|
|
Hi there, this is your daily ☕️ AIpresso.
|
|
|
|
In today's newsletter: 🤖 Dyna's robots fold napkins on their own 🔓 Nvidia open-sources its Run ai software 📱 Liquid AI's new model runs agents on-device 🧠 New method rebuilds AI without retraining 🖥️ Tinygrad PC packs four AMD GPUs Plus: 🎁 10 more stories you might like, 🛠️ 5 engineering & practice picks, 🧰 7 tools, and 📚 5 papers.
|
|
Your agents work while you sleep
Give a Skydive agent an ongoing responsibility and they’ll handle it on schedule, every time.
Have them prep your morning report, research new leads, monitor customer feedback, or keep projects moving overnight. You wake up, the work is already done.
|
🤖 Dyna's robots fold napkins on their own
LINK
|
- Dyna Robotics showed Dyna-2.1, a physical robot agent built on its world-action model that runs long-horizon tasks autonomously for up to an hour, including folding restaurant napkins for 24 hours straight at a 99.4% success rate.
- The underlying Dyna-2 model was pre-trained on over one million hours of human video and uses a diffusion transformer with flow matching to predict future video and then emit smooth, precise robot actions, letting it generalize to unseen scenarios.
- Robots can pick up a new task from as little as 45 minutes of demonstrations, and Dyna already runs deployments folding thousands of napkins daily at Din Tai Fung, though Jason Ma flags dexterous hands that overheat and missing sensory data as current bottlenecks.
|
🔓 Nvidia open-sources its Run ai software
LINK
|
- Nvidia has released the full Run:ai platform as open-source software under the NVIDIA DSX AI Factory, giving developers and ecosystem partners production-grade GPU and AI workload orchestration on Kubernetes.
- Alongside the complete stack, four components ship independently: KAI Scheduler (Apache 2.0, a CNCF Sandbox project) for gang scheduling and GPU sharing, plus the GPU Fractioning Operator, KAI Resource Management, and Workload Map.
- The GPU Fractioning Operator retires Run:ai's proprietary fractioning for CUDA 13.4-native fractions with hard per-workload memory isolation, and the Community Product on NGC carries no paid subscription or Nvidia enterprise support entitlement.
|
📱 Liquid AI's new model runs agents on-device
LINK
|
- Liquid AI released LFM2.5-24B-A2B, a reasoning-only MoE that runs agents on-device with 24B total and just 2B active parameters per token, matching models carrying ~50% more parameters on agentic and tool-use tasks.
- Built on a 40-layer architecture with 64 experts and top-4 routing, it extends context from 32K to 131K tokens and was trained on ~33T tokens, decoding 184 tok/s on an M5 Max and over 21K output tok/s on a single B300.
- Open-weight on Hugging Face today with day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX, it leads on tool use and instruction following, but larger models keep an edge on knowledge and coding benchmarks like GPQA Diamond and SWE-Bench Verified, so pair it with retrieval for knowledge-heavy work.
|
🧠 New method rebuilds AI without retraining
LINK
|
- AWS introduced "priming," a method that converts pretrained Transformers into hybrid state space models by reusing their existing weights, cutting the pretraining cost of exploring hybrid architectures by more than 99%.
- Priming identifies which attention layers do only local work and swaps them for SSM layers, consuming under 0.5% of the source model's token budget while working across Qwen, Llama, and Mistral, dense or MoE, from 8B to 32B.
- A 32B GKA-based hybrid matches its source Transformer on reasoning benchmarks while decoding up to twice as fast at 128K context with half the KV cache; AWS released the model zoo, kernels, and a vLLM plugin under Apache 2.0.
|
🖥️ Tinygrad PC packs four AMD GPUs
LINK
|
- Tinygrad's redesigned Tinybox G4 workstation can house four AMD Radeon RX 9070 XT cards, pairing a server-grade EPYC motherboard with water cooling in a deliberately oversized enclosure for heavy deep-learning workloads.
