Karpathy Maps AI Job Exposure Across 342 Occupations as Agent Memory Systems Get a Major Rethink

March 15, 2026 · 25 sources

The AI community is deep in the weeds on agent infrastructure, with ByteDance's OpenViking proposing a file-system metaphor for agent memory and multiple projects pushing local inference forward. Andrej Karpathy's AI job exposure scoring project sparked widespread conversation, while Claude Code's ecosystem continues to expand with local GPU support and marketing applications.

Daily Wrap-Up

The conversation today orbited around a central tension: AI agents are getting dramatically more capable, but the infrastructure to make them reliable is still being figured out in real time. ByteDance open-sourced OpenViking to treat agent context like a file system instead of a flat vector store. Cognee shipped self-healing skills that watch their own failures. A new terminal called Slate is trying to wrangle multi-model agent swarms into something usable. Everyone agrees agents are the future, but nobody agrees on the plumbing yet.

Meanwhile, Andrej Karpathy quietly dropped one of the more sobering projects of the week: an LLM-scored breakdown of all 342 Bureau of Labor occupations ranked by AI exposure. Software developers landed at 8-9 out of 10. The average across all jobs was 5.3. That number landed differently depending on who was reading it, but the fact that the entire pipeline is open source means anyone can audit or challenge the methodology. It's the kind of project that matters more for the conversation it starts than the scores it assigns. On the lighter side, someone apparently sold their entire house in five days using ChatGPT, which Greg Brockman quote-tweeted with the energy of a proud parent.

The most practical takeaway for developers: if you're building with AI agents, study OpenViking's tiered loading architecture (L0/L1/L2 context levels) and consider whether your current approach of dumping everything into context is costing you both tokens and accuracy. The file-system metaphor for agent memory is emerging as a pattern worth understanding now.

Quick Hits

@Crypto0637 shared Superheat's $2,000 electric water heater with built-in ASIC Bitcoin miners. Same energy usage, but it mines crypto while heating your water. The future of appliances is weird.
@5le recommends a portable charger that can juice up a MacBook on the go.
@OwlcatGames announced an action RPG based on The Expanse, now available to wishlist on Steam.
@nateliason and Cameron Sorsby are launching a high school for entrepreneurs with a wild promise: make $1M by graduation or get full tuition refunded.
@doodlestein shared an apparently excellent prompt with no further context. The image did the talking.
@TheAhmadOsman flexed what can only be described as the final boss of LocalLLaMA home server setups.
@unusual_whales published a guide for connecting OpenClaw agents to real-time stock and options data.
@RayFernando1337 told the story of being pulled into Apple's secret Apple Watch update system project, back when the prototype had no recovery port.
@gdb quote-tweeted the ChatGPT house-selling story with a note about building confidence through AI usage.

Agent Memory and Infrastructure

The biggest technical theme today was rethinking how AI agents manage context and memory. The flat RAG paradigm, where you embed everything and do a similarity search, is showing its cracks at scale. @ihtesham2005 broke down ByteDance's OpenViking framework, which organizes agent context under a unified viking:// protocol: "Memories, resources, skills all organized in directories with unique URIs. Agents can ls, find, and navigate context like a developer working a terminal." The tiered loading system is the real innovation here, with L0 providing one-sentence abstracts, L1 offering ~2k token overviews for planning, and L2 loading full details only when needed. Most agents today stuff everything into context and hope for the best. This approach loads only what's relevant, when it's relevant.

On a parallel track, @iruletheworldmo highlighted Cognee's work on self-improving agent skills that "observe their own failures, inspect what went wrong, and amend themselves automatically." And @steipete pointed to community plugins solving memory problems in OpenClaw, specifically the qmd memory plugin for agents that get forgetful after context compaction. @browser_use retweeted research on when to apply compaction to agent conversations, claiming an 80% cost reduction. These aren't competing ideas so much as different layers of the same problem: agents need better memory, better context management, and ideally the ability to learn from their own mistakes. The community is converging on solutions from multiple angles.

