How to use this document: Ravi will share this with a new Claude session to provide context. The new session should read this fully before starting work. The key files, decisions, and priorities below are the result of extensive research and deliberation — don't re-derive them from scratch unless Ravi asks to revisit something.
1. Who is Ravi
- Current role: ML Engineer at Meta. 13+ years of software engineering experience.
- Location: Bay Area. Open to staying in Bay Area. Previously considered Seattle for state-tax reasons (Anthropic SPV will trigger a tax event), but Bay Area is now the primary target for the next job.
- Anthropic SPV: Ravi holds an investment in Anthropic via a dual-layer SPV. This is a financial investment, NOT employment — Ravi has never worked at Anthropic. The SPV will eventually trigger a taxable event (either when shares convert or cash is distributed).
- Core worldview: AGI is coming within a few years. This is treated as settled fact, not speculation. All career and investment decisions are reasoned from this premise.
- Communication preference: Be completely honest. Push back when you disagree. No hedging or sugarcoating.
2. The two parallel tracks
Ravi is pursuing two things simultaneously:
Track A: Find the best next job
Priority order for what matters in the next role:
- The thing the company is working on is genuinely useful for the future (according to Ravi's estimate of what matters in an AGI world)
- Great people — really talented, nice, inspiring, good moral character
- Learning valuable skills — things that will compound in value
- (Lower priority) Equity upside — a 10x is fine, doesn't need to be 100x
Track B: Build toward AI-powered investing
Long-term goal: build an AI system that identifies long-term investment opportunities with near-zero downside risk (an "AI Warren Buffett"). This is being pursued in parallel — slowly building skills, tools, and knowledge while working the day job. The existing AWB (Automated Warren Buffett) project lives at /Users/ravf/projects/work/research/investments/awb/.
The two tracks connect: Learning LLM training deeply (Track A) directly enables building a better AI investor (Track B). The ideal company for Track A is one where the skills and knowledge also accelerate Track B. This is why companies doing RL, continual learning, or AI-for-science rank highest — their methodology transfers to investing.
3. Company research done so far
Extensive company research has been done and is hosted at ravikant.dev/companies/. All files live in /Users/ravf/projects/work/companies/.
Pages built
| Page | File | What it contains |
|---|---|---|
| Main candidate list | index.html | Sortable table of ~30 companies with 100x/Learn/InvAlign/Fit scores. Hub page linking to all other views. |
| Bay Area Sweep | bay-area-sweep.html | Most comprehensive view. 76+ Bay Area startups across 15 categories, scored on Useful/People/Skills/Fit with weighted total. Filterable by category and min funding. This is the primary reference. |
| Seattle-Compatible | seattle.html | Seattle-filtered view with Tier A/B/C classification. Less relevant now that Bay Area is the target, but kept for reference. |
| Anthropic Pathway | anthropic-pathway.html | Deep dive: interview process (5 stages), values interview prep, comp data, referral strategy, 8-12 week prep plan, specific Seattle-listed roles. |
| Together AI | together-ai.html | Deep dive: founders (Tri Dao, Chris Ré, Percy Liang), products, funding, comp, competitive position, honest fit assessment. |
| AI Chip Startups | ai-chips.html | Probability table for 14 chip companies, why most fail (CUDA moat), what Anthropic uses (TPU + Trainium), roles + prep plan, concentration risk warning. |
| Continual Learning | ai-continual-learning.html | Adaption Labs, Sakana AI, Liquid AI, Cartesia, Periodic Labs, DeepMind CL team. Why CL is the right bet for investing thesis. |
| LLM Training Strategy | llm-training-strategy.html | The project plan. 3 portfolio projects, 12-week timeline, required reading, company mapping. See section 4 below. |
Top companies by weighted priority (from Bay Area sweep)
| Company | Category | Why it ranks high |
|---|---|---|
| Periodic Labs | Small Frontier Lab | AI scientist — RL on real experiments. Strongest mission overlap with investing thesis. Fedus + Cubuk founders. |
| Physical Intelligence (π) | Robotics | Best robotics ML team (Levine, Finn). Genuinely important for the future. |
| World Labs | Robotics | Fei-Fei Li. World models / spatial intelligence. |
| Anthropic | Frontier Lab | Best safety culture. Real Seattle office. Concentration risk with SPV. |
| Harmonic | Small Frontier Lab | Math reasoning as key unlock. Vlad Tenev + Tudor Achim. |
| Guide Labs | Small Frontier Lab | Interpretability as product. Julius Adebayo (Google Brain/MIT). |
| Mechanize | AI Infra | RL environments for job automation. Direct methodological transfer to investing. |
| Sequent Research | AI Safety | Geoffrey Irving (RLHF pioneer). Nonprofit alignment org. Berkeley. Just announced Jun 2026. |
| Together AI | AI Infra | Tri Dao, Chris Ré. Best ML-systems learning. SF-only. |
| insitro | AI for Science | Daphne Koller. ML drug discovery. South SF. |
4. The LLM training project plan
Full details at ravikant.dev/companies/llm-training-strategy.html. Summary:
Three portfolio projects (each → GitHub repo + blog post)
| # | Project | Time | Cost | Key skill | Maps to |
|---|---|---|---|---|---|
| 1 | Pretrain 125M model from scratch → post-train with GRPO | 4 weeks PT | ~$200-500 | Pretraining + post-training RL | Anthropic, Periodic Labs, Harmonic, Mechanize, OpenAI, Physical Intelligence |
| 2 | LLM serving system from scratch (continuous batching, PagedAttention, speculative decoding) | 3 weeks PT | ~$50-100 | Inference systems | Together AI, Fireworks, Anthropic, Groq, Anyscale, Modal |
| 3 | Scaling law reproduction with FSDP (train 5 models, plot IsoFLOP curves, fit Chinchilla power law) | 3 weeks PT | ~$500-1500 | Distributed training + scaling | Anthropic, OpenAI, Thinking Machines, Lila Sciences, Together AI |
If only one project: Project 1 (GRPO) is the highest-leverage single project. It covers the most in-demand skill and applies to the most target companies.
Timeline: 12 weeks part-time alongside Meta
- Weeks 1-2: Reading sprint (papers listed in the strategy doc)
- Weeks 3-6: Project 1 (pretrain + GRPO)
- Weeks 7-9: Project 2 (inference serving)
- Weeks 10-12: Project 3 (scaling laws)
- Week 13+: Apply (batch applications, 8-10 companies in a 2-week window)
Required reading (before projects)
DeepSeek-R1 (GRPO), vLLM/PagedAttention, Chinchilla, Llama 2+3, FlashAttention 1+2, InstructGPT (RLHF), DPO paper, Karpathy's "build GPT from scratch." Full list in the strategy doc.
Application strategy
- Batch applications after all projects done — competing offers create leverage
- One role per company, with referral if possible
- Start referral outreach at week 6 (after Project 1 ships) — don't wait for all 3
- Blog posts go on ravikant.dev + LinkedIn
5. The investing track (parallel, slower)
- Existing work: AWB project at
/Users/ravf/projects/work/research/investments/awb/— SEC filing fetcher, cached data (~21 GB in Google Drive), analysis exports - Stock Watch dashboard:
/Users/ravf/projects/work/research/investments/stock-watch/ - AI Stock Universe:
/Users/ravf/projects/work/research/investments/ai-stock-universe.html - Investment reports: Various HTML reports under
/Users/ravf/projects/work/research/investments/ - SuperInvestors app: Separate project at
/Users/ravf/projects/superinvestors-app-deploy/— 13F tracker - Approach: Don't rush. Build tools, study companies, develop long-term theses. The AI-investing thesis is long-duration (years, not months). Keep building the infrastructure while doing the career projects.
6. Key decisions already made
- ✅ Bay Area is the primary target geography (Seattle was deprioritized)
- ✅ NYC is hard-excluded
- ✅ Priority order: useful work > great people > learning > equity upside
- ✅ 10x equity is fine; doesn't need to be 100x
- ✅ "I'll take any role" was pushed back on — aim for real ML/RE engineering roles that match seniority
- ✅ Project 1 (GRPO) is highest priority if only doing one project
- ✅ Anthropic concentration risk acknowledged — already hold SPV exposure
- ✅ "BUILD YOUR OWN" (AI-for-investing company) remains on the list as the baseline to beat in 2-3 years
7. Technical setup
| Item | Location |
|---|---|
| Work hub repo | /Users/ravf/projects/work/ → GitHub: ravikant0909/work |
| Companies research | /Users/ravf/projects/work/companies/ |
| LLM learning content (book) | /Users/ravf/projects/work/docs/llm-learning/ |
| Investment research | /Users/ravf/projects/work/research/investments/ |
| Live site | ravikant.dev (Cloudflare Pages) |
| Deploy command | /Users/ravf/projects/work/bin/deploy (never run wrangler pages deploy directly) |
| CLAUDE.md | /Users/ravf/projects/work/CLAUDE.md — read this first for repo rules |
8. What the next session should work on
Depending on what Ravi asks, the most likely next steps are:
- Start Project 1 (GRPO) — set up GPU cloud, build tokenizer, start pretraining. The strategy doc has week-by-week breakdown.
- Deep-dive a specific company — Ravi may want Anthropic-pathway-style research on Periodic Labs, Harmonic, Physical Intelligence, Mechanize, or another company from the sweep.
- Investment research — continue AWB work, stock analysis, thesis development. Separate from the career track.
- Add more companies to the sweep — Ravi finds new startups to add (like Sequent was added).
- Referral outreach planning — identify specific people at target companies to reach out to.
Things to NOT redo: The company research, scoring, and project plan are done. Don't re-derive them unless Ravi explicitly asks to revisit. Don't re-explain the AGI worldview, the investing thesis, or the career priority order — they're all documented above and in the various HTML pages. Start from where we left off.
9. Open questions (not yet resolved)
- Which specific startup did Ravi mean by "Anthropic signed with a startup" for compute? (Unresolved from the chip deep-dive.)
- Jerry Tworek's startup — existence not confirmed. Ravi mentioned it but no source provided.
- Adaption Labs — founders and funding not verified. Needs LinkedIn/Crunchbase check.
- Astor (AI-powered investing, Seed, SF) — found in the sweep but not researched. Could be closest to the investing thesis of any real company.
- GPU cloud provider for projects — not yet set up (Lambda, Together, Modal all options).
- When to start: no specific start date committed for Project 1.