☕️ AI robot learns tasks from one demo

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🤖 AI robot learns tasks from one demo

🤝 Marvell strikes $12.2B chip deal with Google

🔥 Mojo programming language goes open source

🧠 Cerebras new chip beats GPU racks 30x

🐦 Ornith-1.5 open models can self-improve

Plus: 💡 5 strategies & tactics, 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

🤖 AI robot learns tasks from one demo LINK

Generalist AI unveiled GEN-1.5, a robotics model that learns a new task from a single 3-to-12-second demonstration loaded into its context window as a "physical prompt," then executes with no additional training.

Across ten tests like opening a jar or pulling cash from a wallet, the model averaged 59% success zero-shot, rising to 83% after ten training steps on five minutes of data.

The model chains two prompts into longer sequences, uses simulation demos, and imitates human hand movements, behaviors Generalist says emerged during eight months of pretraining, though the tasks are short and all results come from the company itself, none independently verified.

🤝 Marvell strikes $12.2B chip deal with Google LINK

Marvell has granted Google a warrant to buy up to 59M of its shares, a stake worth about $12.2B if fully exercised, tying Google deeper to custom silicon it procures through fiscal 2033.

The deal supplies Google's TPU stack with AI inference accelerators, storage and network interface controllers, memory interface controllers, and near-memory compute, part of Google's effort to reduce reliance on Nvidia's general-purpose GPUs.

Most of the stake vests incrementally, with ~57.6M shares split across 240 segments that each unlock only after Google generates another $500M in custom-chip revenue for Marvell, so ownership scales with actual procurement rather than being guaranteed upfront.

🔥 Mojo programming language goes open source LINK

Qualcomm has open-sourced the full Mojo compiler and Modular's AI software stack under the Apache 2.0 license with LLVM exceptions, releasing the code on GitHub weeks after closing its ~$3.9 billion Modular acquisition in late July.

Announced at ModCon 2026, the Modular platform runs inference across Snapdragon chips, AMD and NVIDIA GPUs, Apple silicon, AWS Trainium, and Google TPUs, with AMD taking the stage to show support despite being a potential competitor.

Mojo, created by LLVM and Swift author Chris Lattner, hit its 1.0 milestone on Aug. 12, targets CPUs, GPUs, and AI accelerators with Python interoperability, and is adding Windows support alongside existing macOS and Linux builds.

🧠 Cerebras new chip beats GPU racks 30x LINK

Cerebras launched its CS-4 rack-scale system built on the new WSE-3 Turbo chip, generating GPT-OSS 120B output ~30x faster than a comparable GPU rack in the company's own benchmark.

Each WSE-3T wafer holds 4 trillion transistors, 900,000 cores, and 44 GB of on-wafer SRAM, delivering 125 PFLOPS dense (250 with sparsity) while cutting latency from 5ms to 2ms.

A single CS-4 rack packs three WSE-3T chips for 750 PFLOPS and 10x throughput per watt over CS-3, scaling to clusters for 50T+ parameter models, though the 30x figure comes from Cerebras' own GPU comparison.

🐦 Ornith-1.5 open models can self-improve LINK

Ornith has released Ornith-1.5, a family of open models that closes a self-improvement loop by proposing their own training tasks, generating a scaffold for each, and producing the rollouts used for reinforcement learning.

The lineup spans a 397B MoE flagship, a 35B MoE activating 3B per token, and a 9B dense model with a quantized Mobile build, all trained via GRPO across a three-stage cycle that targets a 0.2 success rate to keep tasks at the capability frontier.

The 397B scores 85.1 on Terminal-Bench 2.1, 56.0 on DeepSWE, 92.8 on GPQA Diamond, and 86.6 on BrowseComp, which Ornith reports as on par with Claude Opus 4.8, though its numbers come from the company's own published tables.

💡 Strategies & Tactics

> [AINews] Death of Params: Z.ai CEO Jie Tang on GLM 5.3 and the new Post-training Scaling Law: Z.ai CEO Jie Tang argues that model quality now hinges on training data and post-training reinforcement, not parameter count, as shown by GLM 5.3's gains.

> Building Federated Multimodal AI Workflows with NVIDIA FLARE: Train multimodal AI across organizations that can't share raw data by exchanging only lightweight model adjustments, cutting network and memory costs while keeping data local.

> Post-Train NVIDIA Cosmos 3 Edge for On-Device Robot Control: Explains how to adapt Nvidia's compact Cosmos 3 Edge world model so a robot can run manipulation policies entirely on its onboard computer, no data-center GPU needed.

> Qwen3.8-Max Just Passed Claude Fable 5 on the Frontend Leaderboard. We Compared Them on 10 UIs: Alibaba's cheaper Qwen3.8-Max matches Anthropic's pricier Claude Fable 5 on frontend UI design tasks, giving developers a lower-cost alternative for interface work.

> GraphRAG: How AI Answers Questions Hidden Across Many Documents: GraphRAG answers questions whose answers span an entire document collection by mapping how entities connect and pre-summarizing clusters, where ordinary similarity search fails.

Other news you might like

• Developing NVIDIA Holoscan applications with CLI, skills, and AI coding agentsLINK

• CISA warns of hackers exploiting critical MLflow vulnerabilityLINK

• LFM2.5 Q4\_0 Checkpoints from Quantization-Aware DistillationLINK

• India’s Murf AI launches Falcon 2 to challenge voice AI leadersLINK

• Repo Radar: deepsec, the One Repo Worth Your WeekLINK

• Nvidia launches NeMo Switchyard to cut AI model costs by 74%LINK

🧰 Trending tools

Attyn: an AI cursor tool that rewrites text, transcribes speech, explains on-screen content, and visualizes answers directly inside your apps.LINK

Clears: converts user stories into reviewed pull requests across multiple repos, automating routine coding so engineers focus on higher-value work.LINK

Replay QA for Teams: continuously tests your GitHub repo on every update, catches bugs, explains root causes, and hands your coding agent the fix.LINK

Zetik: monitors news, podcasts, papers, and code for you, then delivers concise briefings via feeds, push notifications, newsletters, or RSS.LINK

Skriptr: pairs students with an AI research and writing agent in a collaborative workspace, keeping them in control of their work.LINK

Hoplite: migrates your local coding agent setup to the cloud, letting agents run in parallel and keep working even when your laptop sleeps.LINK

📚 Trending papers & reports

Video generation speedups come from a training-free shortcut that skips most of a video model's internal comparisons yet keeps output quality, running roughly 1.5x to 2.6x faster while touching only about a quarter of the work.LINK

AI agent audit trails map an agent's claims back to the exact actions, files, and checks behind them, so reviewers can quickly verify whether an agent's work is correct instead of reconstructing it by hand.LINK

Number-crunching inside language models shows that Llama 3.1 genuinely tracks the structure of number sequences, computing differences between values on its own, rather than just pattern-matching, which supports trusting these models for time-series forecasting.LINK

Vector-symbolic decision-making gives AI agents a compact, fixed-size memory of past experience that speeds learning without adding runtime overhead, and stays reliable even when hardware errors flip stored bits.LINK

Emotion-reading video models keep working accurately even when facial, voice, or language cues go missing or noisy at deployment, using a stable-teacher approach that prevents the system from drifting and degrading over time.LINK

See you tomorrow for a new dose of ☕️ AIpresso!