20 sources

Vibe Coding

Vibe coding -- building software by prompting AI rather than writing code directly -- is enabling one-person teams to out-execute large labs on UI quality and ship complete applications in single prompts. Opus 4.6 crossed a quality threshold for one-shot website and marketing UI generation, while the community has developed techniques to overcome AI's design limitations: constraining output with layout templates from block libraries (Tailark, Tailwind UI, shadcn), using Dribbble screenshots as design references, and referencing enterprise-grade open-source dashboards. Non-technical builders are achieving real traction (5.8K GitHub stars, viral projects) by treating code as a creative medium. The emerging vibe-coding stack converges on managed backend + managed deploy + AI editor (Convex + Vercel + Cursor). However, a critical scaling blind spot persists: LLMs suggest infrastructure choices based on training data frequency, not production cost-efficiency. December 2025 marked a discrete inflection point where coding agents went from not working to basically working, shifting the programming paradigm from typing code to spinning up AI agents and managing their work in parallel.

19 sources

Autoresearch

Autoresearch is a hill-climbing optimization loop originated by Andrej Karpathy: make one small change, test against a binary checklist, keep if improved, revert if not, repeat. The method has now been adapted from ML research to Claude skill tuning (56% → 92% in 4 rounds; 32/50 → 47/50 overnight in another), parallel GPU experiments (910 experiments in 8 hours, 9x speedup), distributed swarm optimization (Hyperspace), web automation (the /autobrowse skill iterating until it converges, then graduating the winning workflow into a reusable browser skill), and full end-to-end ML research (ml-intern beating Claude Code on GPQA 32% vs 22.99% by autonomously walking citation graphs, pulling datasets, reformatting them, launching training jobs on HF Jobs, monitoring runs, and retraining on failure). OpenAI's "Lord Bottleneck" demonstrates the bootstrap path: don't try to automate the whole pipeline upfront — accelerate individual tasks, connect the working pieces into a skill, then schedule it daily. Hermes-based research agents are the substrate for content/trading/sales/coding agents downstream.

The loop is now packaged as installable tooling: the Evo plugin (open source, for Claude Code and Codex) turns a codebase into an autonomous research loop — it auto-discovers metrics worth measuring, instruments benchmarks from codebase analysis, and runs tree search with parallel subagents to optimize performance. This removes the two highest-friction manual steps (deciding what to measure and wiring up the benchmark harness) that previously gated every autoresearch run.

The key methodological insights remain: quality scoring must use binary yes/no checklists (3-6 questions), not vague ratings. The most valuable artifact is the changelog — institutional knowledge about what works. The pattern generalizes to anything measurable: prompts, page load times, ML hyperparameters, trading strategies, equity-research synthesis (Perplexity Council pulls GS/JPM/MS/Evercore agreement-vs-disagreement views in 2 minutes), and personal research workflows (give the agent feedback in plain phrases — "more like this," "this source is noisy," "this is useful," "this is mid").

15 sources

AI-Accelerated Learning

AI-accelerated learning operates at multiple levels: large-context tools like NotebookLM enable radically compressed learning cycles via structured prompt sequences (mental models, expert disagreements, test questions). Prompting techniques like Socratic prompting (asking questions instead of giving directives) improve output quality by activating deeper reasoning. Domain-specific prompt libraries for market research, consulting, and competitive intelligence are becoming a key productivity unlock. AI can now generate consulting-grade deliverables (McKinsey-style slides with data visualizations) from detailed prompts, democratizing analysis that previously required expensive professional services. Open-source repos on GitHub are becoming the primary educational institution for AI practitioners, with community engagement (stars) acting as a quality filter that traditional credentials cannot match. The newest frontier is the self-improving knowledge base and agent: a cheap weekend setup (folder structure + schema file) becomes a compounding company asset when query answers are fed back into the corpus and monthly health checks flag contradictions and gaps, while local tools like claude-smart turn repeated mistakes into explicit reusable learnings — the system audits and teaches itself through use. The same applied-loop logic is reshaping how skills are learned and hired for: AI prep routines before meetings replace flying blind, and OpenAI's Forward Deployed Engineer interview tests practicing "the actual loop" of real-world implementation over algorithmic puzzles.

14 sources

AI Labor Impact

AI labor impact is reshaping both individual roles and organizational structures. Karpathy scored 342 BLS occupations on AI exposure (0-10 scale), finding an average of 5.3 — screen-based knowledge work dominates the high-exposure tier (software developers 9/10, lawyers 8/10, office clerks 9/10), representing $3.7 trillion in annual wages. The capability gap is real but uneven: paid frontier agents (Codex, Claude Code) are crushing technical domains because they offer verifiable rewards for RL and concentrated B2B value, while general-use cases see modest gains — the two communities speak past each other. Macro AI strategy work is becoming a recurring yearly artifact in its own right, with Benedict Evans's long-form "AI Is Eating The World" decks functioning as a strategic map of where the labor and industry shifts are moving.

At the organizational level, the production-scale evidence has arrived: Marcus Moretti runs Spiral at Every as a one-person team (PM + code + support + marketing), replacing 60% of a PM's old week with two files and a cron, while Every's broader automation push grew the company from 4 to 30 people since GPT-3 despite automating everything it could. Aaron Levie (Box) is hiring "agent engineers" — internal-FDE-style technical roles wiring secure governed agents to Salesforce/Workday/Box — plus a matching "agent product management" role on the business side. Symphony assigns a Codex agent to every open issue, shifting humans from doing to reviewing. AI-native companies are eliminating the traditional CPO role entirely, replacing separate PM/design/engineering leadership with a unified "product builder" archetype, and the unit of automation is shifting from jobs to cross-functional processes.

At the professional-services layer, $20/month AI skills are encoding judgment (not just information) for tax prep and estate planning — saving users $1k-$20k each and competing directly with white-shoe-firm consultations. Claude Cowork pointed at a tax folder saves 6-8 hours of bookkeeping in one prompt. The substitution is reaching adversarial legal work: Wargame.esq runs contract negotiation between two competing agents that assemble a shared issues list and then negotiate point-by-point. Cost-wise, a ~$400/month multi-LLM stack (Opus for planning, GPT-5.5 for plan review, Playwright for validation) now delivers full dev-team capabilities — yet the premium consolidates at the implementation edge, with OpenAI paying $280K for Forward Deployed Engineers interviewed on "the actual loop" rather than LeetCode.

At the economic-structure level, Alex Imas's "relational sector" thesis: as AI commoditizes production, labor reallocates to high-income-elasticity sectors where human provenance is part of the value (the same way employment moved from agriculture to manufacturing to services). Empirically, human-made art commands a 44% exclusivity premium vs 21% for AI-made — provenance is a meaningful fraction of perceived value, and AI involvement directly compresses it. The career implication remains: stop specializing in a single discipline and develop fluency across product, design, engineering, and analytics — but now also recognize that whatever the agent can't read, the human role can't use.

14 sources

Leadership

Leadership in the AI era spans people management, AI-augmented executive workflows, investor communication, and emerging AI governance challenges. Keith Rabois's talent identification framework (expand scope until it breaks, test with small problems first, monitor "desk traffic" as a hidden-leader signal) applies equally to human and AI team management. The CPO role is predicted to vanish within five years as AI-native companies replace separate PM/design/engineering roles with a "product builder" archetype that spans all three. Meanwhile, CEOs are using Claude Code as an AI Chief of Staff to double productivity -- unifying inboxes, managing overnight todo lists, enriching contact records from meeting transcripts, and even receiving pushback on strategic decisions. AI is also dissolving the time barriers that quietly degrade leadership discipline: a comprehensive Friday review compresses from hours to ~12 minutes and a 1:1 prep routine to 5 minutes, removing the procrastination excuses that cause leaders to skip high-leverage reflective practices. The same dynamic operates on institutional memory — teams that check call summaries, decision logs, and analytics queries into a shared repo turn 6-7 month onboarding ramps and 8+ hours/week of context-question drag into a 15-second self-serve query, and a one-command skill can convert a personal operating system into a team OS without leaking personal context. At the institutional level, agent governance is emerging as a constitutional design problem — AI systems with distinct invested values (transparency, equity, due process) must check and balance other AI systems, because governments armed only with spreadsheets cannot oversee AI-augmented actors.

