Xiaopu Peng @ ICML

31 posts

Xiaopu Peng @ ICML banner
Xiaopu Peng @ ICML

Xiaopu Peng @ ICML

@SuperNyx1024

Help build @Zai_org

Singapore Katılım Ekim 2025
145 Takip Edilen35 Takipçiler
Ivan Leo
Ivan Leo@ivanleomk·
my life now is just fire off prompt fire off another prompt voice dictate its pretty good
English
2
1
23
3K
Sarah Sachs
Sarah Sachs@sarahmsachs·
The other day @colebemis asked me if I thought Notion was a cult. My answer was I’m so deep I have no way to answer that as a reliable narrator. But that’s kind of the point. To work here is to believe it. Which means every conversation and debate is high trust, because we all have the same red-stained teeth.
Sarah Sachs tweet media
Camille Ricketts@camillericketts

In our experience, Kool-Aid flowing is one of the best indicators of real magic at a company. For all the folks picking where to go next out there, here’s what we look for:     •    Evidence they’re winning. Revenue definitely. But also early headliner customers (Turbopuffer, Baseten, Stripe are all great at highlighting their’s), known talent on the early team (Pierre Computer, OpenAI, Anthropic, Thinking Machines all have notably stacked early crews), and (especially if pre-revenue) accolades that industry leaders pay attention to (Physical Intelligence winning the Robot Olympics).     •    A mission that matters to you. You’ll feel pride in your bones making this thing happen.     •    Zaggy vision for the world. The best companies to join are non-consensus and right. Industry experts laughed Etched out of the room. Anthropic was seen as too soft and safety-focused. Do you believe in the future a company is building?     •    Founder aura. Do you feel the pull to impress them and be in their midst? That counts for a lot… because you will have to be.     •    Internal stage for excellent work. Does this company have a way they regularly elevate folks inside for doing interesting things, making things, doing deals? How high is the bar for what they feature at All Hands? What earns a Slack shout out from the CEO? Will you have constant visibility to learn from your highest performing teammates - and maybe become one, too?     •    External platforms encouraged. Do you see people at the company tweeting about their work? On stages at conferences? Hosting their own events? Writing blog posts? Think @jeff_weinstein at Stripe. @ryolu_ at Cursor. @sarahmsachs at Notion. @oneill_c at Baseten. It’s a signal the company invests in people and their ideas.     •    Rituals. Raises eyebrows on the outside, beloved on the inside. The best kind of cult. Vercel wears all black. Hundreds of cursors swarm Figma All-Hands slides as they’re being presented. Stripes used to run up Bernal Hill and played Werewolf in the evenings. Sub-20 Notion ate heaping piles of pasta every Friday at the same restaurant (and yet never needed a nap). Memories are sticky.     •    Strong merch game. Unusual swag people actually want to wear - a sign of deeper care and creativity. Our recent favorite: Gumloop has people bring in their favorite they’d wear everyday if they could and get it embroidered on the spot.     •    Culture heads indulged. Those folks who want to make the place awesome are told to “go for it!” Ben Barry churning out posters at early Facebook (that we can all still recite). Back in the day Dropbox’s famed Blackops squad stocked in-office karaoke studios, made an employee holiday catalog full of odd offerings. Good energy begets good energy.     •    Other people are telling the story. The most honest reviews you’ll get of a place are outside-in, but they come in all forms. Andrej Karpathy tweeting. 100 community members sharing what they built. Customers name dropping the company on stage. The timeline’s inner circle commenting on product launches. See: who responded when Midjourney Medical went live. Join a company that makes friends. Its network will be yours.

English
5
1
44
6.8K
Su Park
Su Park@sunotsue_·
back home in seoul city for #ICML 🇰🇷 I’m Su, based in SF, making TerminalBench tasks by day. let’s grab coffee if you’ve been thinking deeply about data efficiency, data attribution, or simulation lately - I’ll take you to the cutest ones (I grew up here!) here till monday!
Su Park tweet media
Seoul, Republic of Korea 🇰🇷 English
1
0
17
1.1K
Xiaopu Peng @ ICML
Xiaopu Peng @ ICML@SuperNyx1024·
@ValsAI It looked like a great event! I got to meet Rayan at the RL + Agents lunch during ICML, really enjoyed the conversations!😆
English
1
0
1
142
Vals AI
Vals AI@ValsAI·
📢Vals is at ICML! Last night we hosted a small dinner and it was one of the highlights of the trip. Grateful for everybody who attended and already looking forward to more conversations. If you're at ICML, DM us to say hi!
Vals AI tweet mediaVals AI tweet mediaVals AI tweet media
English
3
0
41
5K
Xiaopu Peng @ ICML
Xiaopu Peng @ ICML@SuperNyx1024·
@rishi_desai2 Just saw this and realized I missed it, looks like it was great! Hope there’ll be a next one I can join.
English
1
0
1
30
Sasha Krassovsky
Sasha Krassovsky@bztree·
I'm at ICML this week if anyone wants to meet!
English
5
0
33
3.6K
Xiaopu Peng @ ICML
Xiaopu Peng @ ICML@SuperNyx1024·
Spent this week at @icmlconf talking to the people actually building benchmarks. @OpenAI retracting SWE-Bench Pro (~30% broken) just made months of private consensus public. Saturation and contamination feed each other. The more trusted a benchmark, the faster it leaks, and broken tasks finish it off. Fixed-answer evals were always going to die this way. @ArtificialAnlysis already dropped SWE-Bench Pro from its Coding Agent Index. The way forward is the paradigm @Datacurve is pushing with DeepSWE: contamination-free, from-scratch, long-horizon. What benchmark are you actually betting on right now?
OpenAI@OpenAI

