Hacker Newsnew | past | comments | ask | show | jobs | submit | RA_Fisher's commentslogin

Yes, you're referring to transfer learning. It's an empirical question that also depends on the nature of activities.

Oof. It’s so sad when businesses with in-demand products suffer from a lack of successful economic mechanism design.


They had a very successful economic mechanism design for many years. It's one that thousands of businesses have relied on for decades. Create great free product, create rich tutorials and docs, people look them up and you can sell variety of services from there.

Then ai companies stole the data, put a $20 pricetag on it and wiped out all those businesses essentially overnight.


If those "variety of services" are worth anything they should remain popular. There are countless ways to advertise them to people. I'm guessing many devs who were regularly browsing their documentation had ad blockers anyway.


Typically here's how it worked:

Company creates a free core product, possibly open source, etc.

Company creates rich help docs and other useful resources.

Customers who use their core free product are frequently engaged with the company and it's helpful resources.

Said company offers additional services such as coursework, cloud services, analytics, etc.

Since customer knows and trusts the core free product, and knows the company is trustworthy, they may be inclined to purchase something.

Now, few coders use these learning resources (ai fetches it for them), they do not know anything about the company and they never engage with the company to see their additional offerings.

Therefore the company now has to advertise by spending heavily on Google/meta/etc. to a cold audience that doesn't even know who they are, because everyone using their system/sdk/library/etc. doesn't know anything about them because ai just steals all their resources and regurgitates it to the coder.


> Now, few coders use these learning resources (ai fetches it for them)

This isn’t the problem. The problem is that you don’t need to pay for (or care about) their services when an LLM can easily replicate them.


Europeans use American iPhones and Android bc they wouldn’t add a backdoor.

Didn’t the UK want a backdoor added?


Anybody who feels that their communications are secure on their mobile devices is deluded. There are countless access methods at many levels, even in the US.

There are laws in every country (including the US) to ensure that communications can be monitored when such monitoring is "authorized."


Yes indeed, I call it the revenge of statistics! Pseudoreplication rears its ugly head once again. Nonparametrics are certainly useful, but once you know things about the DGP, it’s often possible to do something “traditional” that’s better. Probabilistic programming, for life!


AI be used to make reviewing it easy, eh?


Exactly, it’s unrepresentative of AI. It’s damaged AI.


All AI is flawed and prone to "hallucinations"(doing exactly what it was designed to do) that's why Microsoft considers it an "Entertainment" product.


I prefer to say AIs are prone to misremember things, as well as humans do. The more you read and learn, the more material you have to get confused, unfortunately.

We might want to work around the certainty with which it misremembers things. I have a very large interval of "I have no idea", or "I don't remember" and rarely caught myself misremembering things (my children are far better at that). I assume my grandchildren will eventually improve their parents' scores by a wide margin.


Everything is flawed. But if you're testing the effect of "AI advice" you need to use an AI that has a normal failure rate without tailoring the questions to alter that rate, and/or compare AI versus other sources of information that have the same failure rate.


Proofing open math conjectures


....Dangerously suggestive post, when you show you don't have a proper idea of "AI" or won't use the term 'AI' properly.


Yes, definitely a case of bad-think.

Certainly not double-plus-good like AI.

Dangerous.


It seems perfectly representative of AI and AI users.


Why didn't they use a model people actually use?


Because they would have had to dig a little more to find trivia it gets wrong. They don't care about "which AI is the best for little facts about movies," they care about "what do people do when the AI gives them a response".


But surely people's willingness to listen to an AI is contingent on their past experience with this model


These people were not told what model answered the questions.


We need a new Pelican, bc earlier models were disadvantaged relative to later models. It's not a very useful comparison anymore.


I wrote about what it's useful for (not much these days) yesterday: https://simonwillison.net/2026/Jul/16/kimi-k3/#what-can-we-l...


It’s well-known Brockman and his wife donated $25 million to MAGA Inc., the main super PAC supporting Donald Trump.

The main higher-level factor is our patriarchal culture (and more bad things tend to stem from places where that’s intense).


Not sure what this has to do with Sam being hired at YC 10 years ago


Fair. An interesting question: how quickly can we detect something without being thwarted by anisotropy / the multiplicity of backward paths? ie- retrodiction

Let’s organize the temporal order a bit. This is what some research turned up.

“Groups of senior employees, concerned with Altman’s leadership and lack of transparency, asked Loopt’s board on two occasions to fire him as C.E.O., according to Hagey.”

“As Mark Jacobstein, an older Loopt employee who was asked by investors to act as Altman’s “babysitter,” later told Keach Hagey, for “The Optimist,” a biography of Altman, “There’s a blurring between ‘I think I can maybe accomplish this thing’ and ‘I have already accomplished this thing’ that in its most toxic form leads to Theranos,” Elizabeth Holmes’s fraudulent startup.”

https://ghostarchive.org/archive/veLhW

I read he’s a vegetarian out of concern for animals, and that’s a good moral sign (but perhaps less relevant here).

Are / were there precursor concerns arising to signal back then? I don’t know.


I'm quite confident that the error was:

"Despite potential warning signs, we believe Sam to be an ethical person"

and not

"We know that Sam is an unethical person, but that's either not a problem, or an actual asset when it comes to running our venture fund"

The former is definitely still an error, but it's quite different from the accusation being leveled on this thread.


I agree. I don't think there's much public information out there about the due diligence involved in his hiring.


This is super, but students will have access to AI during the test in real life, so it's ironically less realistic to remove it (thinking of the "... GPT-4 actually harmed subsequent performance by 17% when the tool was removed ..." part).

I'm more curious how students perform on the test with vs. without AI.


So Opus might be correct?


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: