I am working on a company-brain type product (agentwork.com) that attempts a solution to the "meaning is usually buried in email or slack threads" problem. Ingests data and builds a memory, but then also reaches out (when given permission) to team members when the answer is not in the sources. Using this information to build additional memories and pull out important information to get it codified. Would love to get your feedback on it.
A lot of tech people seems to have changed their mind on this. Some ways to reason about that are:
1) They are doing it for opportunistic reasons. They can't afford to be enemies with Trump.
2) They legitimately changed their opinion about a wide array of things they used to believe enough to outspoken about.
3) They believe that while Trump's core beliefs are not aligned to theirs, the alternative is worse. And potentially they believe there is a need for some sorm of over-correction to fix what has happened over the last 4 years.
"We picked the latter, which also gave us our performance metric - percentage of generated comments that the author actually addresses."
This metric would go up if you leave almost no comments. Would it not be better to find a metric that rewards you for generating many comments which are addressed, not just having a high relevance?
You even mention this challenge yourselves: "Sadly, even with all kinds of prompting tricks, we simply could not get the LLM to produce fewer nits without also producing fewer critical comments."
If that was happening, that doesn't sound like it would be reflected in your performance metric.
To find two prime numbers that sum to 123, let’s consider the properties of prime numbers and the nature of their sums:
1. Prime Numbers Greater Than 2 Are Odd: Except for 2, all prime numbers are odd.
2. Sum of Two Odd Numbers Is Even: Adding two odd numbers results in an even number.
3. Sum of an Even and an Odd Number Is Odd: Adding an even number (like 2) and an odd number results in an odd sum.
Given that 123 is an odd number, the only way two primes can sum to 123 is if one of them is 2 (the only even prime) and the other is an odd prime. Let’s test this possibility:
• 2 + x = 123 ⇒ x = 121
However, 121 is not a prime number (since ).
Next, consider the possibility of adding two odd primes:
• Odd Prime + Odd Prime = Even Number
Since adding two odd primes yields an even number, it’s impossible for two odd primes to sum to 123 (an odd number).
Conclusion: There are no two prime numbers that sum to 123.
Answer: There are no two prime numbers whose sum is 123; such primes do not exist.
I found the ratio of rice and water to be the main issue - and the rice cooker didn't fix that. I stumbled upon a recipe with a different approach that I use now.
Take plenty of water and get it to a boil. Add any amount of rice to the boiling water, and let it boil for 8 minutes. Then drain the water, remove the pot from the stove and put the lid on for a few minutes. Reduces the need for measuring quantities.
"I’ve keep taking screenshots of everyone who’s asked about acquiring my account. One interesting pattern: the majority of these requests come from profiles without any photos. I find it so weird that people are so eager to get a username when they don’t even share content!"
If I had an account with a huge amount of followers, then I would also not initially reach out from that main account in order to negotiate the price.
Feels like a lot of commenters here miss the difference between just doing chain-of-thought prompting, and what is happening here, which is learning a good chain of thought strategy using reinforcement learning.
"Through reinforcement learning, o1 learns to hone its chain of thought and refine the strategies it uses."
When looking at the chain of thought (COT) in the examples, you can see that the model employs different COT strategies depending on which problem it is trying to solve.
Back in my early university days I did a short project (https://www.dropbox.com/scl/fi/ch33p2xaq7xavgu9uk0qv/index.p...) on how to generate "hard" mazes. While there are many algorithms to create mazes, there is no real literature (maybe for good reason) on how to create mazes that are hard for humans to solve.
But humans have already trained on an incredible number of games (including reality) when they play No Man's Sky for the first time. What they say here is that training on N-1 games makes you better at the Nth game. So you just continue to scale this up.
Definitely feels like we are advancing towards AGI quite rapidly. As another commenter mentioned (https://news.ycombinator.com/item?id=39693035), the OpenAI DotA game was a big milestone for me.
If you think about it abstractly, humans are basically models that take input from our senses, do some internal processing of that and then take actions with our bodies. SIMA is the same - it takes input from video, and takes action through keyboard actions. There is nothing against introducing additional types of input and taking different actions.
The ability to train on one game and transfer that knowledge to a different game should allow future models like this to train in games, by reading text, watching videos etc, and then transfer all of that knowledge to the real world.
Having a simple clear focus is hard to undervalue. Since doing YC 6 years ago there has never been a second of doubt about the priorities (imo).
YC is there to make money for its investors. And the way to do that is to invest in the best startups and make them huge.
The only source of conflict then comes from YC vs. the startups. For example, do you enforce strict legal structures on all companies. And here the dominant priority is of course YC itself. But this is expected, and given the rationality also easy to work with.