We audited SWE-Bench Pro, one of the most widely used AI coding benchmarks, and found it no longer reliably measures frontier coding capability. We find 30% of SWE-Bench Pro tasks to be broken, and are retracting our previous recommendation that the research community use it as a leading coding eval. openai.com/index/separati…

English
0
0
1
278
Xiaopu Peng @ ICML
Xiaopu Peng @ ICML@SuperNyx1024·
@OpenAI Fixed-answer benchmarks don't just saturate. Satuation and contamination feed each other. The more trusted a benchmark, the faster it leaks. Broken tasks just finish it off.
English
0
0
0
1.6K
OpenAI
OpenAI@OpenAI·
We audited SWE-Bench Pro, one of the most widely used AI coding benchmarks, and found it no longer reliably measures frontier coding capability. We find 30% of SWE-Bench Pro tasks to be broken, and are retracting our previous recommendation that the research community use it as a leading coding eval. openai.com/index/separati…
English
317
545
7.8K
2.2M
Ben Rank
Ben Rank@full__rank·
Tomorrow (Thu) we will present PostTrainBench at ICML From 10:30-12:15 in Hall A #2007 (poster) Drop by to learn more about how to build evals for automated AI research and upcoming changes to PostTrainBench @maksym_andr
English
1
3
26
6K
Xiaopu Peng @ ICML
Xiaopu Peng @ ICML@SuperNyx1024·
@iaindunning joked that he was "boring" at @icmlconf. He was anything but.🤣 His idea that stuck with me: Market events could be like tokens for a new AI modality. Just as LLMs learn from tokens, Large Market Models could learn from orders, trades, and market interactions. #ICML
English
0
0
2
77
Xiaopu Peng @ ICML
Xiaopu Peng @ ICML@SuperNyx1024·
@ivanleomk Hey Ivan, I've been exploring the similar topic recently too, but sadly won't be in nyc!
English
0
0
0
25
Ivan Leo
Ivan Leo@ivanleomk·
going to nyc end of the month, would love to meet anyone working in finance, evals and agents :)
English
4
0
36
3.5K
Chubby♨️
Chubby♨️@kimmonismus·
Huge: China’s MiniMax Plans to Launch 2.7-Trillion Parameter Model (MiniMax Pro) tl;dr MiniMax is preparing a 2.7T parameter open-source model, potentially launching as early as Q3. That would make it far larger than any Chinese model currently on the market, and over 6x bigger than MiniMax’s current M3 model. China’s open-source AI wave is not slowing down. Looks like they just started. Via The Information
Chubby♨️ tweet media
English
81
105
1.2K
96.9K
Yulun Du
Yulun Du@Yulun_Du·
Curious why Hy3 didn’t benchmark against Kimi K2.6, especially since the Hy3 preview models did compare against our Kimi models. That said, congrats to the HY team — strong work. (Though I have to admit, from the benchmark table, this feels at least as much like a GLM-5.2 advertisement as an Hy3 release. 🧐🧐😂
Yulun Du tweet media
elie@eliebakouch

new open (apache 2.0!) model from @TencentHunyuan, Hy3 is only 295B total 21B active and competitive with MUCH bigger model on benchmarks huggingface.co/tencent/Hy3

English
28
13
291
52.3K
Xiaopu Peng @ ICML
Xiaopu Peng @ ICML@SuperNyx1024·
Congrats to the @TencentHunyuan team on the Hy3 release. One clear takeaway from @aiDotEngineer last week: the conversation is shifting from benchmark scores to real-world workflows. Great to see more teams moving in that direction.
English
0
0
0
181
Lee Robinson
Lee Robinson@leerob·
Personal update, I'm starting a new role at Cursor! I'm moving into ML, working on training Composer. I'll be researching how to improve model behavior and personality.
English
439
106
7.6K
318.1K
Xiaopu Peng @ ICML
Xiaopu Peng @ ICML@SuperNyx1024·
Great to see GLM-5.2 now avaliable in Cursor! I've been using Cursor since early 2023 and have always felt that GLM-5.2 and Cursor would be a natural match for coding. Strong coding performance meets one of the best builder environments. Excited to see what people build with it.
Lee Robinson@leerob

You can now try GLM 5.2 in Cursor! Excited to see more useful open models, thank you to Fireworks for partnering here. Results from our evals ↓

English
0
0
1
181
Lee Robinson
Lee Robinson@leerob·
You can now try GLM 5.2 in Cursor! Excited to see more useful open models, thank you to Fireworks for partnering here. Results from our evals ↓
Lee Robinson tweet media
English
166
150
2.9K
310.3K