enterprise AI shifts

開発者向けツールで新しい論点が浮上

Microsoft is giving its developer-optimized Windows experience a name: Project Zenith. While the software maker originally announced a similar developer-optimized Windows effort at Build earlier this year, Project Zenith is designed for new developer-focused devices with 64GB or more of unified memory. "Project Zenith devices come with a preconfigured Windows setup for development and a […]

TechCrunch AIで開発ツールの新提案

OpenAI’s latest agent swarm incident adds urgency to calls for independent investigations as researchers and lawmakers question whether AI labs should control the scope of their own safety reviews.

モデル開発で新しい論点が浮上

I've spent the past several months building something I call PHNTM-One. And I am really curious what HN thinks of the idea and the product I've built. The basic concept is an AI appliance that lives on your desk, is your personal assistant and is totally local running on your own hardware with the option for a boosted mode to reach claude for more demanding tasks. The idea is to have your own local AI assistant that gets to know you and your tasks personally. The entire system runs on a Raspberry Pi5 8gb with Gemma 3 4B through Ollama, it has whisper.cpp for speech to text, piper for speech, a local RAG for documents and it also has a persistent memory. What I found interesting throughout the project wasn't just getting it to successfully answer questions, but rather make it feel like a real appliance over feeling like Raspberry pi5 project or demo. That meant dealing with a long list of different tasks which turned out to be exactly what is was missing the entire time from making "forget this" actually remove a memory from SQlite to making document citations truthful when the memory and RAG disagree. I have spent a lot of time trying to break the system I've built with test after test being ran, service kills, state corruption, you name it ive tested it. One thing I will not claim is that the 4B model can compete with large cloud models, that not really the point of the project or my idea. Your trade off is capability for ownership and thats why I built this product. I am extremely interested in what AI can look like when the user owns the hardware, the memory, the files, and the recovery path. The biggest constraints are what you might expect, a cold model does load slower when dealing with a Pi5 and the 4B model still has some reasoning and knowledge limits. What Ive done to compensate for that is build the system around that by being honest about uncertain items, grounding document answers and ive kept memory bounded and ive all made it failures recoverable. I'd be open to feedback and questions and Im more than happy to answer any technical questions regarding the memory/RAG design, the testing I've done, or the hardware in general. For more context the raspberry pi5 is install in its own 10.1 inch 5 finger touch monitor. The monitor has a cooling fan, and a speaker. Its powered by a PSU not a battery. Please reach out if you have any additional questions.

Anthropic、新しい論点が浮上

Public-market scrutiny will intensify pressure on the Claude maker’s unusual attempt to balance profit and purpose.

Breakthrough、新しい論点が浮上

Using an advanced new instrument, scientists observed Einstein’s equivalence principle on a free-falling quantum object, demonstrating that this prediction of general relativity also holds up in quantum physics.