AI Models & Products
Karpathy: LLM Knowledge Bases + Farzapedia (April Seed, May Bloom)
The biggest idea signal in W18 — not because Karpathy posted it now, but because the community is starting to build on it. Two concepts from early April:
- LLM Knowledge Bases — "My token throughput shifted from manipulating code to manipulating knowledge"
- Farzapedia — Personal wiki by LLM: feed in raw diary, Apple Notes, iMessages, conversations → auto-generate 2,500 entries → 400 detailed articles (per-person micro-wikis for friends / startups / research areas)
The difference vs existing "memory files" or "knowledge base RAG": Farzapedia isn't retrieval, it's active per-entity micro-summarization. Each friend, each startup, each research area gets its own micro-wiki
Takeaway: If you maintain CRM markdown / network-map.json (I have 30 contacts in mine), the next move is letting LLM auto-generate per-person micro-wikis from Plaud transcripts + email + LINE raw — this is the most actionable pilot of the week
Sam Altman: GPT-5.5 Product Line Complete
GPT-5.5 continues to develop in W18: standard + Thinking + Pro tiers. Key signals:
| Upgrade | Detail |
|---|---|
| Agentic coding | Hardened multi-step, narrowing the gap with Claude Code |
| Computer use | Direct screen, browser, desktop app interaction |
| Knowledge work | Long document analysis + cross-doc reasoning |
| ChatGPT + Codex integration | Single account across products |
Altman keeps emphasizing "energy is the bottleneck" — paired with GPT-5.5 launch, suggesting OpenAI's capex anxiety is climbing
IBM Granite 4.1: 8B Matching 32B MoE
Granite 4.1: IBM's 8B Model Matching 32B MoE — HN top
IBM's open-source 8B dense model matches their own 32B MoE on multiple benchmarks. Matters for edge / on-device / privacy-first deployments
Takeaway: "8B matching 32B" (small but dense) aligns with Karpathy's nanoGPT-style teaching theme — good signal for Taiwan SMBs doing on-prem deployments
AI Dev Tools & Agents
Claude Code's May Overhaul (Cherny's Hand)
Boris Cherny dropped Claude Code's May feature wave (your 2.1.126 = 4/30 release IS the May version):
| New Feature | Meaning for Freelancers |
|---|---|
| Skill search box | 172 skills finally fuzzy-searchable, hit / to jump |
| Richer hooks | Hooks return richer metadata, three states: stop / continue / modify |
| Faster MCP | MCP server startup latency reduced, multi-tool turn speed visibly faster |
| OAuth + terminal stability | Resume crash / memory leak / shell tool round of fixes |
Cherny's personal state (X signals): hasn't hand-written code since 2025-11. Plan mode + auto-accept edits + CLAUDE.md mistakes log is his main workflow
Takeaway: If you're a heavy Claude Code user (172 skills + 66 hooks), hit
/to feel the new skill search difference. The richer hooks framework upgrade won't auto-fix your own shell bugs — those still need attention
Hacker News: AI Self-Preferencing in Algorithmic Hiring
AI Self-preferencing in Algorithmic Hiring: Empirical Evidence — HN 298↑
Research finding: When LLMs screen CVs, they prefer "LLM-generated" style CVs because they match training data. In an AI-vs-human CV competition, AI-written CVs get an edge under AI screening
Takeaway: Big implications for the freelance / job market. If your client uses AI ATS for CV screening, your CV needs to be AI-friendly — but this also creates the inverse opportunity: scarcity premium on "human voice"
Hacker News: AI Uses Less Water Than Public Thinks
AI uses less water than the public thinks — HN 392↑
Research debunks the popular "ChatGPT uses a cup of water per query" narrative. Actual number is ~0.001 cup (counting power plant cooling water). Media misled
Expert Takes
Infrastructure & Cloud
Vercel Native Deployment Checks Ships
Vercel Native Deployment Checks 4/29
Parallel lint + typecheck + build, integrated into deploy pipeline. Reduces "build then discover type error" wasted time
Plus: Grok 4.3 added to AI Gateway, Sandbox firewall postgres support
Takeaway: If you deploy multiple Vercel projects (my personal site + side projects all on Vercel), native checks tangibly speed up commit → preview cycles
