT even more useful?
Suggested solutions:
- Memory-Augmented Models:
- to allow the assistant to retain historical interactions and access corporate documentation
- and ensure contextually relevant, seamless responses by simulating human-like memory
- Retrieval-Augmented Generation (RAG technique):
- to retrieve external information during conversations
- and provide informed answers, precisely tailored to specific user needs
- Tool-Use Agents:
- to allow the assistant to interact with other external systems (email, phone, booking services)
- to enable proactive management, thus turning the assistant into an action-oriented resource
Promising, right?
Recommendation systems: guiding end-users to success
One of the drawbacks of trained LLM systems is the inability to ask questions (or lacking shared context). While skipping philosophical debates about consciousness, intent, motivation, and other human capabilities, it’s clear that even this basic conversational skill isn’t part of their default design.
Let’s discuss a real-world use case: an assistant for students to help them choose the most suitable college.
Consulting agencies, for instance, help students better navigate their options when choosing between colleges. By analyzing their scores, personal interests, and other important factors (location, courses, and opportunities), those firms can provide tailored advice.
On the other hand, many students are turning to public LLM systems, ChatGPT particularly, to get faster help. But, obviously, these experiences fall short – LLM systems, while excelling at providing general information, can’t access latest admissions, specific requirements, and other nuanced criteria.
Suggested solutions:
- Multimodal Models: to parse text from diverse formats (images, tables, scanned documents)
- In-Context Learning: to teach the model to ask follow-up questions by guiding the response
- Prompt Engineering: to craft input prompts to improve the quality and accuracy of responses
By combining these tools, artificial intelligence can simulate a human consultant’s ability to ask right questions, guiding students through their life decisions with personalized, up-to-date advice.
How we can help
Let’s improve your products and enhance your operations with the vast capabilities of large language models. We integrate, tune, train, and customize – no matter your industry, project scale and peculiarities, and more – to meet your unique business situation.
Abto Software doesn’t just “do requirements” – we examine your needs and suggest game-changing solutions.
Our services:
- LLM strategy and consulting
- LLM development and integration
- LLM fine-tuning
- LLM support and maintenance
Our expertise:
- Artificial intelligence
- Machine learning
- Deep learning
- In-context learning
- Prompt engineering, and more
- Computer vision
Leveraging large language models completely friction-free.