- The base G4 starts at $7,000 using a 32-core EPYC chip with 32 GB DDR4 and PCIe 4.0; adding four 9070 XT cards costs $4,000, pushing a complete build to $11,000 before upgrades.
- Noctua fans plus liquid cooling keep the system under 50 dB, and Tinygrad builds each unit to order, though delivery runs anywhere from one to eight weeks after purchase.
|
|
Stop typing AI prompts. Start talking.
You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.
🛠️ Engineering & Practice
|
> Are harnesses worth the cost?: Adding an executable testing harness makes a cheaper AI coding model as reliable as a top-tier one while costing 33% less per run.
|
|
Other
news & articles you might like
-
Nvidia commits $250M of compute to UK AI safety research
LINK
-
VB pulse agentic orchestration report
LINK
-
The theory 2026 AI survey results
LINK
-
Anthropic acquires Opaque
LINK
-
Nvidia's China AI-chip share falls as Huawei closes gap, report finds
LINK
-
Apple researchers turn idle Macs into a virtual GPU cloud with ‘Minions’
LINK
-
Releasing ai generated software
LINK
-
Anthropic tests its next model under Claude Opus 5.5
LINK
-
Meet Helix World: a real-time audio
LINK
-
We’re powering European scientific discovery with advanced AI compute and frontier models.
LINK
|
|
📚 Trending papers & reports
|
> Confidence calibration gets baked into a model's normal training in one pass, so its predictions report trustworthy certainty levels without the usual extra data, tuning steps, or added components.
LINK
|
|
> What-if predictions can be answered straight from observed data, skipping the error-prone step of mapping cause and effect first, cutting loan-decision prediction error by over 20x and scaling to larger problems without retraining.
LINK
|
|
> Team-safety credit assignment pinpoints which cooperating robots or vehicles caused a shared safety breach, boosting safe task completion up to ~5x in transport and lifting traffic throughput by as much as ~68%.
LINK
|
|
> Graph generators secretly memorize their training data, reconstructing training examples far better than new ones even when standard similarity checks find no copying, a hidden overfitting problem that cannot be fixed by stopping training early.
LINK
|
|
> Feature cleanup tool lets recommendation systems safely retire unused data inputs without hurting results, flagging deprecable feature sets daily and tracing which models depend on each one, trimming up to 59 features in one tested model.
LINK
|
|
|
🧰 Tools & repos
| ElevenLabs: turns any script into studio-quality speech in seconds, with 5,000+ voices across 70+ languages. Free to start. GENERATE FOR FREE |
|
Articos MCP: tests SaaS messaging, positioning, and landing pages against simulated ICP-matched personas, delivering audience signal in under 30 minutes with 86% human accuracy.
LINK
|
|
Inferock: a local proxy for OpenAI, Anthropic, Gemini, and OpenRouter calls that tracks token usage, failures, and retries to reveal overbilling.
LINK
|
|
Playground by Google Labs: experimental hub for testing early Google AI products like search upgrades, Gmail and Docs assistants, and NotebookLM.
LINK
|
|
AgentSDR: open-source, self-hosted sales workspace running email, LinkedIn, and WhatsApp outreach sequences with data enrichment and a CRM that classifies replies automatically.
LINK
|
|
Comcent: browser-based voice platform linking Twilio or SIP providers with call flows, queues, AI transcription, sentiment scoring, and voice bots for calling teams.
LINK
|
|
Make0 Brain: a shared memory store for AI agents across ChatGPT, Claude, and Cursor, keeping your context in sync via MCP and imported conversations.
LINK
|
|
You can check the previous tools here, or add your tool here
|
|
|
Our AI Academy has 330+ step-by-step tutorials on ChatGPT, Claude, Perplexity, and every tool that matters. No fluff — just practical workflows you can use at work. Try it free for 7 days.
|
|
💬 How did you find today's edition?
We read every reply — just reply to this email and let us know how we can improve!
|
|
