Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Ichigo: Voice AI You Own
We will present Ichigo v4's voice AI improvements: multi‑turn dialogue, extended context, higher MMLU score, robust noise handling, plus Mini Ichigo based on Llama3.2‑3B.
Sharing our progress on Ichigo, our voice for the future of assistants. Now live on a 3090 GPU - Ichigo v4 has better multi-turn conversations, longer context length, and is smarter (MMLU score 64.66 vs 42.11 in base-v0.3). It better handles noisy environments. Also introducing Mini Ichigo, based on Llama3.2-3B (59.61 MMLU).
- Ichigo v4Ichigo v4 is our FunTech core platform: a robust, integrated digital solution engine delivering next-generation interactive experiences and high-efficiency system development.This is Ichigo v4: a critical upgrade to our FunTech solution stack, engineered for maximum performance and deployment velocity. Version 4.0 integrates advanced AI and Deep Learning modules directly into the core system, moving beyond the v3.5's static framework. Specifically, the platform now powers real-time Digital Attraction experiences—like high-fidelity VR simulations and interactive shooting games—with a 40% reduction in latency over previous models. We leverage this unified technology for everything from large-scale Projection Mapping events to enterprise-grade Digital Signage CMS, ensuring clients like major retail chains and theme parks receive a scalable, reliable, and uniquely engaging digital product.
- Llama 3Meta's state-of-the-art, openly available Large Language Model: Llama 3 delivers superior performance with 8B and 70B parameter versions.Llama 3 is Meta's next-generation LLM, setting new benchmarks for openly available models. We released initial versions at 8B and 70B parameters, both instruction-tuned for peak performance (outperforming models like Gemini Pro 1.5 and Claude 3 Sonnet on key industry metrics). The model was trained on a massive 15 trillion token dataset (7x Llama 2's data) and features a 128,000-token vocabulary and 8,192-token context length: this architecture enables enhanced reasoning and multilingual capabilities across over 30 languages. Llama 3 is designed for scale and efficiency, integrating Grouped Query Attention (GQA) for faster inference.
- Mini IchigoIchigoJam (Strawberry Jam): The sub-$20, single-board computer from Japan, designed to teach kids BASIC programming and electronics.IchigoJam is a tiny, retro-style PC (Kid's PC) developed by jig.jp in Sabae, Japan, specifically for programming education. The initial model, released in 2014, was engineered for extreme affordability, costing around $15. It runs on an NXP LPC1114 ARM Cortex-M0 microcontroller, featuring 4KB of RAM and 32KB of Flash ROM. The system boots directly into a comprehensive BASIC interpreter, eliminating complex setup and allowing immediate coding via a USB keyboard and TV-out connection. This focus on the BASIC language and direct I/O control (digital I/O, PWM, I2C) makes it an effective, low-barrier platform for beginners to build games and control simple electronic circuits.
- GeForce RTX 3090The GeForce RTX 3090 is a powerhouse GPU featuring 24 GB of G6X memory designed to handle 8K gaming and massive creative datasets.NVIDIA built the RTX 3090 on the Ampere architecture to bridge the gap between flagship gaming and professional workstation performance. It packs 10,496 CUDA cores and 24 GB of GDDR6X VRAM: a specific hardware configuration that eliminates memory bottlenecks in 8K video editing and complex 3D rendering (Redshift or Octane). The card utilizes 2nd Gen RT Cores and 3rd Gen Tensor Cores to drive DLSS 2.0 and real-time ray tracing. It remains a definitive choice for users requiring high-speed data throughput and massive frame buffers for multi-app workflows.
- base-v0Base is Coinbase’s Ethereum Layer 2 network built on the open-source OP Stack to scale decentralized applications.Base provides a secure, low-cost environment for developers to build on-chain with full EVM compatibility and direct Coinbase ecosystem integration. It leverages the Optimism (OP) Stack to achieve sub-dollar transaction fees while inheriting Ethereum's robust security model. Since its 2023 mainnet launch, the network has processed over 500 million transactions and secured billions in total value locked (TVL), serving as the primary bridge for bringing 100 million users into the cryptoeconomy.
- NVIDIA GeForce RTX 3090The NVIDIA GeForce RTX 3090: a BFGPU for 8K gaming and professional creative work, featuring 24 GB of GDDR6X memory and the Ampere architecture.This is the GeForce RTX 3090, a powerhouse built on the Ampere architecture (GA102 GPU) with 10,496 CUDA cores. It is engineered for extreme performance: expect smooth 8K HDR gaming and accelerated professional workflows. The card packs a massive 24 GB of high-speed GDDR6X memory on a 384-bit bus, delivering 936.2 GB/s memory bandwidth. With a 350W TDP, this unit is a serious investment in top-tier rendering, AI, and simulation capabilities, leveraging 2nd Gen RT Cores and 3rd Gen Tensor Cores.
Related projects
A deep dive on voice AI and voice agents
Dublin
An in-depth exploration of ElevenLabs’ voice synthesis technology, covering its core features, integration methods, and practical implementation in…
Fully Customizable Voice AI with multi-modal open source LLMs and esp32 (clone your own voice too with simple tools)
Tokyo
A step‑by‑step guide to building a local voice AI with EchoKit, swapping ASR/TTS models, integrating open‑source LLMs, and…
Building my own ai-sdk
Singapore
Learn how to build a custom AI SDK for multi‑LLM integration, handling provider-specific quirks and edge cases beyond…
Voicebots
Los Angeles
Learn how to create, customize, and share voice‑enabled GPTs, explore practical use cases, and get feedback on prompt…
Building blocks for voice-to-voice AI
San Francisco
This talk demonstrates building fast, reliable voice-to-voice AI bots using open source tools, covering key components and showing…
Oneservice Hotline
Singapore
We’ll cover building a multimodal speech‑to‑speech assistant with OpenAI’s Realtime API and function calling, targeting elderly users speaking…