Yeah it is so absurd. The going cost of a politician in local California politics is on the order of $30-60k a year. Very inexpensive when they vote on policies shaping billions of dollars worth of sales.
Clerky is great, Stripe is really buying all of the great product companies. I feel like with Clerky and Stripe Atlas they, Stripe now controls all of the early incorporation infrastructure now. This feels deeply problematic.
Fable is insanely annoying. It treats me like an incompetent, suggesting when something is in my expertise that I'm actually wrong. And then it will simply refuse to engage or do what's asked. Anything related to present day American Politics is strictly off limits. How is this truth seeking, or rational? Model misalignment increases when post-training is done on non verifiably domains. Scary times indeed.
Yeah, this is correct. There are so many large multi-trillion dollar companies coming to IPO, which if your are passive index holder and you are trying to track the market it is correct for these companies to be included. And besides SPY has chosen not to fast track where QQQ has. It is a free market, and folks are free to NOT buy QQQ. So I'm not sure why this is a point of debate.
He created Django, what do you mean he's not an engineer? Also 'low-effort??' his posts are extremely in-depth, clearly very thought through with a significant amount of time and energy. Additionally he does perform multifaceted checks across LLMs in many of his other blog posts.
It is more than just books published pre-2022, but I've noticed that it is necessary to buy books __printed__ before 2022. Recently I've noticed aggressive misspellings and misquotations in old books printed in '25/'26. I believe these errors are being introduced because the publisher is using an LLM or some similar tool to correct any prior errors but is introducing more hallucinations. Or perhaps the cause is people caring less than they used to?
Yeah I know some of my team members have invested a lot of time in this. Could definitely be worth chatting with them on what improvements could be made here. We're starting to deprioritize our consumer facing agent harness, in favor of more infrastructure level improvements we are making.
Interesting. We definitely have long-running agents where certain preferences are key. However, some of the preferences are likely going to be shared universally across our customers. Is there some way of triaging this feedback into permanent improvements in agent performance?
I'm curious who the ideal customer of this should be. If we're a startup with our own harness, are we a good fit? What would qualify us or disqualify us from being a good user?
This is like saying - 'I don't understand why anybody would ever buy a pizza from a store. All you have to do is get some flour, water, tomato sauce, cheese, mix the flour and water together, whip the dough, add tomato sauce, put cheese on top of it, cook it for 20 minutes, and then serve it.'
So __much__ value is in the fact things are easy. Money is __not__ the most valuable thing in the world.
Sorry, I think you mean abiding*. But laws are not some moral edicts handed down by god. They can and often are wrong or seriously misguided. Laws can and should be broken if and only if the agent at hand has a thorough understanding of why they are violating the law. Breaking a law and antisocial behavior are not necessarily equivalent.
Holy shit this is horrible. It really shows the true cost of having a disciplined public society. People love to hate on SF, and the homelessness. But I think it’s a society that prioritizes individual freedom which allows for both this outcome and the entrepreneurial environment we see.
One of my mentors created Blackboard. It used to be very very good, but he sold it to private equity, and they immediately fired all of the customer support and developers, 3xd prices overnight leading to the 'blackboard sucks' problem. This gave the opening for Canvas to eventually come on to the scene and dominate.
I asked claude to decipher this and it refused. I asked gemini and it was permissible. Very interesting to see ROT-13 banned via Anthropic as a 'prompt injection' risk.
Anecdotes __are__ data. How much weight you ascribe to it as being representative is different. But you cannot disqualify it as 'not data'. It is usually a leading indicator of what could potentially show up in these more robust datasets.
No - because the AI will be super human. No human even at $1mm a year would be competitive with a $100k/yr corresponding AI subscription.
See people get confused. They think you can charge __less__ for software because it's automation. The truth is you can charge MORE, because it's high quality and consistent, once the output is good. Software is worth MORE than a corresponding human, not less.
Why is this being hidden off of the main HN pages? There are clearly enough points for it to be significantly up weighted. I don’t understand the censorship.
I think this is an open question still and very interesting. Ilya discussed this on the Dwarkesh podcast. But the capabilities of LLMs is clearly exponential and perhaps super exponential. We went from something that could string together incoherent text in 2022 to general models helping people like Terrance Tao and Scott Aaronson write new research papers. LLMs also beat IMO and the ICPC. We have entered the John Henry era for intellectual tasks...