ozm · map.name

AI & Machine Learning — July 2026

From July 18 to 26, you gathered research claims, lectures, courses, repositories, and industry posts into a compact record of AI’s moving parts. With no locations attached, its places are the model, the agent, the knowledge graph, the codebase, and the page. Memory, structure, cost, and clear language keep returning.

From July 18 to 26, you collected a compact argument about AI’s working parts. With no locations attached, the thread moves among lectures, courses, repositories, research claims, essays and industry signals. What holds it together is a recurring turn away from magic: toward training, memory, graphs, lean prompts, cleaner prose and tests.

[1]

Andrej Karpathy just broke the entire premise of modern AI: "Agents aren't magic. They're distillation at scale." 99.99% of your LLM's capacity is wasted on garbage data it never needed. source

The thread repeatedly dismantles the agent as a single mysterious intelligence. Karpathy is quoted calling agents “distillation at scale” [1]; Stanford and Anthropic appear in a claim about weak agent memory and 42% lower performance [2]. DeepMind’s 180 agent-team setups shift attention to configuration [3], while prompt caching makes the same practical turn toward recurring costs [4].

[2]

whoever shared the "i have adhd" skill with me thank you it's made my claude replies so good pic.twitter.com/f8ymvXoe2d — jacky source

[3]

Andrej Karpathy dropped a 22-minute lecture that makes every “prompt engineering expert” look outdated. He exposed what most people still don’t understand: ChatGPT is not trained in one step The base model only predicts … source

[4]

Andrej Karpathy just broke the entire premise of modern AI: "Agents aren't magic. They're distillation at scale." 99.99% of your LLM's capacity is wasted on garbage data it never needed. source

[5]

STANFORD AND ANTHROPIC SPENT $3.1M TO PROVE YOUR AGENT PERFORMS 42% WORSE THAN IT SHOULD - AND FOUND THE FIX most agents have the memory of a goldfish - 30 seconds and everything is forgotten - a graph is elephant memory … source

[6]

[7]

new AI research dropped and it should worry everyone currently using AI agents... Google Deepmind researchers built 180 different agent team setups, gave every single one the same budget, and let them compete on the same … source

[8]

Ontologies for AI, by AI As more vendors and people start moving towards the realm of data modeling, knowledge engineering and semantics, inevitably the water gets muddied. source

[9]

8 ways to cut your agent bill: (and when each one pays off) 1) prompt caching → the system prompt and the tool schemas never change between calls, so stop paying full price for them. source

[10]

Google just released a 1-hour course on building agentic knowledge Graphs from scratch: • 00:00 - Introduction to AI agents • 08:09 - Foundations of multi-agent systems • 13:25 - Introduction to Graph engineering • 23:27 … source

[11]

Have you built a language model? You should. It's so much fun to chat with something you made. Anyone can do it, too. I made an app that teaches the fundamentals and gives you everything you need to build your own: pic. source

[12]

GitHub - petergyang/no-ai-slop: Removes 20+ patterns of AI slop from any piece of writing. Removes 20+ patterns of AI slop from any piece of writing. - petergyang/no-ai-slop source

Ontologies bring data modeling, knowledge engineering, and semantics into one contested vocabulary [5]. Google’s course connects agents, multi-agent systems, and knowledge graphs [6], while a Karpathy lecture stresses that ChatGPT is not trained in one step [7]. The report that roughly 80% of a Claude Code system prompt was removed adds a quieter lesson: architecture also means deciding what can be left out [8].

[13]

Ontologies for AI, by AI As more vendors and people start moving towards the realm of data modeling, knowledge engineering and semantics, inevitably the water gets muddied. source

[14]

[15]

[16]

AI is moving into the physical world. We're excited about a new wave of startups rebuilding the systems that power the real world, from education and healthcare to defense, finance, infrastructure, and work itself. pic. source

[17]

[18]

8 ways to cut your agent bill: (and when each one pays off) 1) prompt caching → the system prompt and the tool schemas never change between calls, so stop paying full price for them. source

[19]

& ;He& ;s an analog man in a digital AI world& ;: Nolan’s ‘The Odyssey’ made back its $250M budget in 3 days without industrywide AI cost cutting | Fortune The three-hour epic and Nolan’s $6 billion personal box office … source

[20]

We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them. — Thariq source

[21]

One way you can recognize AI slop is when the ideas and the diction don't match — when the ideas are completely ordinary, but the diction is that of someone announcing a brilliant discovery that they're really excited … source

[22]

The Bitter Lesson source

[23]

whoever shared the "i have adhd" skill with me thank you it's made my claude replies so good pic.twitter.com/f8ymvXoe2d — jacky source

[24]

[25]

For my first post, I’m sharing a letter signed on why open models matter. AI will transform every industry, power every company, and be built by every country. source

[26]

GitHub - petergyang/no-ai-slop: Removes 20+ patterns of AI slop from any piece of writing. Removes 20+ patterns of AI slop from any piece of writing. - petergyang/no-ai-slop source

[27]

We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them. — Thariq source

[28]

[29]

AI agents can write code many times faster than a human. What this means is that you, the programmer, have a large amount of time to use those agents to write unit tests, to write acceptance, tests, to write property … source

[30]

🚨 Apple just solved the biggest problem in AI. pic.twitter.com/8H0jhxw8Ab — Peter Dedene 🚨 Apple just solved the biggest problem in AI. pic.twitter.com/8H0jhxw8Ab — Peter Dedene source

[31]

[32]

What seems to be forming is a quieter standard for AI: less enchantment around agents, more attention to training, memory, structure, cost, language and verification as systems move outward.

[33]

The Bitter Lesson source

[34]

For my first post, I’m sharing a letter signed on why open models matter. AI will transform every industry, power every company, and be built by every country. source

[35]

[36]

🚨 Apple just solved the biggest problem in AI. pic.twitter.com/8H0jhxw8Ab — Peter Dedene 🚨 Apple just solved the biggest problem in AI. pic.twitter.com/8H0jhxw8Ab — Peter Dedene source

What seems to be forming is a less enchanted view of AI: not one marvel, but an arrangement of training, memory, structure, cost, tests, and language choices.