亚历山大·王(Alexandr Wang)在谈到“如何拉开与 99% 人的差距”时,给出了极其朴素甚至残酷的答案:
顶层竞争根本不拼智商,拼的是极端执行力的非对称倾斜。
他的三条核心准则:
1. 绝对聚焦,压榨认知带宽: 绝大多数人的忙碌只是“低密度伪勤奋”。真正的专注是把全部精力砸在唯一的核心支点上,拒绝任何无关干扰。
2. 把“过度交付”做成护城河: 行业标准要求走 1 英里,你就走完 10 英里。当所有人都按合同完成 100% 时,常态化交付 200% 的人会直接卷走全部声誉与溢价。
3. 带靶向的超额输出,拒绝无目的内耗: 努力不是盲目自我感动,而是针对具体目标的持续过载测试。把每一次交付都当成生死决战来打。
商业世界最公平的一点在于:
聪明人到处都是,但能忍受极度枯燥并长期超额交付的人凤毛麟角。
执行力溢出,就是最暴力的护城河。
code_z retweeted
几条教训
1. 不要依赖 agent 读论文或者高质量博文,
后果:思维深度大幅度下降,灵感触发时机大幅度减少,浪费大量的时间去理解 agent 的垃圾表达。
2. 不要把搞科研/技术学习 和 英语学习混在一起。
后果:效率低下,推进缓慢,失去耐心后放弃,灵感和思考都需要时间积累,混合学英语这个目标会大幅度延迟你的思考,让你的脑子在来个任务上来回切换
3. 不要因为 token 用不完就想着 background 跑点东西,把 token 用掉,不影响你的主线
后果:除非你编了稳定的 workflow,否则即时奖励更强的 vibe coding 一定会吸引你的注意力,打断你的深度思考。
总结,一定要感受自己在 agent 时代下的认知水平的提高速度,到底是加快了,还是不变,还是退步了。如若不然会沉迷于 shallow work,最后认知退化,导致阅读长文字、深度思考都没耐心,人就废了。
code_z retweeted
What if embodied agents could learn to plan instead of just learning how to act? 🧠🤖
Introducing Generative Trajectory Planning (GTP)! 🚀
Instead of learning policies through RL, GTP imagines possible futures, plans ahead, and learns to plan better through experience.
code_z retweeted
In language models, turning a generative model into a critic is simple: pass in the output and prompt for a score. Robot policies can't do this - they output physical actions & there's no built-in vocabulary to predict scores.
So when it comes to RL on robot policies, we typically train a separate critic from sparse rewards - blind to everything the policy learned from the actions in the pretraining data.
We think this has it backwards - can a critic learn from those actions the way the policy does? (1/n)
code_z retweeted
Replicated the laundry task in mujoco web, and also self-trained the RL WBC. By Opus 5.5 in 3 hrs.
Agentic real2sim is nearly solved.
Inspired by the discussion between @qiayuanliao and @wenlong_huang, @kun_h____'s explanation, and insights from Pengxiang and Junfeng. I admire @DynaRobotics' thoughtful choices and look forward to what's next.
code_z retweeted
Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions.
CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks.
With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks.
We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed.
Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size.
📄 Blog: contrastive-lm.notion.site
💻 Code: github.com/Contrastive-LM/CL…
🗣️ Discord: discord.gg/5dAQEDJBs
🤗 Data & Models: huggingface.co/Contrastive-L…
More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵
code_z retweeted
What if RL didn’t choose one action at a time?
Reinforcement Learning with Action Chunking
Q-Chunking: run RL directly over short sequences of actions.
Instead of:
aₜ → aₜ₊₁ → aₜ₊₂
the agent chooses:
[aₜ, aₜ₊₁, …, aₜ₊ₖ]
This simple change turns out to be powerful for long-horizon robotic tasks.
code_z retweeted
🤖 One brain, for any robot. But how does a generalist robot achieve mastery on a specific task?