Claude Code Ecosystem Expands

Claude Code dominated the conversation from multiple angles today. @techNmak covered Unsloth AI's guide to running Claude Code entirely on local GPUs, and the guide's value goes beyond the setup instructions. It explains why local inference feels slow (an attribution header breaks KV caching), why Qwen3.5 outputs degrade (f16 KV cache is the default but q8_0 or bf16 performs better), and how to disable thinking mode for agentic tasks. "Fits on 24GB. RTX 4090, Mac unified memory." That's a meaningful threshold for developers who want API independence.

@shannholmberg argued that Claude's superpowers plugin with 83,000 GitHub stars is "the most underrated plugin for marketers right now," noting that almost everyone using it is a developer. @mattpocockuk raised pointed questions about whether OAuth tokens from Claude subscriptions can legally power the Claude Agent SDK for local dev loops, and whether open-source tools built on that pattern can be distributed. The legal compliance docs and public statements from Anthropic appear to contradict each other, which is the kind of ambiguity that chills open-source development. @shanraisshan's best practices repo for Claude Code hit 11.8K stars, and @shareAI-lab's "build a Claude Code clone from scratch" repo reached 9K. The ecosystem is maturing fast, but governance questions are trailing behind the code.

AI's Impact on Jobs and Careers

Karpathy's job exposure project was the most discussed single item today. @JoshKale summarized the methodology: "Scraped all 342 occupations from the Bureau of Labor. Fed each one to an LLM with a detailed scoring rubric. Built an interactive treemap where rectangle size = number of jobs and color = how exposed that job is to AI." The key heuristic is straightforward: if the work product is digital and the job can be done from home, exposure is high. Software developers scored 8-9. Medical transcriptionists scored a perfect 10. Roofers and janitors scored 0-1. The average was 5.3 out of 10 across all occupations.

@TukiFromKL amplified the anxiety with a rapid-fire roundup that included a CEO of a $200 billion company saying "35% of new grads won't find jobs" on camera without flinching, and Meta firing 15,000 people despite $165 billion in revenue. The juxtaposition of AI capability stories (curing a dog's cancer, selling a house in five days) against job displacement data creates a disorienting effect. These aren't abstract trends anymore. They're landing in specific occupations with specific exposure scores, and the entire scoring pipeline is open source for anyone to verify.

Agentic Workflows and Tooling

Developers are actively experimenting with how to orchestrate AI agents in practice. @jamonholmgren teased what he calls the "Night Shift workflow," claiming it's "about 5x faster, better quality, I understand the system better, and I'm having fun again." He contrasted it with previous workflows that "left me exhausted, overwhelmed, and feeling out of touch with the systems I was building." Details are still forthcoming, but the sentiment resonates: raw speed from agents isn't enough if the developer loses comprehension of their own codebase.

@realmcore_ showed off Slate, a terminal UX designed for agent swarms that lets you "literally use Opus 4.6 and GPT 5.4 at the exact same time." Making multi-model orchestration intuitive is a hard UX problem, and most existing terminals aren't built for it. @victormustar demonstrated a different approach, using the Hugging Face CLI as the sole interface for an autonomous AI engineering agent. The agent autonomously fine-tuned a model for Japanese, running 23 SQL queries to audit data quality across 153K examples before curating 4,096 training samples. "Feels less contaminating: an agent browsing the web for solutions would defeat the whole purpose of autoresearch."

@sharbel compiled this month's fastest-growing GitHub projects, and the list reads like a snapshot of where developer energy is flowing. OpenClaw leads at 122K stars as a personal AI assistant. Superpowers hit 30.7K stars as a plug-and-play agent skills framework. But the more surprising entries tell a bigger story: RuView (30.4K stars) turns regular WiFi signals into real-time human pose detection with no cameras or sensors. MiroFish (17K stars) bills itself as a "swarm intelligence engine that predicts anything," and @k1rallik noted that its creator, a student, "outranked OpenAI on GitHub with 10 days of vibe coding." And then there's Heretic at 7.6K stars, which "removes guardrails from any language model automatically." The open-source AI ecosystem is moving fast enough that next month's list will look completely different.

