Know Thy Frenemy:

Understanding LLMs –
Past, Present, and
Future

Barak Shoshany
Department of Physics, Brock University
 Motivation
• Many professors and students use LLMs.
• Other talks focus on implications of LLMs.
• This talk focuses on LLMs themselves.
• Understanding LLMs better will help professors:
• Optimize use.
• Dispel misconceptions.
• Develop situational awareness.
• Incorporate in courses.
• Instruct students on proper use.
Part I: Past
 Neural networks 1/5
• Curve fitting: Find function that
approximates data.
• Minimize error.
• Example: Approximate by
polynomial.
• More parameters = better fit.
 Neural networks 2/5
• Neural network: “Extremely
sophisticated curve fitting.”
• Mathematical model of brain
(≈100 billion neurons).
• Width: # of neurons per layer.
• Depth: # of hidden layers.
• Deep network: multiple hidden
layers.
• Deeper layer = more abstract.
 Neural networks 3/5
• Each connection has a weight
(≈importance).
• Each neuron does a weighted
sum of previous neurons.
• Result passed through non-
linear activation function.
• Universal approximation
theorem: Any function can be
approximated by a neural
network with enough neurons.
 Neural networks 4/5
• Example – image to text (simplified):
• Input layer = pixels.
• Early hidden layers = edges, orientation, colors.
• Middle hidden layers = textures, motifs.
• Late hidden layers = objects, scene context.
• Output layer = text description of image.

