The Model Is Not the Whole System

i gather tweets that are interesting that signal the shape of latest llm capabilities and news around them

Prediction became assistance

Andrej Karpathy’s lecture traces the crucial shift from predicting tokens to behaving like an assistant [1]. His separate observations about agent efficiency and model training place that shift inside a larger question: how much useful work can be extracted from the underlying computation [2]? An argument that machine-learning progress comes primarily from increased compute rather than better algorithms pushes the thesis further [3]. Scale supplies capability, but it does not by itself produce a dependable agent.

[1]

a lecture by Andrej Karpathy discussing the training and functionality of ChatGPT, particularly its evolution from predicting tokens to becoming an assistant source

[2]

a tweet discussing Andrej Karpathy's insights on the efficiency and potential of AI agents and model training source

[3]

an article discussing the concept that the best way to make progress in machine learning is through increased computation rather than algorithmic improvement source

The surrounding system starts to matter

Research from Stanford and Anthropic moves the focus to memory performance [4], with Graph Engineering presented as a way to improve retention and agent results [5]. That makes the warning about casually using words such as “ontology” and “context” more than semantic fuss: vague data models become vague agent behavior [6]. Even the removal of a significant portion of the Claude Code system prompt became an experiment worth studying [7]. Around the same core model, prompt caching and model routing can reduce bills [8], while delegating testing can improve programming efficiency [9]. The gains increasingly come from engineering the environment, not merely choosing the model.

[4]

a discussion about research conducted by Stanford and Anthropic on improving AI agent memory performance source

[5]

a tweet discussing research on agent performance improvements and memory retention through Graph Engineering from Stanford and Anthropic source

[6]

a tweet discussing the complexities of data modeling, knowledge engineering, and the misuse of terms like 'ontology' and 'context' in the AI industry source

[7]

a discussion about system prompts in AI, specifically the removal of a significant portion of the Claude Code system prompt and the insights gained from this change source

[8]

a tweet outlining strategies to reduce agent billing costs in machine learning applications, including prompt caching and model routing source

[9]

a tweet discussing how AI agents can improve programming efficiency by handling testing tasks source

Capability becomes a question of control

Competitive teams of agents at Google DeepMind extend the system beyond one assistant [10], while startups are applying AI across education, healthcare and finance [11]. Jensen Huang’s case for open models connects that expansion to safety, innovation and national sovereignty [12]. Yet the clearest quality controls remain strikingly ordinary: one tool removes more than 20 recurring patterns from AI-generated prose [13], and another post identifies bad AI writing by the mismatch between ordinary ideas and inflated diction [14]. One person even credits a skill for managing ADHD with improving Claude’s responses [15]. Better output still depends on someone noticing what the system needs.

[10]

a summary of new AI research by Google DeepMind on competitive AI agent teams and their performance source

[11]

a tweet from Y Combinator discussing the emergence of startups transforming industries like education, healthcare, and finance through AI source

[12]

a tweet from Jensen Huang discussing the importance of open AI models for safety, innovation, and national sovereignty source

[13]

a GitHub repository for a tool that removes over 20 patterns of AI-generated content deemed unnecessary or poor-quality in written text source

[14]

a tweet discussing the recognition of poorly written AI-generated content, highlighting the mismatch between ordinary ideas and exaggerated diction source

[15]

a tweet expressing gratitude for a skill related to managing ADHD and its positive impact on responses generated by Claude source

Also in this thread

[16]

a tweet announcing the release of a 1-hour course on building agentic knowledge graphs, covering AI agents, multi-agent systems, and environment setup source

[17]

a promotional tweet about an app that teaches users how to build their own language model and engages in chat conversations source

[18]

a news article discussing Christopher Nolan's film 'The Odyssey', its financial success, and its commentary on the impact of AI in filmmaking source

[19]

a tweet discussing the emergence of startups focused on transforming various sectors such as education, healthcare, and finance through AI source

[20]

a tweet discussing Apple's recent breakthrough in artificial intelligence solutions with a related image source

The durable advantage may not be the agent itself. It may be the care taken around it: what it remembers, how it is instructed, what it costs, how it is tested and whether a person can still recognize a bad sentence.