🚀 Introducing CLIFT (Closed-Loop Iterative Fine-Tuning), a framework that turns Gemini Robotics On-Device into a humanoid specialist using only a managed fine-tuning API. [1/6]
🌐 thomaschen98.github.io/clift…
📄 arxiv.org/abs/2607.29172
code_z retweeted
📚 @saturdayrobotic Robotics & World Models Reading Club 18 Recap: keynote @Guanming717 (@gen_instinct), @ffeng01 (@UCSanDiego, @AetherLab_AI).
Scaling video generation alone will never produce reliable embodied intelligence. The next generation of world models will be defined by causality, latent actions, verification and task alignment — not bigger diffusion models.
DreamZero, ImageWAM, FastWAM, LeWorldModel, TC-WM, WAV and Unified Latent Action Models all point in the same direction.
DreamZero jointly models video and actions instead of treating control as an afterthought. Frames are VAE-encoded into latents that enter Causal DiT blocks together with action noise and proprioception+language. Training uses joint video-action flow matching under teacher forcing. Inference keeps a KV cache, samples action chunks autoregressively from real observations, executes them asynchronously in closed loop, and decodes future frames only on demand. The joint prediction is deliberately factorized into a video term plus an inverse-dynamics term.
ImageWAM turns image-editing foundation models into world-action models. A frozen LLM encodes the language instruction; the current observation is VAE-encoded with noise and fed into a tunable image-editing backbone. A lightweight tunable Action Expert then reads the edited future observation and outputs the full action sequence, leaving the powerful visual priors frozen.
FastWAM asks a blunt efficiency question: does action prediction really need to attend to future video? Training still mixes three masking regimes — joint video-action denoising, video denoising plus inverse dynamics, and action conditioned only on current observations. At inference the winning strategy predicts actions without ever attending to future video tokens, treating video prediction merely as an auxiliary training signal. Joint training still helps representation learning; decoupling at inference delivers much faster real-time control.
LeWorldModel learns planning-oriented latent dynamics instead of pixels. An encoder maps observations to latents; a predictor rolls those latents forward for many steps under actions; a cost module compares the trajectory against a goal latent. An auxiliary regularizer projects the latents onto random univariate directions and forces the distribution toward normality, producing a compact space directly usable for model-based control.
“Is latent all you need?” The answer is nuanced. Asymmetric denoising, sparse future imagination and dense action refinement let the model dream only when necessary. Heatmaps show predicted action hotspots tightly aligned with objects and robot end-effectors, proving that latent representations can focus computation on causally relevant regions while ignoring irrelevant background.
Classic video generators fail three basic tests: precise action control, object consistency and physical consistency. Choppy knife-cutting sequences and physically implausible 3D navigation are not edge cases — they are symptoms of missing causal structure.
Three interlocking questions therefore dominate:
Can we recover the hidden state factors behind observations?
Can we recover the latent actions that actually drive system dynamics?
How do we use those representations to build self-improving systems?
Hu & Shum (2013) give a concrete answer to the first: under mild assumptions a short temporal block of observed trajectories is already sufficient to recover the latent context up to an invertible transformation. When clean latent factors are absent, stochastic residuals still matter and the model becomes a pseudo-Bayesian filter. Empirically this approach ranks at or near the top on Kitchen (~70 %), Maze2D (~160 %), Walker (~118 %), LIBERO-object (~93 %) and LIBERO-long (~62 %), beating DD, DF, LDCQ, Diffuser and DP.
Latent-factor identification proceeds by feeding raw trajectories through a sequential encoder to obtain latents, then a sequential decoder that reconstructs the trajectories. These latents are subsequently used by Ada-Diffuser-Planning and Ada-Diffuser-Policy modules that inherit the same causal inductive bias (diffusion I/O, masks, inverse dynamics).
