Paul Graham’s August 23 tweet is not a job posting. It is a “learn the substrate” post dressed as nostalgia.
If he were 17 today, @paulg wrote, he would learn how to build LLMs from scratch and train as powerful a model as whatever hardware he could access. Follow-up: he would not start a startup yet — better ideas grow out of understanding later, not out of what you knew at 17.
The thread hit 833K views on X and 326 points on Hacker News. The next day, Yann LeCun posted a competing “if I were 17” answer — ~194K views by August 24 — that reframed Graham’s substrate advice as incomplete for physical intelligence.
explainx.ai’s read: Graham is describing curiosity and depth in token land, not a career plan. HN is describing labor markets and GPUs. LeCun is describing a second curriculum for everything transformers do not natively touch. All three can be true — and they point at different stacks.
TL;DR
| Question | Direct answer |
|---|---|
| Literal advice? | Learn transformers by building; defer startup |
| Career advice? | Only indirectly — depth beats “wrapper startup” later |
| HN main pushback? | Few training jobs; hardware gate; survivorship bias |
| HN main agreement? | Toy training teaches more than API-only use |
| Best resources? | Karpathy, Raschka, nanoGPT, CS336, rasbt repo |
| Relation to prior PG post? | Aug 3 = rewards; Aug 23 = what to study |
| LeCun's counterpoint? | Study why LLMs write essays but can't clean rooms — pursue architectures beyond LLMs |
| Is "no hands in datacenter" the whole story? | LeCun: even a humanoid wired to a datacenter LLM still won't clean your room |
| Does videogame RL at scale fix it? | LeCun called that pipedream "charmingly naïve" |
| Where does "world model" fit? | LeCun's hint — token predictors lack consequence modeling; see world models guide |
What Graham actually said
Two tweets, August 23, 2026:
- If I were 17, I'd learn how to build LLMs from scratch, and then train ones as powerful as I could with whatever hardware I could get access to.
- Notice that what I would not do is try to start a startup. Instead I'd build the foundation of knowledge to base a startup on later.
The replies are the usual PG mix — play computer games with the model, maybe build robots, “to learn how wheels are made” when someone asks why not use existing APIs.
None of that changes the core claim: at 17, depth beats incorporation.
That is consistent with how he talked about learning HTML in 1995 — not because every teenager became a web startup founder, but because the substrate of the next decade was the web.
LeCun's counterpoint: essays yes, bedrooms no
On August 24, 2026 (~12:10 AM), Yann LeCun posted his own age-17 answer. It is not a rejection of learning — it is a different foundation:
I would try to figure out why LLMs can write my essays but not clean my bedroom. Then I would study topics in college and grad school that could help solve that problem. I'll figure out a set of methods and architectures beyond LLMs that can quickly learn to perform physical tasks as efficiently as humans and animals. That last item is also what I would if I were 30, 40, 50, or 66 years old 😉
Same “foundations before hype” family as Graham. Different substrate: not transformers and loss curves, but sensorimotor intelligence and whatever comes after next-token prediction.
What the replies argued — and how LeCun answered
| Reply | Claim | LeCun's pushback |
|---|---|---|
| Dante Blank | LLMs can't clean rooms because they run in datacenters without hands — don't infer a new theory of intelligence from that | You can hook a humanoid to a datacenter LLM and still won't get a room-cleaning robot |
| Nicholas Pipitone | Train RL on videogames at ~30 Hz across millions of titles | "This is charmingly naïve" |
| Samir Alibabic | Give LLMs a body and they can clean | "Nope. Not a chance in hell." |
| Ralph Martello | Collect a year of sensor-suit data from 100 people | Why do young animals and humans learn without 150 years of perceptual data from others? |
| george_barcino | Multiple simultaneous sequences beyond words — transformers should handle it | "The word 'just' is doing a lot of heavy lifting… Hint: ask yourself why 'world model' is becoming such a buzzword" |
The through-line: bodies are necessary but not sufficient. LeCun is not saying "wait for better robots." He is saying token predictors trained on discrete symbols do not natively model continuous physical consequence — the same argument he made in the July 2026 physical-agents thread and at AMI Labs around JEPA-style world models.
Two curricula, one honest split
| If your 17-year-old goal is… | Graham's path | LeCun's path |
|---|---|---|
| Ship software agents, evals, harnesses | Build LLMs from scratch → understand logits, context, fine-tuning limits | Still useful — but not the bottleneck LeCun cares about |
| Robotics, manipulation, embodied AI | Teaches you why API wrappers lie; does not teach contact physics | Study architectures beyond LLMs; sample-efficient motor learning |
| "Digital AGI" in knowledge work | Strong overlap — depth beats wrapper startups | LeCun: optimistic about AI, "just not about LLMs" for physical domains |
Graham's advice and LeCun's advice do not contradict if you treat them as environment-specific. The mistake is reading either tweet as a universal major.
