Joo et al. present findings on the use of the 310LR-C Dual-Mode Lever in soft electrostatic actuators, resulting in a significant performance improvement for jumping robots. This research could inform future developments in robotics and actuator design.
Joo et al. (Nature Communications) used our 310LR-C Dual-Mode Lever in key experiments on ultralight soft electrostatic actuators. The result? A jumping robot achieving 60% greater jump height:
This tweet outlines the technical architecture of an AI agent, highlighting dedicated resources, intelligent routing, and cost-saving measures. Senior engineers may find the details on prompt caching and cross-session memory particularly relevant for optimizing AI system performance.
10/
the tech under the hood for the curious:
- dedicated CPU/VM per agent (not shared)
- claude sonnet 4.6 with intelligent routing between anthropic models for cost and task efficiency
- prompt caching (90% cost reduction)
- cross-session memory with daily summaries
- 30+
This tweet outlines practical strategies for optimizing AI model inference, emphasizing infrastructure considerations like model quantization and prompt caching. Senior engineers will find these insights valuable for building robust AI systems that can handle real-world demands.
Treat AI like infra, not lipstick.
Playbook:
• quantize models; host near users
• vector DB for embeddings + rerank
• batch & cache prompts (deterministic)
• async workers + circuit-breaker
Save 3–10× on inference. Build for failure. #AI #SRE
The LeWorldModel demonstrates effective ball localization and next-frame prediction after extensive training, showcasing advancements in action conditioning. This could inform future model development for real-time game AI applications.
After 100 epochs of JEPA training on ≈6,000 frames, the LeWorldModel has learned to:
Localise the ball with measurable structure
Predict next-frame ball position through the ARPredictor
Localise the paddle with action context confirming working action conditioning
Meta has released its first model from the Superintelligence Labs, which may indicate a shift in their AI strategy. Senior engineers should evaluate its capabilities and potential integration into existing systems.
Top stories in AI today:
- Meta Superintelligence Labs ships first model
- HeyGen’s Avatar V solves AI’s identity drift
- Build an automated ad generator with this tool
- Anthropic simplifies the agent-building system
- 4 new AI tools, community workflows, and more
This tweet discusses a novel soft robot design that utilizes heat-responsive materials and embedded electronics for movement without traditional mechanical systems. Senior engineers may find the innovative approach to robotics and materials science relevant for future applications in AI and automation.
A new origami-inspired soft robot uses heat-responsive materials and embedded electronics to move, fold, and reshape itself, without motors, pumps, or bulky mechanical systems.
@Princeton
This tweet presents a comparative benchmark of three AI models on a Raspberry Pi 5, highlighting performance differences in cold-start, sustained throughput, and RAM usage. Senior engineers may find the insights useful for selecting the right model for edge deployment.
72 hours of LiteRT-LM vs Ollama vs llama.cpp on a Pi 5 8GB ($160 board, post-DRAM-hike pricing).
Clean result:
→ LiteRT-LM wins cold-start by ~30%
→ Ollama wins sustained tok/s
→ llama.cpp still holds the RAM headroom past 4k context
No single "best edge runtime" on Pi. Pick
The tweet discusses Aave's transition plan to shift risk management to decentralized infrastructure, highlighting a significant move in DeFi. Senior engineers should note the implications for on-chain finance and risk management systems.
If you believe global finance belongs onchain, you cannot rely on centralized, off-chain risk silos.
@LlamaRisk
’s transition plan for Aave shifts risk management to neutral, trusted infrastructure.
DeFi will win with
@aave
V4.
The tweet highlights the adoption of Chinese open source AI models by notable companies like Cursor and Cognition, indicating a shift in the AI landscape. Senior engineers should note the implications of this trend on competition and innovation in AI infrastructure.
Silicon Valley is quietly running on Chinese open source AI models.
Here are the receipts:
→ Cursor confirmed last month that Composer 2 is built on Moonshot's Kimi K2.5
→ Cognition's SWE-1.6 model is likely post-trained on Zhipu's GLM
→ Shopify saved $5M a year by
The tweet outlines significant infrastructure progress, including the integration of OpenAI's API and various systems into DGX Spark. This is relevant for engineers focused on building robust AI systems and infrastructure.
45 min in. Infrastructure phase nearly done.
OpenAI API wired (gpt-5.4 confirmed live)
Gemma 4 26B pulled to DGX Spark #1 (17GB, MoE)
Nemotron replicated to DGX Spark #2 (86GB)
Codex CLI installed
Firecrawl MCP wired
1Password CLI resurrected
Cron roster: 7
The tweet discusses Gemma 4's use of shared KV cache layers, which allows it to run on a laptop but also highlights a limitation in cache reuse for llama.cpp. This insight into architecture could be relevant for engineers working on efficient AI system designs.
There is a catch nobody is talking about.
Gemma 4 uses shared KV cache layers - the last layers reuse K/V tensors from earlier layers instead of computing their own. That is why it fits on a laptop.
But that same architecture breaks cache reuse in llama.cpp. Every request
Gemma 4 demonstrates impressive efficiency with 27B parameters, outperforming Llama 3.1 at 405B levels. This benchmark highlights the trend towards more efficient models without the need for extensive infrastructure.
Gemma 4's a beast hitting Llama 3.1 405B levels on benchmarks with just 27B params.
That's efficiency on steroids, no data center apocalypse required.
Open source winning big.
ConvApparel is a new dataset aimed at improving LLM-based user simulators by quantifying the 'realism gap.' This could be relevant for engineers focused on enhancing conversational agent training methodologies.
Introducing ConvApparel, a new human-AI conversation dataset, as well as a comprehensive evaluation framework designed to quantify the "realism gap" in LLM-based user simulators and improve the training of robust conversational agents.
Read all about it →
goo.gle/41k5eff
KellyBench tested frontier AI models in a simulated betting market, revealing that all models lost money, with varying degrees of ROI. This highlights the challenges and limitations of current AI models in real-world applications, which is crucial for engineers to consider.
Interesting new benchmark called KellyBench which put frontier models in a simulated Premier League betting market for a full season. Every model lost money.
- Claude Opus 4.6: -11% mean ROI, avoided ruin
- GPT-5.4: -13.6% mean ROI, avoided ruin
- Grok 4.20: -88.2% ROI, went
Fastly's integration of Compute and Semantic Caching optimizes AI agent performance by reducing operational costs at the network edge. This could be relevant for engineers looking to improve the efficiency of deploying AI models in production environments.
$FSLY Fastly optimizes Claude Managed Agents by moving intelligence to the network edge. Integrating Fastly Compute and Semantic Caching significantly lowers the cost of running frontier models / AI agents. Claude Opus 4.6 charges per token for every interaction, for example.
The announcement of MCP as a universal standard for AI agents indicates a significant shift in open-source AI, potentially impacting how AI systems are built and integrated. Senior engineers should monitor this trend as it may influence future infrastructure and development practices.
AI Agents just became the fastest-growing category in all of open-source AI.
Here's why this matters — and why 2026 is the year of the agent:
MCP (Model Context Protocol) changed everything.
Anthropic open-sourced MCP in late 2024 as a universal standard for connecting AI