“A kitten playing with a

ball of yarn.”
 Neural networks 5/5
• To determine the weights, we do supervised training.
• Can start with collection of image-text pairs.
• Available data: ≈100 billion pairs.
• Initial weights completely random.
• Feed image to network, get guess of text (nonsense at first).
• Measure how wrong guess is (loss).
• Back-propagation: Trace error backwards through layers, adjust
weights to correct.
• Repeat for all pairs (epoch) and multiple epochs, minimize loss.
 How LLMs work 1/5
• LLM = Large Language Model.
• Probabilistic model to predict next token in text.
• Token: letter / symbol / word / part of word.
• Example:
The cat → The cat is → The cat is fluffy
• Pre-training: Self-supervised learning on trillions of tokens (web,
books, code, etc.)
• Self-supervised: Just take existing text and hide a token.
The cat is _____ → try to predict → fluffy? → minimize loss
 How LLMs work 2/5
• After pre-training, LLM just “autocompletes”; not yet a chatbot.
Are cats fluffy? → Are cats fluffy? What about dogs?
• Instruction-tuning: Train with samples of chats:
User: Is salt salty?
Assistant: Yes, salt is salty.
• Reinforcement Learning from Human Feedback (RLHF): Humans give
“thumbs up/down” to chat replies.
• Base model turns into helpful chat/instruct model.
User: Are cats fluffy?
Assistant: Some are, some aren’t; depends on the breed.
 How LLMs work 3/5
• Current LLMs: based on transformer architecture.
• Attention heads can “look back” on all tokens (within context
window).
• “Pay attention” to earlier words, find links and connections.
• Work on all tokens in parallel – much faster than linear models.
• Can run on GPUs (Graphics Processing Units).
• Scale up easily; larger models are better.
• Latest models have trillions of weights!
• Transformers caused the AI “boom”.
 How LLMs work 4/5
• Embedding: Each token = point (vector) in very high-dim space.
• ≈10,000-30,000 dimensions.
• Naturally occurs as side-effect of training.
• Represents context-agnostic semantics.
• Transformer layers refine this to true context-aware meaning.
• Example: Embedding of “spring” represents “springness”.
• Flowers bloom in spring → season context.
• The spring absorbs shocks → object context.
 How LLMs work 5/5
• Subspaces of embedding space correspond to abstract concepts.
• “gender” axis: “she” and “he” in opposite directions.
• “plurality” axis: “dog” and “dogs” in opposite directions.
• Can do “arithmetic”: king - male + plural = queens.
• Embeddings + transformers = true mathematical representation of
language.
• Impossible to do manually!
• Original use for transformers: machine translation.
English → embeddings → French
Part II: Present
 The LLM zoo 1/6
• Understanding model types:
• Non-reasoning models: Predict next token with no prior planning.
• Reasoning models: Perform long and complicated reasoning, solve
problems at PhD/research level. More expensive.
• Big models: Trillions of weights. Expensive to use. Large knowledge base.
• Small (“mini”/”flash”) models: Fewer weights. Distilled from big models.
Cheaper. Less knowledge but good performance in key areas.
 The LLM zoo 2/6
• OpenAI models:
• ChatGPT is just the chat interface; many different models!
• Non-reasoning models:
• GPT-4o: Older model, good for fun stuff (chatting, images), free (limited use).
• GPT-4o-mini: Small model, not as good, free.
• GPT-4.1: Improved model, accessible only via API.
• GPT-4.5: Largest model, broader knowledge, higher EQ, paid only.
• Reasoning models:
• o3: Best model available, PhD+ level in all disciplines, problem solving, complex
coding; paid only. o3-pro available soon, pro plan only.
• o4-mini: Newer but smaller model, free (limited use). Do not confuse with 4o-mini!
• o4-mini-high: Thinks longer, paid only.
 The LLM zoo 3/6
• OpenAI capabilities (vary by model, not all free):
• Text: Chat, read, analyze, edit, generate, summarize, translate…
• Code: Read, edit, debug, generate, refactor, explain, write tests…
• Data (CSV / JSON / Excel / etc.): Parse, analyze, extract, visualize…
• Science/math: Explain, teach, reason, solve problems at PhD+ level…
• Image: See, edit, generate; photos, art, slides, diagrams…
• Voice: Listen, generate, “advanced voice”…
• Video: Watch, camera/screen sharing…
• Tool use: Autonomously search the web, run Python code…
• Custom instructions, memory…
• Context windows: 128K for GPT-4o, GPT-4.5, o3, o4-mini, 1M for GPT-4.1.
 The LLM zoo 4/6
• OpenAI other features:
• Deep Research: Spends up to 30 mins collecting data from dozens of
resources. Specialized early o3. Free (limited use of light version).
• Create custom GPTs that follow specific instructions. Paid only, can be
used by free.
• Create recurrent tasks that run autonomously. Paid only.
• Sora: Generate videos based on script. Paid only.
• Operator: Browse the web and perform tasks. Pro plan only.
• Plus plan: 20 USD/month, pro plan: 200 USD/month.
 The LLM zoo 5/6
• Google models:
• Gemini 2.5 Pro: Latest model, reasoning. Free (limited use).
• Gemini 2.5 Flash: Smaller model, reasoning. Free (limited use).
• Capabilities: Text, code, data, science/math, image, voice, video, code
execution, web search, custom “gems”, Deep Research, Google app
integration, 1M context window.
• Gemini Advanced plan: 27 CAD/month.
 The LLM zoo 6/6
• Anthropic models:
• Claude 3.7 Sonnet: Latest model, non-reasoning. Free (limited use).
• Claude 3.7 Sonnet Thinking: Reasoning variant. Paid only.
• Capabilities: Text, code, data, science/math, image input only. 200K
context window.
• No voice, video, code execution, custom bots. Web search in US only.
• Pro plan: 17 USD/month, max plan: 100-200 USD/month.
 Benchmarks 1/2
• GPQA Diamond: Graduate-Level Google-Proof Q&A.
• 198 PhD-level MCQ in biology, physics, chemistry.
• Humans + web access: 22%.
• Gemini 2.5 Pro: 84%.
• OpenAI o3: 83%.
• OpenAI o4-mini-high: 78%.
• Claude 3.7 Sonnet Thinking: 77%.
• OpenAI GPT-4.5: 71%.
• Gemini 2.5 Flash: 70%.
 Benchmarks 2/2
• Humanity’s Last Exam: 2,500 PhD-level MC or short answer
questions in math, physics, biology, medicine, humanities, social
science, computer science, engineering, chemistry, and more.
• Leading expert human (estimate): 6-8% (mostly MC guesses).
• OpenAI o3: 20%.
• OpenAI o4-mini-high: 18%.
• Gemini 2.5 Pro: 17%.
• Gemini 2.5 Flash: 12%.
• Claude 3.7 Sonnet Thinking: 10%.
 Common misconceptions 1/7
• Misconception: “LLMs only predict the next token, so they can’t do
_____”.
• Rebuttals:
• How else would you write text?
• Humans also predict the next token.
• By predicting the next token in training, neural net encodes syntax,
semantics, world knowledge, etc.
• Analogy: predict next move in chess.
 Common misconceptions 2/7
• Misconception: “LLMs just memorize the entire Internet; no better
than Googling”.
• Rebuttals:
• High scores in Google-proof benchmarks.
• Neural net doesn’t store Internet pages verbatim; uses them to internalize
general meaning.
 Common misconceptions 3/7
• Misconception: “LLMs can’t even multiply two numbers”.
• Rebuttals:
• Was true of ChatGPT in 2023. Now LLMs just use Python to multiply.
• Most humans would use a calculator.
• Training doesn’t focus on this skill; could theoretically be improved, but
there’s no need.
 Common misconceptions 4/7
• Misconception: “I tried an LLM once and it couldn’t do _____”.
• Rebuttal:
• Was true of ChatGPT in 2023. Now LLMs are much more capable.

