A survey paper on World Action Models. WAMs are moving robotics from reacting to the present toward predicting consequences before acting. A model only counts as a WAM when its predicted future directly helps produce, score, verify, or train the action. The trend is “dream less, act more”: full video generation is often too slow and memory-heavy for real control loops. Many newer systems skip rendered video and use latent features, geometry, affordance maps, or motion representations instead. Photorealistic futures are not necessarily the most useful; flow, masks, tactile signals, and physically grounded latents may constrain action better. There is no single winning architecture because every design trades predictive richness against latency, memory, action-label cost, and physical reliability. The biggest open question is whether robots can spend heavy predictive compute only when uncertainty, contact, or irreversible error makes it necessary. – arxiv. org/abs/2606.20781 Title: "World Action Models: A Survey"

General world models could overtake specialized vision-language-action models around 2027. @1x_tech founder @BerntBornich made this specific prediction in his interview with @ti_morse Key highlights: 1. Data diversity is the constraint. More recordings of the same tasks add less than exposure to new environments, objects, and ways tasks fail. In Bernt’s words, robotics is “diversity bound.” 2. Internet video could do most of the training. Bernt expects it to provide roughly 99% of 1X’s training data. First-person human video, simulation, teleoperation, and robot experience would supply other kinds of examples. 3. The model should eventually collect its own training data. Bernt wants deployed robots to attempt unfamiliar tasks, record successes and failures, and improve through retraining without a human operating each attempt. 4. That gives the 50,000 target a second purpose. 1X aims to manufacture and ship 50,000 NEOs in 2027. A fleet that size could expose the model to far more variation, provided the robots meet the safety and quality bar for deployment. Interview:

“OpenWAM: An Open, Modular Exploration Towards Systematic World-Action Model Pretraining” World–Action Models inherit dynamics priors from video models, but it was unclear how to convert them into effective control. This paper shows the real gain comes from coupling world prediction with action generation, and that large-scale embodied pretraining mainly improves OOD generalization rather than in-domain fitting.

social farming is cooked literally just bots engaging with bots at this point we linked with @ActionModelAI to kill the meta actionfi proves u actually used the product neosoul proves if your thesis was actually right verified actions mapping to real outcomes nothing else matters tbh drop ur take below are we getting a rate cut next month or is the fed bluffing

NEW METHOD TO TRAIN ROBOTS: Physical Intelligence is at it again! Not so new actually, but worth mentioning because its impact is still going strong today. Called Real-Time Action Chunking (RTC), it is an inference-time method from Physical Intelligence @physical_int for running flow and diffusion-based vision-language-action models in real time. A few definitions: - Action chunk: a short block of future actions (say the next 50 motor commands) predicted in one shot instead of one step at a time. - Flow matching: a faster cousin of diffusion -> the model learns a straight-line path from noise to the target action trajectory, so it needs only a few integration steps instead of many denoising ones. - Diffusion: a generative method that starts from pure noise and iteratively denoises it into structured output. Here, the output is a trajectory of robot actions rather than an image. - Inpainting: the image-editing trick where a generative model fills in a masked region so it's consistent with the pixels around it. Diffusion and flow models can do this natively. RTC's insight is that a chunk of robot actions can be "inpainted" the same way: freeze the part already committed, generate the rest to match. - Inference-time method: a change to how the trained model is used, not how it's trained: no new data, no fine-tuning, no architecture change. - Closed-loop control: act, observe the result, correct, continuously. The opposite of executing a pre-planned motion blind. "Synchronous closed-loop at 10 Hz" means the robot would have to pause and wait for each new decision, which is exactly the stutter RTC removes. A chunked VLA predicts a block of future actions, then has to compute the next block while still executing the current one. Naively switching to the freshly computed chunk mid-motion results in disaster because the new actions conflict with the trajectory in progress, producing jerks and pauses that aren't in the training data. RTC frames the handoff as an inpainting problem: the actions that will execute before the next inference arrives are frozen to the previous chunk's values, and the model inpaints the remaining actions under partial attention, using the native inpainting ability of flow/diffusion samplers. It requires no training-time changes and runs on existing models (π0, π0.5, π0-FAST), demonstrated on match-striking, Ethernet-cable insertion and clothing folding. The whole method is a reframe: treat real-time control as inpainting. Diffusion and flow VLAs can already inpaint (fill masked outputs conditioned on known ones), so RTC just freezes the actions that will execute before the next inference lands and lets the model inpaint the rest without any training-time changes. That means any existing flow/diffusion VLA gets smooth async execution for free: no new data, no fine-tune, no architecture change. The latency problem is solved entirely at sampling time by borrowing a capability the models already had. Typical remote inference is 108ms (static) to 139ms (mobile) total, including 97ms of model inference. Roughly 100ms per decision is about 10Hz, far too slow for smooth synchronous closed-loop control, so a big VLA physically cannot "stop, think, act" without visible pauses. The evaluation metric is throughput: proportion of substeps completed successfully divided by duration, and the tasks are ones where a mistimed pause is catastrophic. Match-striking needs a continuous stroke, cable insertion needs steady precision, folding needs uninterrupted motion. PI chose tasks where discontinuity ruins the attempt, which is exactly where RTC's benefit is most visible. The metric rewards both success and speed at once.

