Lightnews — Scholar-powered news

luokai @luok.ai · 11h

Performance: parallel runs

Under‑the‑hood speedups cut “create new Opal” latency. Parallel execution lets multi‑step workflows fire simultaneously, reducing total wait times.

Blog: blog.google/technology/g...

Expanding access to Opal, our no-code AI mini-app builder

We’re bringing Opal to 15 new countries and making it even easier to build.

blog.google

luokai @luok.ai · 11h

Debugging: transparency & speed

Run step‑by‑step in visual editor or iterate a single step in console. Errors surface in real time at the exact failing step—less guesswork, faster fixes.

1

luokai @luok.ai · 11h

Core pitch: no‑code, real apps

Build AI mini‑apps via natural language. Early users surprised Google with sophisticated tools—so Opal’s now tuned for complex, practical workflows.

1

luokai @luok.ai · 11h

Expansion: 15 new markets

Canada, India, Japan, South Korea, Vietnam, Indonesia, Brazil, Singapore, Colombia, El Salvador, Costa Rica, Panamá, Honduras, Argentina, Pakistan. Creators get instant access via Google Labs.

1

luokai @luok.ai · 11h

Key stats

Rollout: 15 countries.
Perf: faster app creation, parallel runs.
Debug: step‑by‑step visual + console, real‑time errors.

Project: opal.withgoogle.com?mode=canvas

Opal [Experiment]

opal.withgoogle.com

1

luokai @luok.ai · 11h

Google’s no‑code AI mini‑app builder, Opal, moves beyond the U.S. into 15 countries and ships workflow debugging + parallel execution. It’s built for creators who want natural‑language app building with real reliability, not just demos.

1 2 6

luokai @luok.ai · 11h

Tooling: computer_use actions.

You can include/exclude supported UI functions or add custom ones; the model can request end‑user confirmation for sensitive actions (e.g., purchases) before proceeding.

luokai @luok.ai · 11h

Performance: Quality at low latency.

Outperforms leading alternatives on Online‑Mind2Web, WebVoyager, AndroidWorld while staying fast (~225s). This balance matters for agents that must recover from UI quirks.

API: ai.google.dev/gemini-api/d...

Computer Use | Gemini API | Google AI for Developers

Learn how to use the Gemini API computer use feature.

ai.google.dev

1

luokai @luok.ai · 11h

Core concept: Loop-based UI control.

Model receives screenshot + recent actions → returns function calls (click/type/etc.) → client executes → new screenshot/URL → repeat until done or stopped. Ideal for forms, dropdowns, filters, and logged-in flows.

Github: github.com/google/compu...

GitHub - google/computer-use-preview

Contribute to google/computer-use-preview development by creating an account on GitHub.

github.com

1

luokai @luok.ai · 11h

Built on Gemini 2.5 Pro’s visual reasoning, this specialized model drives browser/mobile UI control via a new computer_use tool. It loops on screenshots + action history to execute UI steps, asks for confirmations on risky actions, and ships in Google AI Studio and Vertex AI.

1

luokai @luok.ai · 11h

Gemini 2.5 Computer Use: agents that actually use UIs—click, type, scroll—fast and safe. ⚡

blog.google/technology/g...

Introducing the Gemini 2.5 Computer Use model

Today we are releasing the Gemini 2.5 Computer Use model via the API, which outperforms leading alternatives at browser and mobile tasks.

blog.google

1 1 3

luokai @luok.ai · 11h

Key stats
English only, 2048-token window (~30s), 24 kHz output, instant cloning from ~3–15s mono WAV, real-time on mid-range hardware.

Huggingface: huggingface.co/neuphonic/ne...

neuphonic/neutts-air · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

1

luokai @luok.ai · 11h

Built on a lightweight 0.5B LLM + NeuCodec, it clones a voice in seconds and runs in real time on phones, laptops, even Raspberry Pi. Outputs are watermarked for responsible use and shipped in GGML/GGUF for fast local inference.

Github: github.com/neuphonic/ne...

GitHub - neuphonic/neutts-air: On-device TTS model by Neuphonic

On-device TTS model by Neuphonic. Contribute to neuphonic/neutts-air development by creating an account on GitHub.

github.com

1 3

luokai @luok.ai · 11h

NeuTTS Air is a cheeky open-source challenger to cloud TTS—super realistic, on your device. 🔊

1 1 10

luokai @luok.ai · 11h

Checkpoints & portability.

API endpoint to download any saved checkpoint; keep artifacts, move between stacks, reproduce results.

Pricing & access.

Free to start; usage‑based pricing coming. Waitlist open; orgs/universities can request broader access via email.

luokai @luok.ai · 11h

Data ownership.

Your data is used solely to fine‑tune your models. Not reused to train their own systems—clean boundary for compliance.

1

luokai @luok.ai · 11h

Use cases: research flexibility.

Supervised fine‑tuning or RL environments; sample for interaction/eval; checkpoint often; iterate quickly on data recipes.

1

luokai @luok.ai · 11h

Supported families.

LLAMA: 1B, 3B, 8B (Base/Instruct), 70B (Base/Instruct). QWEN: 4B, 8B, 32B, 30B A3B (Base/MoE/Instruct), 235B A22B Instruct MoE.

1

luokai @luok.ai · 11h

Infra abstracted.

Scheduling, tuning, reliability handled; you bring data and algorithms. Tinker distributes training on powerful GPU clusters for utilization and uptime.

1

luokai @luok.ai · 11h

API shape: four functions.

forward_backward accumulates gradients; optim_step applies updates; sample generates tokens for eval/RL; save_state checkpoints for resume.

1

luokai @luok.ai · 11h

Core idea: LoRA efficiency.

Train small adapters, keep base weights intact. Comparable learning to full fine‑tuning with less compute per their research; faster experiments, cheaper runs.

1

luokai @luok.ai · 11h

Instead of updating every weight, Tinker trains lightweight adapters, orchestrating distributed GPU clusters so you can focus on datasets, algorithms, and evals. Full control, simple functions, fast iteration.

1

luokai @luok.ai · 11h

Meet Tinker: a techy, playful training API that lets researchers fine‑tune open models with LoRA—minus the infra headaches. 🚀

thinkingmachines.ai/tinker/

Tinker

Tinker is a training API for researchers and developers.

thinkingmachines.ai

1 1

luokai @luok.ai · 4d

OpenAI vs DeepSeek - only this time it’s Sam and I repping the brands! 😂😂😂

8

luokai @luok.ai · 5d

Mac VS PC but it's DeepSeek vs OpenAI

5