• Misconception: “I tried an LLM today and it couldn’t do _____”.

• Rebuttal:
• Choose the right model. GPT-4o mini can’t do much, o3 can do a lot.
 Common misconceptions 5/7
• Misconception: “LLMs can’t learn anything new”.
• Rebuttals:
• Latest models learn on the spot: read documents, web search.
• Examples:
• First upload docs/manual/tutorial, then ask question.
• Enable search to get up-to-date info (o3/GPT-4.5 enable by default).
 Common misconceptions 6/7
• Misconception: “Students cannot use LLMs to cheat on this
assessment because it _____”.
• Rebuttals:
• Requires private course material? Upload notes/slides.
• Requires vision? No problem.
• Requires clicking with the mouse? Operator / Claude computer use.
• Requires higher-order reasoning? PhD-level reasoning models.
• Randomizes numbers? No problem.
• Is Google-proof? See earlier slide.
• Is proctored remotely? Interview Coder app, or just use phone.
• Is timed? LLMs are faster than humans.
• Is run through AI detector? Doesn’t work.
• Is proctored in person? Earpiece + smart glasses (soon).
 Common misconceptions 7/7
• Misconception: “Student LLM use hurts learning outcomes”.
• Rebuttals:
• Very true if student just cuts and pastes!
• Numerous studies show LLM tutors improve student performance (in
exams with no LLM available).
• My own experience:
• Students said my AI chatbot helped a lot.
• ASTR 1P02 students who used the chatbot got 6 points higher grades on average.
• (Note: Does not imply causation.)
 Create your own chatbot 1/2
• Easiest way: custom GPTs.
• Gemini gems cannot be shared.
• Create custom instructions: course material, pedagogical
preferences, logistics…
• Limitations:
• Must have paid plan.
• Free users must use GPT-4o mini, an inferior model.
• Even paid users must use GPT-4o, not a reasoning model.
• Privacy concerns.
 Create your own chatbot 2/2
• Harder way: use an API.
• Much more customizable, including choice of model.
• Students can use for free.
• Limitations:
• Must have good programming skills (HTML/CSS/JavaScript/Python).
• Must host own website.
• Must pay whenever it’s used (per token).
• Workaround: Gemini 2.5 Flash (500 req/day free)!
Easy way:
Interface: ChatGPT
Model: GPT-4o/4o-mini
Hard way:
Interface: My own website
Model: Gemini 2.5 Flash (API)
Part III: Future
 To be discussed in
another workshop…

Any questions?