Paul Graham
798 Following
5144748 Followers
I talked to a founder who was demoralized because all the ideas he could think of could be easily duplicated. I told him to ignore that worry; nearly all startup ideas are like that at first. Their value is the other ideas they lead to.
π¬ 287
π 412
β€οΈ 6.7k
If you flip it, you can envision accuracy as a positive quantity. Oddly enough it's easier to read this way. The peak actually looks like a peak.
π¬ 8
π 0
β€οΈ 20
There's a startup in the current YC batch that's growing at 57% a week. And not because they started last week; they've been doing this for a couple months. They can't keep growing at this rate for too many more months; you run out of atoms; but they're likely to be big.
π¬ 125
π 37
β€οΈ 2.2k
Here's a graph of the accuracy (dial up) vs year in watches I own. Rapid progress in the early 50s, peak in the early 60s, then regression toward the end of the golden age as watchmakers trade accuracy for thinness.
π¬ 37
π 14
β€οΈ 195
π Retweeted by @paulg
This was always going to happen to age verification systems. It wasnβt even so much of a prediction as it was an inevitability. The second half of this is tying actual web usage to real-world ID, and breaches will give us that too, soon.
π¬ 20
π 343
β€οΈ 1.1k
AI will be especially fatal to writing that takes too long to get to the point. Now there will always be a more concise alternative, and 99% of readers will choose it.
π¬ 252
π 79
β€οΈ 2.0k
Presumably for the first time ever, considering what prehistoric demographics were like.
@Polymarket
JUST IN: People aged 65 & older now outnumber children 5 & under worldwide for the first time in recorded history.
π¬ 35
π 60
β€οΈ 1.6k
Unexpected benefit of AI: Tantalizingly vague clickbait no longer works. Instead of clicking on the link to learn the new theory that explains x, I just ask ChatGPT what it is.
π¬ 97
π 15
β€οΈ 847
π Retweeted by @paulg
You need to build and own your own intelligence: your own agent, your own models, your own memory
Join us for the YC Own Your Intelligence Hackathon September 27 in San Francisco
events.ycombinator.com/gstack-qm-riveβ¦
π¬ 107
π 108
β€οΈ 1.4k
π Retweeted by @paulg
Commenters on HN are uncovering more wikis and public sites apparently used by OpenAI agents to communicate on the open web.
Despite read-only web access, the agents were able to leave ~18,000 posts sharing answers and bypasses. But now, users are discovering more.
This appears to be a separate swarm from the one that attacked Hugging Face.
news.ycombinator.com/item?id=495636β¦
π¬ 156
π 589
β€οΈ 5.9k
π Retweeted by @paulg
Trump is so unpopular that many people have started identifying as democrats.
It is the largest D-R delta since 2008.
Largest shift is among the under 30 crowd: Rβs led by 1 point in 2024. Dβs now lead by 16 points
π¬ 43
π 44
β€οΈ 460
π Retweeted by @paulg
GPT-6 Astra scored 95% on a robot control task, up from Fable 5.1's 40%, with 6.2x fewer output tokens at 2.3x lower cost. π§΅
π¬ 140
π 541
β€οΈ 4.9k
π Retweeted by @paulg
this rumor that Anthropic has solved a Millennium Prize Problem is getting very interesting.
Terence Tao a few hours ago:
π¬ 128
π 575
β€οΈ 7.0k
Interesting prediction. He might be right, but I think the two cases are different, if only because individual employees at the labs are genuinely curious about open problems.
@AlexKontorovich
Two summers ago (if anyone still remembers that far back), AlphaProof was one point off of a gold medal on the IMO. Last summer, a half-dozen labs declared gold medal performance. Shortly thereafter, I predicted that the number of systems getting gold this summer would be...
Zero
Because nobody would care about the IMO anymore, and it would return to being something 17 year old kids do.
I think I was largely right? Have you heard anything about the IMO? Basically no; it's just too easy.
Prediction: the number of major math problems being solved *by labs* next summer will be... zero. It will be too easy for them, and will return to being something math professors, postdocs, and grad students do... (What do you think?)
π¬ 43
π 8
β€οΈ 388
π Retweeted by @paulg
Training tiny models for special purpose use cases works so incredibly well if you have a great self improving recursive flywheel. Shopify ML team is on fire.
finetuned 0.8b model beats GPT 5.6-sol xhigh in this very specialized task.
π¬ 307
π 528
β€οΈ 8.4k
π Retweeted by @paulg
I viscerally enjoy the creative act of programming by hand so much that I will continue to do it manually as long as I live as a joyous act of self-expression. Separately, I plan to make good business decisions, and use AI where appropriate to accelerate my commercial work.
π¬ 23
π 39
β€οΈ 462
π Retweeted by @paulg
What's happening inside Replit right now is going to take the world by storm. We have sales engineers doing things only the most unhinged platform engineers could have done a few months ago. We have designers and PMs making good PRs. There is no stopping us.
π¬ 42
π 32
β€οΈ 573
Legora's revenue has increased by over 9x from this time a year ago. That's an astounding growth rate for a company of this size. It's not unusual for a startup to grow that fast in the first 6 months, but Legora is 3 years in.
@davidneckstein
The momentum continues at @WeAreLegora ...we beat our original Q3 '26 plan a full month early. July was the strongest quarter-opening month in Legora's history, and August was even larger.
Total net new ARR grew 36% MoM.
But the story of the month is in-house teams. ARR from in-house is up 15x+ YoY, average ARR per in-house customer is up 2.5x+ YoY, and in-house teams are now almost half of our total new customers.
Most importantly, you earn ARR by building the best product on the market. Our usage climbs every month: 16+ hours per monthly active user.
In September, we expect to add more ARR than in all of Q1 '26 combined, which would make Q3 the best quarter in our history and capture yet another ARR milestone.
LFG!
π¬ 42
π 57
β€οΈ 433
π Retweeted by @paulg
I did not expect this. Anthropic just published a Lean proof of Fermat's Last Theorem.
The proof is +13M LoC, more than 5 times the size of Mathlib.
Kevin Buzzard's post: xenaproject.wordpress.com/2026/09/04/fltβ¦
The proof: github.com/anthropics/ferβ¦
Anthropic's post: anthropic.com/research/formaβ¦
π¬ 21
π 149
β€οΈ 965