For builders already on the Graham path, the LeCun thread is a reminder: training literacy in language models does not transfer to torque, friction, or long-horizon planning in clutter. That is why YC Paper Club's robotics gap post and DYNA-2's scaling-law read keep landing — the sensorimotor void is still open even while HN debates nanoGPT on a GTX 1660.
What HN got right (and Graham elides)
The top-voted skeptic thread on HN is not anti-learning. It is anti-confusing learning with a labor market.
Concrete counters from the discussion:
- Almost no companies train frontier models. Most teams fine-tune, route, or wrap APIs. Training jobs are tiny and credential-heavy.
- Hardware is the gate. Meaningful pretraining wants serious GPU spend; a teenager without budget hits a long plateau before “interesting.”
- Survivorship bias. Graham’s advice comes from someone who won a lottery and advises lottery entrants.
Fair. Also incomplete.
What HN also got right (the pro-learning camp)
Several high-signal comments share the same shape: I trained a toy model / built a bad web server / retyped a database — never made money from that exact skill, but I understand the layer above forever.
Examples from the thread:
- Browser in 2000 vs LLM in 2026 — few people employ browser authors; many employ people who understand the web.
- rasbt/LLMs-from-scratch — cited repeatedly as the weekend-to-month on-ramp.
- GTX 1660 training runs — you can learn on modest hardware if you accept “not frontier.”
- Felix Rieseberg’s Language Model Builder — interactive, no account, free tutorial HN surfaced in the Show thread.
The middle position explainx.ai agrees with: build small on purpose. Not to compete with OpenAI — to stop treating the model as magic.
A sane curriculum (if you actually are 17 — or 37)
Graham did not specify steps. HN and the existing canon do:
| Stage | What | Outcome |
|---|---|---|
| 1. Mechanics | Matrix math you can use; Python; one MNIST or tiny char model | Loss goes down; you touched gradients |
| 2. Transformer | Karpathy “Zero to Hero” + nanoGPT or Raschka book | You know what a block is |
| 3. Scale reality | Try fine-tuning a small open model locally (Ollama, llama.cpp) | You feel data and compute limits |
| 4. Harness | Loop engineering, evals, context | You learn where APIs lie |
Skip stage 4 and you risk becoming the person who forwards model output unread — the meat proxy pattern applies to “I don’t know how it works but the loss looked low.”
How this connects to Graham’s other 2026 LLM takes
Same month, different lens:
| Date | Post | Lesson |
|---|---|---|
| Aug 3 | Math vs writing / verifiable answers | Training follows graders |
| Aug 23 | This post | Substrate before startups |
| Aug 24 | LeCun reply | Token training ≠ physical intelligence |
| Aug 18 | 1980 AGI test | Threshold labels lie |
| Jul 7 | Fable vs GPT-3 speculation | Capability jumps reshape plans |
Building an LLM from scratch does not contradict “math trains faster because it is verifiable.” It explains why reading the code and loss function matters when you later build products in ungraded domains like writing or design. LeCun adds the parallel claim: essays are verifiable in tokens; bedrooms are not — and scaling the former does not automatically unlock the latter.
What not to do with this tweet
- Do not read it as “drop out and train a 70B.” That is Twitter hero fiction.
- Do not read it as “startups are bad.” Graham runs a startup factory; he said not yet.
- Do not skip evals because you once ran nanoGPT. Training literacy ≠ shipping literacy — same gap as cheap answers vs trust.
Related on explainx.ai
- Yann LeCun: LLMs vs physical agents and Moravec's paradox
- What are world models? Starchild-1 and beyond
- Paul Graham: LLMs crush math but lag at writing
- YC Paper Club: why robotics still isn't solved
- DYNA-2 — robotics scaling laws, read honestly
- Build your personal AI system — local hardware and workflows
- What is llama.cpp? Run models locally
- Qwen 3.6 27B local with llama.cpp and OpenCode
- Ollama $88M — open models go mainstream
- 2x not 10x — realistic LLM coding gains
- Students choosing AI startups over internships
- Andrew Ng — AI engineering skills map
- Meat proxy — relay etiquette for model output
Sources
- @paulg on X — August 23, 2026
- @ylecun on X — August 24, 2026 reply thread (~194K views)
- Hacker News discussion
- Andrej Karpathy — neural networks Zero to Hero
- Sebastian Raschka — Build a LLM From Scratch
- rasbt/LLMs-from-scratch
Career-market claims in the HN thread are commenter opinions, not hiring data. Graham's tweets are the primary source for his advice. Follow @explainx_ai for builder-oriented reads on model craft.