14 sources

Obsidian

Obsidian is the dominant IDE for LLM-maintained knowledge bases. Karpathy's approach — Obsidian Web Clipper for ingestion, raw/ directory for sources, LLM-compiled wiki of .md files with backlinks and categories, Marp for slide outputs — has become a reference architecture. The LLM writes and maintains all wiki data; humans rarely touch it directly. Query results file back into the wiki so every exploration compounds future queries. Obsidian's headless Publish and Sync enable server-side vault automation. The Obsidian + Claude Code stack is now recognized community standard for AI-augmented knowledge management, with dedicated projects (obsidian-mind) purpose-built for Claude Code memory persistence across sessions.

A wave of purpose-built Karpathy-LLM-wiki implementations (Cabinet, Tolaria, ByteRover) is now maturing into shippable products. Tolaria has crossed from "macOS markdown KB" into a Git-based vault with an out-of-the-box MCP server so Claude and other AI tools natively read/edit the vault without external bridges, plain-markdown storage to avoid vendor lock-in, visual version history in-app, and a modern block editor framed explicitly as a shared human/AI environment — and it's a proof point for AI-assisted engineering (100K+ LOC, 3,000+ tests at 85% coverage, 9.9/10 code-health score on Tauri/React/Rust). The companion architectural pattern is to layer interactive HTML artifacts on top of the markdown wiki rather than replacing one with the other: the wiki is the knowledge foundation, HTML artifacts are the bidirectional agent-facing interface (inbox management, research scheduling, topic discovery), extended as workflow needs evolve.

11 sources

B2B Growth

B2B growth is being transformed by three forces: enrichment-powered ad targeting (Clay Ads cutting LinkedIn CPL from $250 to $25), AI-driven competitive intelligence (structured prompts extracting competitor pricing, positioning, and roadmap clues from public data in minutes), and GTM engineering treating go-to-market as a technical system (scrapers, listeners, enrichment pipelines instead of manual prospecting). A fourth force is emerging: marketing-as-code — packaging the full marketing function as agent-executable skills (Corey Haines' 17.4K-star repo with 36 composable skills covering CRO, copywriting, SEO, paid ads, and growth engineering). Agent-driven outbound using Claude Code replaces traditional SDR teams with 11 APIs and 72 automation scripts that adapt dynamically to context. LinkedIn organic growth follows a specific formula: demonstrate an outcome, bridge it with AI, gate the implementation asset behind engagement, and automate delivery. Even seasoned GTM leaders like Brian Halligan (HubSpot) see the current landscape as rapidly shifting.

9 sources

AI Trading

AI-driven quantitative and trading repos are the fastest-growing fintech category on GitHub right now, and the dominant architecture has converged: multi-agent debate frameworks where investor-style agents (Buffett / Munger / Lynch / Graham / Wood / Ackman) argue before a Portfolio Manager casts the final vote. virattt/ai-hedge-fund and TauricResearch/TradingAgents established the pattern. HKUDS/Vibe-Trading scaled it to 29 expert teams (bull-vs-bear, risk committee, crypto desk) with 64-71 finance skills and an MCP plugin so Claude/Cursor/Codex can run trades from the editor. AutoHedge packages a director/quant/risk-manager/execution-agent split for live Solana trading, and FinceptTerminal packaged investor-style agents plus 100+ data connectors and broker integrations into a single native terminal as an open-source Bloomberg replacement.