Knowledge Management with AI

@Atenov_D published a detailed breakdown of a "10x learning speed" setup that chains NotebookLM, Google Drive, Obsidian with Smart Connections, and Claude Code into a continuous knowledge processing pipeline. The core insight is that breaking information into atomic notes (one idea per file, tagged and cross-linked) isn't just organization but is the learning itself: "When you open a note weeks later and see three related ideas you'd forgotten, that's active recall. That's what builds memory." Omar Khattab (@lateinteraction) retweeted a related project where someone used DSPy and RLM to automatically refactor a messy Obsidian vault. The pattern emerging here is that AI isn't just helping people consume information faster; it's restructuring how knowledge gets stored and retrieved in ways that compound over time.

Sources

Ray Fernando @RayFernando1337 · Mar 9

I was part of a secret project at Apple that revolutionized the way we update our devices, and it all started with a prototype that had no recovery port. I was sitting in my office, when suddenly, I got pulled into this top-secret project. I mean, it was so hush-hush that I had to sign all these NDAs, and basically kiss my holidays and free time goodbye until we finished it. So, I show up to the meeting, and the manager confirms I signed the NDAs. Then, he pulls off this black cloth, and boom! There's the prototype, right in front of me. My eyes were so big, and my heart was pounding like crazy. I had a million questions running through my mind. Turns out, I was going to be the one making sure the Apple Watch could update without any problems. But get this - when I asked about plugging it into a computer to recover it if something went wrong, the manager dropped a bombshell on me. There was no port! I was like, "HOLY CRAP! What did I just sign up for??" And then, he just walked away and told me to figure it out. Can you believe that? I ended up spending the next several years working with tons of teams across Apple, trying to come up with a solution. It was intense, but we managed to create this incredible update system that would totally change the game not only for the Apple Watch but for all future systems as well. Looking back, it was one of the craziest and most challenging things I've ever done. It taught me a lot about what I'm capable of and how important it is to have a great team by your side. It also taught me how important developing relationships with people are. Always have people’s back no matter what and be a great human being. It returns dividends for life. Stay Hungry. Stay Foolish.

Ahmad @TheAhmadOsman · Dec 19

LocalLLaMA Home Server Final Boss 😎

Eli Schwartz @5le · Jan 19

This is my new favorite gadget to travel with. It can charge a MacBook on the fly! https://t.co/X0S33fuxTv

Owlcat Games @OwlcatGames · Feb 7

Take the helm and step into your Expanse story. WISHLIST Action RPG The Expanse: Osiris Reborn on Steam 🚀

unusual_whales @unusual_whales · Feb 12

Building an agent with OpenClaw? Point it at Unusual Whales Skill MD to level up your agent with real-time stock and option data from Unusual Whales. See the step by step guide here:

Nat Eliason @nateliason · Mar 13

Make $1m by graduation. Or get 100% of your tuition refunded. That's the promise of the new high school for entrepreneurs Cameron and I are launching this fall through @AlphaSchoolATX. We need 2-3 coaches to help make it happen. DM us or apply!

C CameronSorsby @CameronSorsby

We’re launching a new @alphaschoolatx high school for aspiring entrepreneurs. Our promise: Make $1m by graduation, or receive a full tuition refund. Yes, this will be the coolest high school in the world. And we're building the best team in the world to make it happen. We’re looking for 2-3 exceptional coaches to help us guide the students towards achieving this aggressive but achievable goal. You won’t be giving lectures or assigning homework. You’ll be grilling them on their P&L, driving them to the car wash they bought, critiquing their email funnels, pushing them to do things 99% of the world doesn't believe is possible. Job posting is live and DMs are open.

