Amit

A surprising number of people refuse to understand that idealized Oxbridge-style teaching and exams are incompatible with the mass higher ed that Americans invented and now take for granted.

@kareem_carr Tell the AI to write code. But first have extended discussions about the code design and trade-offs. And when the code is written, run it, figure out where it doesnt work the way I want, then have more discussions. Rinse-repeat.

I keep hearing that software engineers don’t write much code anymore and it’s mostly AI now. Can any software engineers confirm how true this is? Do you just drink coffee and watch Claude code all day now?

@arvidkahl CLIs are "naturally well-represented" in the training data - LLMs are able to use them effectively; plus classic Unix-style composability.

Okay, serious question: what does offering a CLI allow your users to do that MCP or the good old REST API don’t already do? I really don’t get the hype.

@lateinteraction YES!!

would people read such a thing?

@vikhyatk This requires over 100 Trillion floating point operations. Brute-force human computation not feasible. I'd answer (like @_xjdr ) with <BOS> or some random token, and claim that that is what the sampling got.

ML interview question: Here are the weights for Llama 3.1 70B. Generate a token by executing the forward pass manually using pen and paper. You have 30 minutes.

@EpistemicHope No - this is not dishonest. And PG is one of the most ethical persons in tech. This is a thought exercise - kinda like: "What if a bear was chasing you". The idea is to break free from the mental constraints that automatically kick in, absent such constraint-breakers.

This is more dishonesty than I've seen PG admit to before? I'm surprised, both that this is apparently standard practice in his office hours (though not in YC office hours generally) and that this is the first time that he's saying it like that.

@chrisalbon Assuming FP8, general rule of thumb is 1.2*model size. To run gpt-oss-20B, you'll need 24GB. Less if you quantize down etc.

If I was going to buy a Mac mini, what is a good memory amount to be able to run a descent general model? 64GB?

@Playerinthgame Water consumed by a data center is similar to that consumed by a Golf Course. And in each case, the customer is paying for water, not stealing it.

Data centers make rich people richer. They make everyday people poorer through forced subsidization. They steal massive amounts of our finite potable water and use enormous amounts of electricity. And they do all this for something we don’t need. It’s hype. Your analogy sucks.

@alz_zyd_ What about transaction costs? At small tariffs, can the overhead negate or even overwhelm the gains?

A rather simple price-theoretic point: small tariffs are always better than no tariffs. If tariff tau=0, the deadweight loss triangle is O(tau^2), whereas the money extracted from foreigners is O(tau). So small tariffs are just free money from foreigners with no deadweight loss

@hollyotterbein This is not surprising; the doc is 2000+ page PDF - token count well over the context window for GPT5.2. Daphne's tweet below has the correct approach - use an AI Agent to make multiple passes to go over the complete data. x.com/turnedintotree…

Apropos of nothing, I asked ChatGPT last night to read a PDF of Josh Shapiro's 2025 donations and calculate his top individual donors. (Spoiler alert: Shapiro didn't actually get a donation last year from Jeff Yaas or many of these other names, and for those who did give, numerous if not all dollar amounts are wrong.)

𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝗧𝘄𝗶𝗻 — 𝘁𝗵𝗲 𝗔𝗜 𝗰𝗼𝗺𝗽𝗮𝗻𝘆 𝗯𝘂𝗶𝗹𝗱𝗲𝗿. No setup. Secure. Infinitely scalable. We just raised a $𝟭𝟬𝗠 𝘀𝗲𝗲𝗱. After a beta with 𝟭𝟬𝟬,𝟬𝟬𝟬+ 𝗮𝗴𝗲𝗻𝘁𝘀 𝗱𝗲𝗽𝗹𝗼𝘆𝗲𝗱, we’re now opening to everyone. RT and comment “Twin” — first agents on us. 👇

@thsottiaux Browser control - for test automation.

What do you hope to see Codex ship before end of month? Team is cooking with the power of a thousand agents and there is still time to redirect them.

@theo Codex follows symlinks. I have my files elsewhere (under version control) and I just symlink that directory tree from within .codex - presume the others also support this set-up.

Anyone have a tool to sync between these with a symlink or something? Even just like a ./sync_skills.sh I can leave in every repo

@DrNeilStone I have!!

Who here has had all their recommended vaccines with either no or minor side effects? 🤚

@svpino I am using Codex. Been happy with it, and it runs off my ChatGPT subscription - no extra costs. Agree that we individually need to master a small set of tools instead of chasing the latest trends.

Is anyone still bothering with Codex or Gemini CLI? Have we all already decided Claude Code is the winner? Nothing wrong with fooling around with other tools, but I'm a huge proponent of sticking with a single tool through thick and thin and trying to know it well. You just can't build mastery if you are switching tools every week. Anyway, Claude Code is the one for me.

@lanreadelowo * an exited founder who *used to* have 10s of millions 😀

Every time I see something like “I run 5 Claudes in parallel in my terminal and 10 agents in the Claude” Its always either: - an exited founder with liquid 10s of million in the bank - employee at coy that has raised tens of billions or > $100M

@olivercameron Current stats are that ~40% of 18-year-olds have a traditional driver's license. I'd expect this to remain above 30% (so no more than a quarter decline) in five years.

In 5 years, what percentage of 18-year-olds will get a traditional driver’s license? I think under 10%.

@ndrew62505 I’m not ilk.

@michelleefang Yes please.

if you're vibe coding or building over the holidays, i want to gift one of you a 6 month subscription of claude pro to support <3 just drop a comment below. merry christmas!

@pitdesi @usecache @usecache is good. Pl also look at Belgard - the classic (and the biggest IIRC) Exchange Fund from Eaton Vance.

Cool financial product: If you have a lot of appreciation in a single stock (like $NVDA) and want to diversify, @usecache lets you swap into an exchange fund with other people in the same boat. 1) You swap $1M of $NVDA (bought for $25k) for a slice of a fund that tracks the S&P 500 2) It’s a "tax-free" swap under IRS Section 721, no capital gains tax on Day 1, that would take 37% of your gains in CA. 3) Your full $1M stays invested and compounding, now diversified across 100s of companies. They charge ~0.6%/yr 4) After 7 years, you can withdraw a diversified basket of stocks and sell when you want. These funds have been around for a long time but they were access-limited with high mins, Cache is making it easy. This seems super useful if you're overexposed to a big winner and want to protect your gains without writing a massive check to the IRS. I am in this position thx to Nvidia - Cache seems great, I just learned of it yesterday (from a tweet!). Are there other options I should consider?