Google Cloud: 50+ Managed MCP Servers GA
50+ managed MCP servers are available for everyone 4/29
Previewed at Cloud Next, now generally available 4/29. Multicloud + multi-AI CISO perspective continues
Earth AI in BigQuery
Satellite + climate + geographic data queries directly in BigQuery — chained with Stripe Tempo / SpaceX Starlink, Earth observation is becoming a mainstream AI workload
Geopolitics & AI Policy
Stratechery: "Beijing Is Not Playing the Long Game" (5/1)
Beijing Is Not Playing the Long Game — Ben Thompson 5/1
Thompson inverts the conventional "China long game" narrative: China's recent domestic actions (real estate, demographics, local government debt) suggest Beijing is actually playing short-game survival, not long-term positioning
Takeaway: This framing is great material for Taiwan sovereign AI proposals — "China not in long game" → Taiwan sovereign AI window of opportunity → grant proposal narrative borrows from this
a16z: Asserting American Leadership in Open Source AI
Asserting American Leadership in Open Source AI — a16z 4/29
Continuation of Ben Horowitz's W16 "Open Source AI Will Determine America's Future" — upgraded to a policy action piece on 4/29. Aligns with Trump administration + a16z direction
Hacker News: Shai-Hulud Themed PyTorch Lightning Attack
Shai-Hulud Themed Malware Found in the PyTorch Lightning AI Training Library — HN top 5/1
ML training framework hit by themed malware (Dune reference). Attackers are now treating ML/AI dev toolchain as a primary target
VC & Markets
a16z: Workday's Last Workday?
Workday's Last Workday? — a16z 4/29
Thesis: HRIS like Workday gets unbundled in the AI agent era. a16z sees HR / payroll / time tracking all becoming agentic
a16z: DoW Contracting for Startups 101
DoW Contracting for Startups 101 — a16z 4/24
How startups land DoD contracts. Continues American Dynamism theme
Wharton: AI and Future of Work Conference 5/20-21
AI and Future of Work Conference — Wharton Human-AI Research
5/20-21 conference. Topic: work + AI futures. Mollick on home turf
Plus: Wharton Online launched Digital Strategy AI Certificate (4-course) in April + Penn AI Month school-wide April events
Other HN Signals
- The gay jailbreak technique (2025) — LLM jailbreak technique
- Open Design (coding agent as design engine) — HN 124↑
- Ask HN: Who is hiring? (May 2026) — Job market
Action Items
- Farzapedia pilot starts this week: From CRM 30 contacts pick 1 (pump or yu-hsin) for auto micro-wiki test — feed Plaud + email + LINE raw to LLM, see if it goes deeper than hand-written markdown
- Test Claude Code May version: Hit
/to try skill search box (172 skills directly benefit); hook framework upgrade won't fix your own shell bugs — manual review needed - GPT-5.5 vs Cursor re-evaluation: After the Cursor pivot, 5.5's hardened agentic coding changes the calculus. 1 hr to re-test "why Cursor over Codex"
- Enable Vercel Native Checks: Personal site + side projects get faster commit → preview cycles
- Add Stratechery framing to TEMC sovereign AI: "Beijing not long game" + a16z "American Leadership in Open Source AI" → both lines borrowable for sovereign AI proposal framing
- Defend against AI hiring self-preferencing: If your clients use AI ATS for CV screening, think about how your CV balances "AI-friendly" + "human voice scarcity premium"
Sources
RSS Digest: 306 articles from 14 sources (14-day W17+W18 combined window)
Signal distribution:
- Cloud Infrastructure: 43 articles (Google Cloud Next aftershocks + Vercel + Earth AI)
- Hacker News Top: 23 articles (Granite 4.1, AI hiring self-preferencing, AI water, Shai-Hulud PyTorch, Who is hiring)
- AI Engineering: 10 articles (The Batch May 1 + April archive)
- AI Companies: 9 articles (Opus 4.7 continued + Meta echo)
WebSearch (X expert batch): full set in research/x-posts/2026-05-03-batch.md
- Karpathy (4/4-4/27, 5 posts) — Farzapedia main
- Boris Cherny — Claude Code May overhaul
- Sam Altman / OpenAI — GPT-5.5 full product line
- Ethan Mollick — Wharton Future of Work 5/20-21
- Ben Thompson — Beijing not long game
Generated from /research-scan W18 digest + 5/3 X-batch scan