Software developer! Working on the next great Metroidvania https://nitter.cf/t.co/0tWjtKXQFc. Helped build DWS https://nitter.cf/t.co/Z1Qfa7B3kV.
Gainesville, FL
Joined March 2009
- Tweets2.6K
- Following1.4K
- Followers494
- Likes11.6K
Hey @OpenAI can I borrow some compute so I can knock out one of the Millenium Problems?
35
I have been saying for years the moment to watch is for when computers can do proofs on unsolved math problems. Difficult open math problems are some of the most difficult problems to solve because of their abstraction and relatively little feedback you get when working on such a problem. OpenAI dropped a crazy amount of proposed solutions to many difficult open math problems github.com/openai/math. This is as big of a moment as walking on the moon. Most people don't even know this happened.
There is a lot of fear and concern about this. The utility of mathematicians might go to zero. There could be a time where new proofs by humans goes to zero. The fear is that AI might one day look at us the same way we look at monkeys. I'm optimistic about our future with AI but do have concerns when there is no more need for "human in the loop".
1
49
Kyle Nunery retweeted
After TEN LONG YEARS, Order of the Sinking Star is out Now! On Switch 2, PS5, and ... it's at the top of Popular New Releases on Steam, already!
Thank you everyone for your support during this long, long, long, long, long development!
Kyle Nunery retweeted
I think I’ve decided not to blog about this because the benchmark client they use to tweet results is so bad and the implementations have nothing in common and I’m convinced the Rails code from the Rails founder can be much improved cuz of bad caching patterns.
I’m done
🤖 Made with AI
Kyle Nunery retweeted
The Campfire ports and benchmarks are invalid. Nothing published is of credibility.
Yes we will be reading code less or none in the future. But that’s not today.
The Campfire ports proved absolutely nothing except one individual got bamboozled by AI’s benchmarks.
Kyle Nunery retweeted
Architecture Is Now an Economic Decision: How High‑Quality Code Reduces AI Token Costs 10x
confidenceengineering.com/bl…
So where is AI coding at this point in time:
1. Beyond super human at generating code within a "consistent" codebase.
2. Beyond super human at generating sql. It can create highly optimized sql and it understands query plans better than any human.
3. Beyond super human at converting code from one dialect to another.
4. Beyond super human at debugging and finding root causes. For example, if you have DataDog and the source available it can find bugs that would of been hours to days in 1-5 minutes.
5. Beyond super human at writing tests of any kind. Unit, integration, e2e, and acceptance tests are trivial for the latest models.
6. Beyond super human at creating human understandable documents for us to review. Creating a Claude artifact to summarize a working feature is only minutes which would of taken hours to days before.
7. Beyond super human on working at multiple tasks at a time.
8. Beyond super human at debugging issues in your browser in real time.
Where you still have to watch it:
1. The smallest missing context and it can go in the wrong direction.
2. Overly complex solutions because it doesn't have the full context or because it's default approach is to build Netflix scale projects.
3. Overly confident language which can be misleading if you aren't an expert on said topic. If you push back it will often self correct.
4. Dead code or leftovers from a migration can confuse it.
5. Propagates bad patterns in your codebase unless you tell it to avoid them.
1
57
I like to do predictions every New Year. My only prediction was that this was going to be the weirdest year ever. Definitely proving to be true.
39
If you could one shot a 90 minute AI movie it would cost around 6k. That's very expensive and the true cost would be 60k-100k. We don't have to worry about AI movies taking over yet.
1
41
Kyle Nunery retweeted
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
I have a polynomial time solution for the 2d Euclidean traveling salesman problem. Mentioning it now before Anthropic publishes a solution.
1
40
Kyle Nunery retweeted
Si se que cuando programas con Go usando IA el resultado no es perfecto, ves patrones antiguos y errores básicos, JetBrains ha creado Modern Go Guidelines para que los agentes generen Go moderno.
github.com/JetBrains/go-mode…
Muy recomendado para programadores de Go.
AI's solution to 87-year-old riddle takes mathematicians by surprise | New Scientist newscientist.com/article/258…
110
Kyle Nunery retweeted
Same plan everyday. This week we had a solid 5 greens days. Yes, we had a few people make some mistakes but over 99% of my Discord was nice and green this week.