Crypto Currency @Crypto0637 · Mar 14

🚨 BREAKING A WATER HEATER THAT PAYS YOU IN BITCOIN JUST DROPPED. Superheat unveiled a $2,000 electric water heater that secretly mines Bitcoin while heating your water. Same energy usage as a normal heater — but the built-in ASIC miners earn Bitcoin in the background, helping offset your energy bill. HEAT YOUR WATER. MINE BITCOIN. LOWER YOUR BILLS. THE FUTURE OF HOME APPLIANCES? ⚡️🚀

Sharbel @sharbel · Mar 14

the fastest growing GitHub projects this month: 1. openclaw/openclaw (122K stars) your own personal AI assistant, runs 24/7 on any OS (what I use to run all my agents) 2. obra/superpowers (30.7K stars) agentic skills framework. plug-and-play tools for AI agents 3. ruvnet/RuView (30.4K stars) turns regular WiFi signals into real-time human pose detection no cameras. no sensors. just WiFi. 4. 666ghj/MiroFish (17K stars) swarm intelligence engine that predicts anything 5. moeru-ai/airi (16K stars) self-hosted AI companion with real-time voice chat runs on your own machine. you own it. 6. shanraisshan/claude-code-best-practice (11.8K stars) the best practices repo for building with Claude Code 7. badlogic/pi-mono (11.8K stars) full AI agent toolkit: CLI, unified LLM API, web UI, Slack bot 8. bytedance/deer-flow (10.4K stars) ByteDance's open-source SuperAgent. researches, codes, creates on its own 9. shareAI-lab/learn-claude-code (9K stars) build a Claude Code clone from scratch. bash is all you need. 10. p-e-w/heretic (7.6K stars) removes guardrails from any language model automatically the pace of AI right now is insane. bookmark this. next month's list will look completely different.

Ihtesham Ali @ihtesham2005 · Mar 14

RIP flat RAG ☠️ ByteDance just open-sourced OpenViking and it exposes everything wrong with how we've been building AI agent memory. Here's what every agent framework gets wrong: Memories live in one place. Resources in another. Skills scattered everywhere. And when you need context, you're doing flat vector search and hoping for the best. That's the problem. OpenViking fixes all of it with one idea: treat agent context like a file system. Everything lives under a unified viking:// protocol. Memories, resources, skills all organized in directories with unique URIs. Agents can ls, find, and navigate context like a developer working a terminal. But the real breakthrough is tiered loading: → L0: one-sentence abstract for quick lookup → L1: ~2k token overview for planning decisions → L2: full details loaded only when actually needed Most agents dump everything into context and pray. OpenViking loads only what's needed, when it's needed. Token costs drop. Accuracy goes up. And retrieval actually makes sense now. Instead of one flat semantic search, it does directory-level positioning first, then recursive refinement inside high-score directories. You can literally watch the retrieval trajectory no more black box. The self-evolution piece is wild too. At the end of every session, it automatically extracts learnings and updates agent and user memory. The agent just gets smarter the more you use it. 9K stars. 13 contributors. Built by the ByteDance Viking team that's been running vector infrastructure since 2019. 100% Opensource. Apache 2.0. Link in comments.

Atenov int. @Atenov_D · Mar 14

Someone sent me a post titled "enthusiast built a setup for 10x learning speed" I wanted to laugh. Then I watched the video and couldn't stop thinking about it. The guy understood something real. He just put it behind a paid course. So I spent a weekend reverse-engineering it from everything he showed publicly. Here's the full breakdown - no course required. > The system has four parts. NotebookLM eats raw material - videos, books, articles. It extracts structured notes. You direct what to pull out, it does the reading. A Google Drive script moves files between NotebookLM and the rest of the system. Set it up once, forget it exists. Obsidian + Smart Connections is where knowledge lives. Smart Connections builds a vector index locally - on your machine and shows a live panel of notes relevant to whatever you're reading right now. Claude Code closes the loop. It pulls notes from Drive, processes them into your vault, and runs the transformation that makes everything useful. Every note gets rewritten using MOC + Zettelkasten. One idea per note. Tagged, titled, linked to related ideas. When you open any note, you see exactly which other notes connect to it. You're walking a graph of your own thinking. Most people store information to retrieve it later. That's a better Google Drive. Breaking things into atomic notes forces you to understand what you're storing. The tags and cross-links aren't overhead - they're the learning itself. When you open a note weeks later and see three related ideas you'd forgotten - thats active recall. That's what builds memory. > Three things that make or break this: The atomic note rule is non-negotiable. One idea per file. Compromise here and the graph collapses into a better-looking mess. Give Claude Code a reference template. A concrete example of what a finished note looks like. Without it, output varies. / With it, every note is consistent - and consistency is what makes the vault navigable a year from now. Write wherever you want. Terminal, phone, voice memo. Then ask Claude Code to process it into Obsidian format. The only requirement is Markdown at the end. Six months in, opening your vault feels less like searching a database and more like consulting a version of yourself that remembered everything. That's the actual 10x. Bookmark this. A few hours to set up. Compounds for years.

A Atenov_D @Atenov_D

How I turned Obsidian into a second brain that runs itself

Peter Steinberger 🦞 @steipete · Mar 14

There's a lot of cool stuff being built around openclaw. If the stock memory feature isn't great for you, check out the qmd memory plugin! If you are annoyed that your crustacean is forgetful after compaction, give https://t.co/C5B7PJxorq a try!

Jamon @jamonholmgren · Mar 14

My current agentic workflow is about 5x faster, better quality, I understand the system better, and I’m having fun again. My previous workflows have left me exhausted, overwhelmed, and feeling out of touch with the systems I was building. They also degraded quality too much. This is way better. I’m not ready to describe in detail. It’s still evolving a bit. But I’ll give you a high level here. I call this the Night Shift workflow.

Jeffrey Emanuel @doodlestein · Mar 14

God, I love this prompt. https://t.co/ha8EhtTlhH

Shann³ @shannholmberg · Mar 14

claude superpowers is the most underrated plugin for marketers right now 83,000 github stars. trending daily. but almost everyone using it is a developer here´s how it works and how to apply it to marketing 🧵 https://t.co/6rYDD7wunT

BuBBliK @k1rallik · Mar 14

me after reading how a student outranked OpenAI on GitHub with 10 days of vibe coding https://t.co/GwQVrxMcBW

K k1rallik @k1rallik

MiroFish: The God View Engine

Omar Khattab @lateinteraction · Mar 14

RT @ramirosalas: My Obsidian vault became a mess over time, so I wrote a program that uses @DSPyOSS and RLM to completely refactor it into…

Tuki @TukiFromKL · Mar 14

🚨 Do you understand what happened in the last 12 hours? > A CEO of a $200 billion company said on camera that 35% of new grads won't find jobs. He didn't even flinch saying it. > Meta made $165 billion last year and is still firing 15,000 people because apparently record profit isn't profitable enough. > Some random guy in Florida sold his entire house in 5 days using ChatGPT. No real estate agent, no commission, no experience. Just vibes and a $20 subscription. > A man in Australia cured his dying dog's cancer with AI after every single vet told him there was nothing left to do. Built a custom vaccine from his couch. > The guy who created Uber and left 300,000 taxi drivers broke is back. Building robots now because apparently ruining one industry wasn't enough. > Tinder wants access to your camera roll. Your drunk photos, your 3am notes app meltdowns, your deleted selfies. They're calling it a "vibe check." > Naval, the man who made hundreds of millions investing in software, just said software is dead. Four words and the entire industry felt it. > And Anthropic removed the limit on how long their AI can think and then doubled everyone's usage for free. Because when the product is addictive enough you give the first taste away. All of that happened today. Not this week, not this quarter. Today. A random Saturday in March. This is worse than you being on meth.

