Let me share why I’m bullish on @StepFun_ai too. We’ve got the best homies here! @liulicheng10 and @MikaStars39 Btw, @liulicheng10 was supposed to publish a technical… blog? Still waiting lol. Someone get licheng to post it! Okay, jokes aside, I love it here. Our whole post-training team is one big family.

DeepSeek is finally cooking up their own merch. Previously: DeepSeek pens, and cultural shirts with the DeepSeek logo printed on the reverse side (note: not the back, but the reverse/inside). Now: the Big Fat Fish. (Note: AIGC-generated characters have no copyright, so it’s actually that logo.)

Introducing SceneBound: How Much Should a Streaming 3D Reconstruction Model Remember? SceneBound is a training-free rule that decides what a streaming 3D reconstruction model keeps in memory. Models such as CUT3R, StreamVGGT, STream3R and ZipMap read a video frame by frame and carry the past in a recurrent state, fast weights or a key-value (KV) cache, and every frame written there shifts the coordinates of later cameras or takes GPU memory. SceneBound stores a frame only when the camera reaches a new place: its predicted pose has to lie at least a radius M from every stored keyframe. Memory then grows with the places visited and stays flat when the camera returns. One radius serves four released models and three memory types. Against the best published training-free rules we test: • in one pass, 9% lower error in CUT3R's recurrent state • in one pass, 18% lower error in StreamVGGT's KV cache • after nine passes of a scene, 32% lower error with 11× less writing in the state • after nine passes, 11% lower error with 2.2× less GPU memory in the cache On 11 KITTI drives, its submaps cut the error of one state by 71%. Paper: Delighted to work with Chuyue @artrix_fsc , Chutian, Jingmin @KleinJoe219384 , Hongliang @0xHongliang , Yuheng @eiaton84855 and Xiuyu @sheriyuo. The discussion with Lilo @EpiphanyFall was also very helpful.

I have a fun theory. Just speculation, so don’t take it as actual model info. If you knew how big Sol and Astra were, you could probably work out Bel’s size too. Then you’d see why I don’t buy rumors like “it’s over 10T parameters and as big as GPT-4.5.” You’d also have a better idea of whether OpenAI could actually release Bel.

Yeah TB-Sci and HLE-Diamond are both pretty solid benchmarks. They show where Chinese models still lag. Multimodal stuff is getting more valuable too, and v5 is at least a solid improvement. Really though, O/A has been prepping for these benchmarks for a while lol. This is basically their optimization direction. Chinese models will catch up soon enough.

We are hosting the tutorial Failure Modes in Agentic AI at #EMNLP2026# in room P3 As agents start to challenge human intelligence, the least to do is to know when they fail before we cannot understand how they succeed So sad I can't join in person but I'll present via zoom. Come join us @ZoeyLi20 @HennyJieCC @ManlingLi_ @qiancheng1231 !

My account is finally back. First, a security reminder: please use a unique, randomly generated strong password for every important account. Do not rely on 2FA alone. 2FA can be bypassed, and I strongly recommend using a passkey. Here’s what happened: On Sep 30, I received an email about a new login to my X account. Unfortunately, I didn’t pay enough attention to it at the time. On Oct 1, my account was compromised despite having 2FA enabled. My Gmail account remained under my control the entire time, and I have never entered my credentials on any phishing/scam website. As far as I can determine, the attacker only had my password. How they managed to bypass 2FA and take over the account is still unclear to me. About 10 minutes after the takeover — while I was still on vacation — I received an email from Kimbo Chen telling me that my account appeared to have been hacked. Huge thanks to him for noticing and reaching out so quickly. When I checked, my password had already been changed and a large number of scam links had been sent through my Chats. What made the situation even worse was that the Chat passcode could apparently be reset directly, while I myself could no longer access my own Chats. I immediately opened my first ticket with X Support. The initial response essentially said that because I could still access the account at that point, they would not take further action. Later that same day, I lost all access to the account. On Oct 6, I submitted another hacked-account ticket. The suspension on this account was subsequently lifted, but during this period @GenAI_is_real was targeted by the scam and nearly had their account compromised as well. This whole situation was crazy. Through the new support ticket, I was eventually able to reset the password and regain control of the account. The account is now back to normal. I sincerely apologize for the impact caused by my poor security practices. More than 100 accounts were affected or contacted by the attacker, including @MIT_CSAIL (which has now blocked me), and @GenAI_is_real, who thankfully avoided losing their account. I’m extremely grateful to everyone who warned me, helped report the account, or offered assistance during the past week. Special thanks again to Kimbo Chen, and to my colleague @frankyouchill for taking time out of an already busy schedule to help me deal with this mess. Again: use unique random passwords, and use passkeys whenever possible. Treat 2FA as an additional layer of protection, not something that makes an account impossible to compromise. Everything on this account should now be back to normal. Thank you all for your help, and I hope everyone stays safe and has a great day.