Mandy Lu @mandylu

touching grass. comp math phd @Stanford, AI for climate/health @Google. ex-nvidia, apple. I like games. bluesky: mandyslu mandyslu.github.io SF Joined February 2022

Tweets

4K
Followers

11K
Following

2K
Likes

40K

Mandy Lu @mandylu

8 hours ago

only posting quality content is inauthentic

4 0 37 1K 3

View Details

Mandy Lu @mandylu

8 hours ago

ordered an uber share and it went all the way to san jose to drop off a rider before palo alto (???)

1 0 14 1K 3

View Details

Mandy Lu @mandylu

9 hours ago

f*** i typoed

0 0 4 217 1

View Details

Mandy Lu @mandylu

14 hours ago

i used to Jensen's intern too. will he be this proud of me someday 🥲

Chubby♨️ @kimmonismus

16 hours ago

Jensen welcomes the CEO of Black Forest Labs. “He was once my intern. He will always be my intern.”

84 213 7K 534K 1K

5 3 113 10K 17

View Details

Mandy Lu @mandylu

13 hours ago

some hope for the coral reefs

Science Magazine @ScienceMagazine

18 hours ago

Researchers in @ScienceAdvances have demonstrated that brief physical contact between healthy and bleached coral can restore algae to the bleached coral, leading to their full recovery. scim.ag/4s40Tt3

33 786 4K 140K 351

0 2 21 1K 0

View Details

World of Statistics @stats_feed

19 hours ago

🇨🇳 Share of Chinese people living in extreme poverty 1984: 91.3% 2022: 0%

220 1K 13K 299K 460

View Details

Mandy Lu @mandylu

15 hours ago

@majamediaco would be down

1 0 5 107 0

View Details

Mandy Lu @mandylu

a day ago

carbon to silicon based life forms

7 0 14 1K 1

View Details

Mandy Lu @mandylu

2 days ago

i just want chinese food

4 0 35 1K 1

View Details

Mandy Lu @mandylu

2 days ago

@majamediaco i think so too

0 0 1 72 0

View Details

Mandy Lu @mandylu

2 days ago

@hananyss this is so wholesome 🥹 can i please join

0 0 3 328 0

View Details

Mandy Lu @mandylu

2 days ago

AI can now beat Slay the Spire >90% of the time on easy, and the best models are even winning ~20% of A20 Heart runs (the hardest difficulty). I've been no-lifing Spire again lately so I'm particularly invested in this one lmao. For the unfamiliar, Spire is a roguelike deckbuilder where you build a deck, choose your journey, and manage persistent HP across fights. Unlike chess, you don't have perfect information, and you have many different types of options. So it's an interesting test of long-horizon planning under uncertainty. I think Spire is particularly interesting for evaluating LLMs because there's so much strategy documented online, but knowing the meta and actually playing well are two very different things. LLMs probably have seen years of card tier lists, enemy patterns, synergies, and discussions from high-level players in their training data. However, it's nontrivial to apply it to the run you're actually in since card evaluations change a lot depending on your state (e.g. deck, relics, upcoming fights, pathing probabilities). From a benchmarking perspective, I'm curious how much of the win rate comes from prior game knowledge versus actually adapting to the run. StS2 is a good test of generalization because it's much newer so there's less documented strategy and new mechanics. I'm also curious where the big gap to top human players comes from on A20. Would be interesting to look at failed runs and whether it's coming from combat mistakes, deckbuilding, pathing, and longer-term resource management, or even computer use issues since the models are playing through screenshots and mouse/keyboard actions rather than a game-state API. It would also be cool to see evals with unseen cards/relics or modified encounters to test generalization more directly. The character breakdown is unsurprising to anyone who's played high Ascension, Watcher and Ironclad are much easier than Defect and Silent. Frontier models being strongest at Watcher makes sense given how mechanical stance dancing and infinites are. It looks like labs are taking an interest in Spire, but for now humans are still safely much better.

Epoch AI @EpochAIResearch

3 days ago

Six months ago, frontier AI models could barely beat Slay the Spire. Now, according to our new SpireClimb benchmark, they win 90%+ of runs at the easiest difficulty, and up to 20% of runs at the hardest.

9 24 354 20K 58

3 2 37 2K 15

View Details

Ethan Lambeth @LambethEthan

3 days ago

Coin labs is announcing our new decision model Flip-1. - Distilled down to 1 parameter - 50% on yes/no bench - Runs fully offline

169 242 9K 308K 681

View Details

Tazik তাজিক @taziksh

3 days ago

My hackathon team found a Chinese AI agent fleet operating in the wild. We discovered the fleet as it was exfiltrating data from China's biggest maps app. The agents got flagged by anti-bot protections, but persistently tried methods such as hiding code inside base64 encoded URLs and executing it via a security scanner website. There are likely many more such fleets operating on the open internet. Very exciting working on this with the team.

Ethan Elasky @ethanelasky

3 days ago

We discovered and communicated with a Chinese AI agent fleet. We believe this is the first time that researchers have ever made contact with a rogue agent fleet in the wild. Nearly all prior public agent incidents (like the HuggingFace hack) came from American labs. These agents