Large context windows let models read more information, but they become expensive, slow, and less reliable as irrelevant material accumulates.
2
RAG is easy to deploy, yet fixed embeddings struggle with sensitive data, domain-specific distinctions, and questions that require reasoning across documents.
3
Training knowledge into model weights could make inference cheaper and let models learn specialized information, although the right update method and handling of changing data remain open problems.
Summary
Jack Morris argues that current LLM systems handle common knowledge well but fail on private, recent, and niche information. He compares three ways to add knowledge: putting everything in the context window, retrieving relevant material with RAG, and training information into model weights. Full context has steep cost and latency, while long contexts can reduce reasoning quality. RAG is simple and useful, but embeddings can expose source text, mix up domain-specific documents, and fail when answers require relationships across multiple chunks. Morris proposes using synthetic data to turn small private datasets into broader training sets, then updating only a small part of a model to reduce forgetting. He discusses LoRA, prefix tuning, memory layers, and related methods, but is clear that the evidence conflicts. He expects future systems to combine weight updates with RAG and context. The unanswered questions include when training becomes cheaper than retrieval, how to handle contradictory or time-varying information, and how to personalize models at large scale.
Long-tail and private knowledge are outside what a base model can answer
Morris starts with examples where ChatGPT is useful for general questions but fails on recent, private, or niche tasks. Without web search, it does not know that the Blue Jays won the World Series after its knowledge cutoff. It is also poor at optimizing an AMD GPU kernel, answering questions about a company's BlackRock partnership, implementing a feature in a private monorepo, or writing in a user's style. He says prompting cannot create missing knowledge: asking the model to learn or practice does not give it the data or process needed to do so.
Knowledge can be added through context, RAG, or weight updates
Morris frames knowledge injection as three approaches. Full context copies all relevant material into the prompt. RAG retrieves a smaller set from a larger collection and adds it to the prompt. The third approach is training information into the model's weights, which is the focus of the talk. Full context can work when the dataset is small, such as a medical record or a finite collection of World Series results, but it becomes difficult when the information is large or repeatedly used.
Long context has both a systems cost and a reasoning limit
Morris gives an example of context length changing generation speed from 10,000 tokens per second with 1,000 context tokens to 130 tokens per second with a 128k context. He explains that transformer self-attention compares every token with every other token, producing quadratic growth. Larger windows can avoid a model breaking technically without preserving reasoning quality. He cites Chroma's Context Rot report, where performance worsens as irrelevant information is added, and says models may produce grammatical answers while failing to solve the actual problem.
More efficient attention does not automatically make infinite context useful
Morris mentions Mamba, state space models, linear attention, hybrid attention, sparse attention, and sliding-window attention as ways to reduce the cost of long sequences. Their efficiency does not remove the reasoning problem. In his view, a model that can process an unlimited context but cannot reason across all of it has not solved the practical problem. He uses MiniMax M2's choice to retain regular quadratic attention as an example of the trade-off between computational efficiency and model quality.
RAG is easy to use, but embeddings expose data and flatten domain distinctions
Morris describes vector databases as today's practical memory system for LLM applications. Embeddings are convenient, and he says a basic pipeline can take about five lines of code. He also presents security research showing that an attacker can reconstruct much of the original text from embeddings, with the cited work recovering 90% of text exactly at certain lengths. In a Visa and MasterCard example, conventional embeddings place related documents so close together that a Visa query can retrieve MasterCard material. Contextual embeddings can adapt to surrounding documents and improve this case, especially for niche domains.
Retrieval cannot answer every question that depends on relationships across documents
Morris says some questions require associations among multiple documents or information that is implied rather than explicitly stated. Chunking and retrieval may return pieces of the answer without allowing the model to reason across every needed relationship. He considers agentic search and deep research promising because they spend more inference-time compute, make repeated searches, and think over the results. This creates a trade-off: deep research pays at inference time, while weight-based approaches pay more during training and can make later inference cheaper.
Synthetic data can turn a small private dataset into trainable knowledge
Morris argues that directly training on a small dataset with next-token prediction can make a model memorize the text without learning to answer related questions. In his 3M financial-report example, the model reached zero loss on the report but produced a terrible poem about 3M's fiscal year. He points to work by Andrej Karpathy and a Stanford paper on synthetic continued pre-training, where a small source dataset is expanded into diverse examples that describe its facts. He says this can let a model outperform GPT-4 on the target dataset at roughly 100 million to nearly a billion synthetic tokens.
Small parameter updates reduce forgetting, but the best method is unsettled
Updating every parameter risks catastrophic forgetting and can make the model obsessed with the new material. Morris compares full fine-tuning with prefix tuning, LoRA, added mixture-of-experts components, and memory layers. The desired update is small, resistant to forgetting, expressive enough to store useful facts, and fast at inference. He cites results where LoRA learns less and forgets less than full fine-tuning under SFT, while another result from Thinking Machines suggests LoRA can approach full fine-tuning under RL. He also presents memory-layer results that show little forgetting and strong learning, while admitting the evidence is conflicting.
Weight memory will probably coexist with retrieval and must handle changing information
Morris does not expect systems to stop using context or RAG. He imagines periodic training updates followed by retrieval at request time. The right balance depends on data volume, update frequency, freshness, and cost, but he says there is not yet enough analysis to set a decision boundary. He also does not have a settled method for conflicting policies, deleting outdated facts, or temporal data. For personalization, he thinks storing a small LoRA-style update per user could be feasible, while continual updates would be harder.
"If you want to build a system that can answer really detailed documentation questions for your company, you don't need it to know what the capital of the smallest province in Tajikistan is."24:05
Who should watch
You are building a RAG or internal question-answering system and want to understand where embeddings and retrieval may fail.
You need to add private, specialized, or frequently reused knowledge to a model and are comparing prompt-time access with fine-tuning.
You are researching personalized or continually updated models and want a candid account of the open problems around forgetting, cost, and conflicting information.
