06 Oct 2025

OpenAI Agent Builder; WiFi Hears Heartbeats; Optimus Could Kick Your @ss

OpenAI Tightens Sora Rules After Viral Launch

OpenAI’s video app Sora sprinted to the top of the App Store and its first week was bonkers: think Pikachu mash-ups, Michael Jackson deepfakes, and celebrity likenesses everywhere doing and saying things they likely never would.

CEO Sam Altman says the company is pivoting (shocker!) from a loose opt-out posture to opt-in controls for rights holders (who can specify if/how characters and styles are used), and is exploring revenue sharing with creators whose IP fuels viral clips. Read Altman's post on the topic here.

The guardrails arrive as Sora’s “Cameos” feature, where users appear in AI-generated scenes, accelerates remix culture - and of course the legal risk.

The company also just announced a massive computing deal with AMD involving chips and data centers.

Breaking: According to the Wall Street Journal, "Under the terms of the deal, OpenAI committed to purchasing 6 gigawatts worth of AMD’s chips, starting with the MI450 chip next year. The ChatGPT maker will buy the chips either directly or through its cloud computing partners. AMD chief Lisa Su said in an interview Sunday that the deal will result in tens of billions of dollars in new revenue for the chip company over the next half-decade."

🧩 Jargon buster: Opt-in vs. opt-out - opt-in means your content can’t be used unless you say yes; opt-out means it can be used unless you say no.

Read more: The Rundown AI, AI Breakfast

Tesla’s Optimus Shows Off Real-Time Martial Arts

Tesla posted a short clip of Optimus sparring in real time - blocking, stepping and kicking without sped-up footage.

The hands still look limited, but balance and recovery are improving. It’s a stress test for quick perception-to-action loops that later translate into lifting, carrying, and navigating uneven spaces. The footwork appears smooth however hands remain mostly idle, indicating the 22-DOF hands are still in the lab.

According to AI Breakfast "Elon Musk confirms the demo is AI-driven, not remote-controlled. Optimus v2.5 is processing inputs and generating responses on the fly, a big step toward robots that can actually interact with humans and handle unpredictable environments."

Musk had previously said Tesla's aiming for 5,000 Optimus robots in 2025 for internal use in its factories. He also hinted that they want to reach 10,000 to 12,000 units worth of parts, with a further goal of 50,000 units in 2026.

🧩 Jargon buster: DOF (degrees of freedom)—how many ways a robot joint can move; more DOF usually means more dexterity.

Your Wi-Fi Router Can Hear Your Heartbeat

Researchers at UC Santa Cruz demoed Pulse-Fi, which uses commodity chips and Wi-Fi signal shifts to estimate heart rate as accurately as medical-grade wearables - no cameras or watches required - using a lightweight LSTM neural network.

That could turn homes and hospitals into ambient health monitors (and raise new privacy questions).

🧩 Jargon buster: LSTM—a type of neural net good at reading time-series data like fluctuating Wi-Fi signals.

OpenAI + Jony Ive Device Hits Snags

The always-on, screen-free device OpenAI is (apparently) building with Jony Ive may be slipping, according to the Financial Times.

Reports cite unresolved basics: how the assistant should “sound,” how to avoid over-talking, whether processing stays on-device or in the cloud and how to deliver low-latency responses without giant server bills.

The unit (reportedly phone-sized and “always on”) is still searching for the right balance of privacy, personality and compute.

🧩 Jargon buster: On-device vs. cloud inference - running the model locally boosts privacy/latency but limits power; cloud scales capability but adds cost and data exposure.

OpenAI Previews Agent Builder at DevDay

At DevDay 2025, OpenAI is unveiling Agent Builder, a visual tool for stitching together AI workflows - imagine drag-and-drop blocks for logic, loops, file search, and MCP connectors - so teams can go from prototype to production without bespoke scaffolding.

There’s also momentum on core models and platform integrations, plus demos where Sora 2 “answers” questions by generating short explanatory videos (early, but telling).

🧩 Jargon buster: MCP - a standard that lets agents securely plug into local tools and cloud apps.

Walmart Turns Every Pallet Into a Sensor

Walmart is rolling out Wiliot ambient IoT “pixels” (battery-free tags) across 4,600 stores and 40+ distribution centers, streaming location, temperature, and motion data to AI systems.

Instead of periodic inventory “visibility,” Walmart gets continuous inventory awareness - spotting spoilage, misplacement or delays in real time.

It’s a blueprint for retailers who’ve long struggled to know what they own and where it is.

Says AI Secret: "After decades of barcodes, RFID tags, and manual audits, Walmart may have finally closed retail’s oldest knowledge gap."

🧩 Jargon buster: Ambient IoT - Tiny, low-power tags that broadcast sensor data to nearby receivers without batteries.

The New Arms Race: Memory, Not Just GPUs

Hyperscalers and AI labs are locking down DRAM (random access memory), HBM, and NAND (memory that holds when you turn off you computer) supply years ahead - OpenAI’s “Stargate” alone reportedly already reserves a massive slice of global DRAM.

Memory makers (Micron, Samsung, SK hynix) are shifting fab spend toward HBM, leaving traditional buyers short. Startups are adapting with leaner models, synthetic data, and inference-first strategies.

🧩 Jargon buster: HBM—high-bandwidth memory paired with accelerators to feed models fast enough.

A Gamer Built a Tiny LLM Inside Minecraft

Creator sammyuri built CraftGPT, a five-million-parameter language model using Redstone logic (439 million blocks!).

It runs painfully slowly and outputs are rough, but it’s an extraordinary proof that tokenization, memory, and compute can be simulated inside a sandboxed game world.

🧩 Jargon buster: Parameters—the “knobs” a model learns; more isn’t always better, but it sets an upper bound on capacity.

Google’s PASTA Learns Your Visual Taste

0:00

/0:19

Google introduced PASTA, a system that adapts image generation to your personal style by watching which of four options you pick over repeated rounds.

Trained on thousands of human sessions and simulations, PASTA’s outputs were preferred 85% of the time versus a standard baseline - especially on abstract prompts.

According to The Rundown AI

PASTA presents users with four image variations per round, observing selections across turns to build a model on unique aesthetic preferences.
Researchers trained PASTA using 7,000 human sessions and 30,000 simulated scenarios, creating a framework that recognizes diverse visual preferences.
Google open-sourced both the interaction dataset and simulation tools, enabling other researchers to develop AI systems that personalize to users.

🧩 Jargon buster: Preference modeling - teaching models what you like by learning from your choices.

Tools & Launches

Gemini 2.5 Flash Image (GA) - Google’s image model (“Nano Banana”) hits general availability with more aspect ratios and character-consistent edits. Early adopters pair it with 3D pose tools and real-time pipelines for sub-10-second creative loops.
IBM Granite 4 (open models) - IBM’s new Granite 4 Mamba-Transformer family aims for lower memory use and efficient agentic workflows - solid for enterprises that need predictable cost and governance.
Perplexity Comet (browser) - Perplexity made its AI-native Comet browser free worldwide, adding a background assistant for power users - part search, part research aide, part memory layer.

Today’s Sources: AI Breakfast, AI Secret, The Rundown AI