🍓

🍓🍓🍓 @iruletheworldmo · Mar 14

bookmark this immediately. cognee just solved the biggest problem with ai skills/prompts, they break silently over time and its hard to notice their fix: skills that observe their own failures, inspect what went wrong, and amend themselves automatically. try not to fall behind ^^

T tricalt @tricalt

Self improving skills for agents

Josh Kale @JoshKale · Mar 15

Andrej Karpathy just dropped a project scoring every job in America on how likely an AI will replace it from 0-10 > Scraped all 342 occupations from the Bureau of Labor > Fed each one to an LLM with a detailed scoring rubric > Built an interactive treemap where rectangle size = number of jobs and color = how exposed that job is to AI The key signal in his scoring: if the work product is fundamentally digital and the job can be done entirely from a home office, exposure is inherently high. The scale: 0-1: Roofers, janitors 4-5: Nurses, retail, physicians 8-9: Software devs, paralegals, data analysts 10: Medical transcriptionists Average across all 342 occupations: 5.3/10. The entire pipeline is open source. BLS scraping, LLM scoring, the visualization. All of it. Much respect for the sensei this is scary and awesome

akira @realmcore_ · Mar 15

I don't think there's a single terminal ux that handles agent swarms well With slate, you can literally use Opus 4.6 and GPT 5.4 at the exact same time But making it intuitive took a ton of work So heres a thread on how it works and how to actually use it 🧵 https://t.co/7WX8qO0kmn

Browser Use @browser_use · Mar 15

RT @mamagnus00: When to do compaction? I could reduce the cost of my agent by 80% with this. https://t.co/NyciPtutrJ

Tech with Mak @techNmak · Mar 15

Claude Code can run entirely on your local GPU now. Unsloth AI published the complete guide. The setup itself is straightforward - llama.cpp serves Qwen3.5 or GLM-4.7-Flash, one environment variable redirects Claude Code to localhost. But the guide is valuable because of what it explains beyond the setup: Why local inference feels impossibly slow: Claude Code adds an attribution header that breaks KV caching. Every request recomputes the full context. The fix requires editing settings.json - export doesn't work. Why Qwen3.5 outputs seem off: f16 KV cache degrades accuracy, and it's llama.cpp's default. Multiple reports confirm this. Use q8_0 or bf16 instead. Why responses take forever: Thinking mode is great for reasoning but slow for agentic tasks. The guide shows how to disable it. The proof it all works: Claude Code autonomously fine-tuning a model with Unsloth. Start to finish. No API dependency. Fits on 24GB. RTX 4090, Mac unified memory.

Greg Brockman @gdb · Mar 15

“The more you use AI, the more confident you’ll become in leveraging those tools”

D Dexerto @Dexerto

Florida man sold his house in just 5 days after letting ChatGPT handle the entire process instead of a real estate agent The AI handled pricing, marketing, showings, and even helped draft the contract https://t.co/t5BfIGN9lZ

Matt Pocock @mattpocockuk · Mar 15

Can I get some questions answered by someone at Anthropic? 1. Can you use an OAuth token generated from a subscription to power the Claude Agent SDK strictly for using Claude Code in a local dev loop? All I want is a more reliable API for parallelizing multiple Claude Code's. 2. If I build an open source tool that relies on this pattern - i.e. for making parallelization easier - can I distribute it so that other people can use it? The reason I'm asking is that the legal compliance docs and @trq212's public statements (below) appear to contradict. https://t.co/avHG1GLLM1

Victor M @victormustar · Mar 15

The hugging face CLI is the cheat code for autonomous AI engineering. Ask your agent to finetune a model for japanese, it will autonomously: 1. hf models ls: 11 searches to find Qwen3.5-2B 2. hf datasets info: 8 lookups across 3 Japanese instruction sets 3. hf datasets sql: 23 queries to audit quality (lengths, dedup, translation flags) 4. hf download: pulled model + 3 datasets, curated 4,096 examples from 153K 5. hf upload: adapters + model card pushed to Hub 6. hf buckets sync: artifacts persisted to storage Also it's quite cool because I give my autoresearcher access to the CLI without any other internet browsing capabilities. Feels less contaminating: an agent browsing the web for solutions would defeat the whole purpose of autoresearch (here it's raw data from HF).

V victormustar @victormustar

Autoresearch: make Qwen 0.5B better at chess 👀 tip: you can /loop claude on the Pi session so it does a nice reporting of what's happening every 15 minutes https://t.co/tqYAZeQcXS