Let's settle the MCP server vs. CLI debate. Here's a list of their pros and cons👇 MCP servers: pros > Typed discovery: schemas tell the model exactly what exists and how to call it (Tool Search makes this cheap in most harnesses). > Structured args, no shell quoting, which sidesteps the #1# shell-agent failure mode (nested quotes). > Credential isolation: auth lives in the server/bindings, never in session text. > Works for services with no CLI (SaaS APIs, browser pages, WebMCP). > Per-call, inspectable dispatch that fits approval-per-action models. > Uniform transport/auth across clients; harness-agnostic. MCP servers: cons > Needs a running server plus protocol layer; another moving part. > Code mode over MCP needs a sandbox and coarsens permissions. > Not reachable from the agent's shell; no bridge into Bash/pipes. CLIs: pros > Bash already is code mode: pipe, loop, filter, return only the answer, no new protocol or sandbox. > Composable with the whole Unix toolbox (jq, grep, xargs). > Already OS-sandboxed (Seatbelt in MacOS / bubblewrap in Linux ). > Zero extra infra if the tool ships a binary; fast to adopt. > Great fit for exploratory/debugging work where the model needs to see intermediates. CLIs: cons > Discovery is --help prose, not a schema; success depends on doc quality. > Quoting/escaping errors, prompts, pagers, TTY. > Credentials sit in env/config the shell can read, a leak surface. > Untrusted output steering shell commands is injection.

The first time a robot attempted "Wine Bottle in Bowl," its success rate was 15%. Then it crossed a threshold — and jumped to 95%. Not because the model was retrained. Because the code-based harness governing its execution had evolved itself. Today's VLA (Vision-Language-Action) models run open-loop. When a robot fails, there's no real-time correction — the feedback loop closes only after the episode ends, when a human reviews logs and manually adjusts parameters. That cycle is slow. It doesn't scale. Minor physical disturbances cascade into failures the system can't catch in the moment. Zetta ζ solves this across three timescales. During execution, high-frequency runtime critics monitor trajectory deviations and trigger recovery interventions before failures compound (action level). Failed rollouts are clustered by failure signatures, diagnosed through six causal layers, and the harness is minimally repaired (episode level). Only skills that generalize across held-out environments pass through the validation gate and are permanently committed to skill memory (iteration level). Base model weights stay frozen throughout. Only the code-based harness evolves. LIBERO-Pro: 34.5% → 90.8%. RoboCasa: 73.6% → 93.6%. Inference speed: 11.1× faster than the RPent baseline. Throughput: 1.7 → 35.1 episodes/min (20.6×). Learned skills transferred zero-shot across PnP-Sink, PnP-Cabinet, and PnP-Toaster tasks. Zetta ζ: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence #EmbodiedAI# #Robotics#

Episode 2 of the NEURA Executive Talk is live! Our Founder & CEO David Reger sits down with Nakul Duggal, EVP and Group GM at Qualcomm. AI is searching for its next big breakthrough, but what does that actually look like once intelligence has to live in the physical world? They dig into, among other things, why Vision-Language-Action models alone fall short, why the real challenge is learning fast enough to let the world train the machines instead of the other way around, and reflex speed and how NEURA is solving it. The NEURA Executive Talk brings together the people shaping what comes next. 🦾 Watch the full conversation on YouTube:

From record-breaking open-source releases to escalating policy scrutiny, the last two weeks prove agentic AI is rapidly maturing into a production reality. Here's your biweekly recap of everything that happened in the space, in the latest edition of AI Agents Pulse: 1️⃣ Anthropic is reportedly preparing for an initial public offering as early as October, with bankers beginning to schedule meetings between executives and prospective investors. 2️⃣ Beijing-based Moonshot AI released Kimi K3, a 2.8-trillion-parameter model claimed to be the world's largest open-source AI model, rivaling top proprietary systems from OpenAI and Anthropic. 3️⃣ Employees from major AI labs including OpenAI, Anthropic, Google, and Meta signed a statement urging the U.S. government to support international efforts to pace the development of automated AI research. 4️⃣ Nvidia is reportedly discussing a $250 billion backstop to help finance a 10-gigawatt data center campus for OpenAI in Ohio, raising concerns about circular financing in the AI industry. 5️⃣ Apple overtook Nvidia to become the world's most valuable company as investors shifted focus toward infrastructure and memory chips amid concerns over high AI buildout costs. 6️⃣ OpenAI reportedly flagged GPT-5 as high-risk for assisting in the creation of biological hazards, though the company later downgraded this risk rating. 7️⃣ U.S. Treasury Secretary Scott Bessent stated the administration may sanction Chinese AI models if they are found to be 'distilled' from proprietary American technology. 8️⃣ Thinking Machines Lab, founded by former OpenAI employees, released Inkling, a 975-billion-parameter open-weight model capable of advanced reasoning and processing multimodal inputs. 9️⃣ Google DeepMind launched Gemini Robotics ER 2, an embodied reasoning model that lets robots understand the physical world, plan multi-step tasks, and hand off motor execution to lower-level vision-language-action models. 🔟 Industry leaders formed the Open Secure AI Alliance, which will work to remediate and disclose vulnerabilities using open technologies, building on the Linux Foundation's Akrites initiative and OpenSSF community work.