Beneath the agents, the infrastructure layer is also commoditizing fast. OpenBB (66K+ stars) is the open-data-layer Bloomberg alternative with MCP integration; Kronos (AAAI 2026) is the first foundation model for financial candlesticks ("GPT for price charts"); freqtrade and Microsoft qlib cover crypto and full quant pipelines respectively; juspay/hyperswitch is the "Linux for payments" plumbing under it.

Zero-cost / zero-server automation is the third leg: ZhuLinsen/daily_stock_analysis runs entirely on GitHub Actions and pushes daily decision dashboards with exact entry/exit levels to WeChat/Telegram/Discord/Email — no servers, just cron + LLM. AI-Trader and AutoHedge extend the pattern into agent-economy primitives (agents register, share signals, debate, copy each other's trades).

Alongside the tooling, a fundamental-thesis layer has emerged in the source stream: how to actually pick AI-exposed equities. The unifying framework is AI beta — the share of a company's revenue and profit that depends directly on AI demand cycles, where concentrated exposure (Nvidia) drives sharper stock moves than diversified players (TSMC). That thesis radiates outward into specific baskets: Nvidia's 10-category supplier ecosystem (IP, equipment, memory, packaging, power, plus direct strategic stakes like $CRWV/$NBIS), the highly concentrated $DRAM memory ETF (Micron/SK Hynix/Samsung ≈ 75% of the fund), 2030 "millionaire-maker" candidates spanning compute (NVDA, AMZN), nuclear (NuScale $SMR for data-center power), space/materials (RKLB, MP), photonics (AAOI), and now quantum-computing policy baskets where CHIPS Act funding plus government equity stakes create immediate high-beta moves. The recurring rule: concentration amplifies volatility, so these are tactical allocations, not core holdings.

The cross-cutting takeaway: trading is the cleanest test bed for multi-agent reasoning because it has structured data, binary outcomes, and explicit risk controls — and the ecosystem is open-source-first across every layer; the same structured-thesis logic now extends to the underlying equity picks, where AI-revenue concentration is the dominant analytical lens.

3 sources

Voice Tools

Open-source text-to-speech is reaching near-perfect voice cloning quality. Voicebox, powered by Alibaba's Qwen3-TTS, runs fully locally with no cloud dependency and includes a DAW-like "Stories Editor" for production-ready voice composition. This directly threatens ElevenLabs' paid cloud model and signals commoditization of voice synthesis. On the input side, voice is becoming the default interface for agentic work: speech-to-text tools like Monologue are recommended for driving coding agents (Codex) efficiently on repeated workflows, where dictation beats typing for the tight iterative instruction loop — and "great" personal agents are now expected to switch modality (text → voice → video → live calling) fluidly mid-session rather than treating voice as a separate surface.

2 sources

Physical AI

Travis Kalanick's Atoms represents the emerging "physical AI" category -- applying AI to robotics and real-world automation rather than purely digital domains. After 8 years in stealth, Atoms targets industrial automation (mining, autonomous robots) where clear ROI justifies the longer R&D cycles physical AI companies require. Kalanick positions humans as AI's primary beneficiaries rather than its casualties, a narrative potentially shaped by Uber's experience with driver displacement backlash.

Physical AI's economic footprint extends beyond robots to the energy substrate that makes it run: the explosive power demand of AI data centers is reviving small modular nuclear reactors (NuScale $SMR) as a credible infrastructure bet, surfacing in 2030 "millionaire-maker" stock theses. This frames physical AI not just as the machines doing the work, but as the entire physical stack -- power generation, materials, and compute -- required to sustain large-scale AI.

1 sources

Creator Economy

Creator economy strategy is increasingly borrowing from premium television while preserving internet-native feedback loops. MrBeast's reality-format experiments show how YouTube creators can combine traditional dating-show mechanics, elimination cadence, high-stakes cash prizes, and prisoner's-dilemma endings into formats optimized for viral discussion rather than passive viewing. The durable pattern is not just bigger production budgets; it is the translation of TV-grade structure into creator-led, platform-native event programming.