Task-centric world models go further. Five architectural families are contrasted; the winning TC-WM injects a task signal that aligns and splits latents during training. History is per-patch encoded and aligned with the current embedding; actions condition a latent-dynamics transformer (positional embedding + transformer + per-patch decode) that predicts future embeddings and proprioception. A trainable encoder-decoder sits on top of frozen vision foundation models. Planning uses either cross-entropy method elite selection or latent diffusion guided by inverse dynamics. The same models outperform TD-MPC2, DreamerV3, MuZero and DINO-WM on CEM (Maze 100 %, Wall 100 %, Push-T 92 %, Cheetah 292) and LDP (Lift 60+ %, Can 62 %, Square 40 %, Hopper 46) while producing markedly more physical contact and fewer floating artifacts than Cosmos3-Nano.
Identifiable world models can also serve as verifiers. Generative models produce blurry objects, blurry arm motion and interaction hallucinations. Identifiable latents are fed through an Ada-Diffuser; Temporal Difference Verification then applies gradient guidance that penalizes physically inconsistent regions, pushing trajectories onto the valid dynamic manifold for reliable test-time guidance.
Unified Latent Action Models abandon robot-specific action spaces. Instead of recovering every hidden factor they recover only the shared latent actions that explain how the world changes across embodiments (partial identifiability, Kong & Xie 2022). A video foundation encoder produces latents; an inverse-dynamics stack of spatio-temporal ViTs and a forward stack of DiTs are trained inside a diffusion process with AdaLN, timestep conditioning, embodiment-ID classification and gradient reversal. At inference a single frame yields a transferable latent action that can be executed zero-shot on a new body.
Zero-shot transfer experiments confirm the point: LAD collapses into incoherent frames, LVP produces plausible video that ignores the source actions, while the latent-action approach successfully executes the same intrinsic behavior on the target embodiment.
Self-supervised skill learning needs neither demonstrations nor rewards. A skill embedding is contrastively aligned with state-action pairs: positives that move toward a goal are pulled together, negatives are pushed away (InfoNCE). Interaction-weighted resampling from the replay buffer focuses learning on meaningful interactions. Locomotion often follows linear-Gaussian dynamics; manipulation exhibits discontinuous “jumpy modes”. Local causal structure learning is the key that bridges these islands.
Hallucination has a clear data-allocation root. Policies only need the narrow distribution of optimal actions; world models need the broad distribution of suboptimal and exploratory actions. Action-free internet video is abundant for learning general dynamics, yet action-labeled robot data remains scarce. On-policy collection (Sailor, VLAW, World-VLA) limits generality; information-maximizing exploration runs into the information paradox where model uncertainty does not correlate with useful learning progress.
The practical learning framework therefore combines an adaptive curriculum that generates increasingly difficult tasks, a skill library of latent skill variables, and MIST-style masking: states are randomly masked, the model must maximize mutual information between the masked states and its predictions of observation and reward. This forces compact, task-relevant, causally structured representations.
WAV (World Action Verifier, Liu et al., arXiv:2604.01985) reframes the remaining problem as verification. Three core ideas: (1) semi-supervised setting that exploits more data, (2) decomposed verification that replaces one hard check with two easier ones (state plausibility + action reachability), (3) goal-oriented cycle consistency that couples inverse and forward models. Diverse subgoals are sampled from action-free video; sparse inverse dynamics extracts action-relevant features; the agent actively seeks the hardest plausible subgoal. On MiniGrid the method achieves the highest correlation with true error, lowest prediction error and strongest action following (near Oracle). On real robots it adapts to novel appearance, novel objects and shifts in policy optimality with only 200 target samples.
The high-level pipeline is now clear: Environment (with exploration that discovers novel representations) → Representation Learning (compress and distill from foundation models) → Structure Learning (incorporate domain knowledge and task feedback) → Decision-Making (adaptivity, compositionality, controllability).
Scaling alone makes world models broader but not necessarily physical or controllable. Causal hidden representations identify what the world is; latent actions identify where the control signals come from; simple causal principles enable models that are more physical, controllable and self-improving.
The ultimate question left open by all of this work remains: can we build a world model that not only dreams of the world, but also lets agents interact inside it, experiment, discover goals, and continuously refine the model itself?




















