AI Reading for Wednesday October 2

OpenAI demoed an agent that would use the voice API to call a vendor and order strawberries - Tom's Guide

A demo: building a voice assistant with the new APIs (can skip to 8:00 for the demo) - YouTube

Gonna power a lot of spam calls.

---

OpenAI DevDay also had new stuff on prompt caching, fine-tuning, and model distillation, to fine-tune gpt4o-mini on outputs from gpt-4o, and other - Image Press

There are some RAG improvements. I really want better metadata support.

---

Some inside dish about Mira Murati's departure. Technical teams did not think o1 was ready, and indeed it's missing key features the other APIs have. She was often the middleman between Altman pressing for releases and tech staff asking for more time. Her loyalty was questioned after the Altman fracas. - Fortune

---

OpenAI billing bug freaks out victims - Reddit

Maybe fixed - status.openai.com

---

OpenAI's new logo will be a plain black circle, apparently - Fast Company

---

When your private chats are included in a public dataset (presumably because you clicked a box), or become discoverable in litigation. - The Atlantic

---

When your transcription assistant captures too much and shares it to the wrong person. - Washington Post

---

When your AI sysadmin bricks your computer - The Register

Was I not supposed to use ‘sudo’? If anybody had told me that sort of thing was frowned upon…

---

Meta ‘probably?’ trains on smart glass data? How can there not be full disclosure if it's training on everything it sees in your private moments? - MacRumors

I can’t wait til my agent, trained on all my private data and moments, calls someone to order strawberries and gets subjected to a little adversarial prompting.

---

Big Tech might come to regret pushing back on regs after an endless stream of AI disasters and sleazy players bring the whole AI enterprise into disrepute. - FT

---

Palmer Luckey wants to profit from an AI arms race. - Business Insider

---

US Army tests armed robot dogs. - Military.com

---

Mustafa Suleyman presented his vision for AI and assistants like Copilot - The Verge

---

Some more Copilot features: Paint gets generative infill and erase, Photos gets upscaling. - Engadget

---

Insiders say Google made progress on reasoning models to match OpenAI's o1. - Bloomberg

---

Brian Chen says Apple Intelligence is useful but not that compelling yet. The promise is pretty clear though, I have this nagging feeling they will charge excessively for the useful stuff when it comes out. - NY Times

---

Apple rethinking Vision Pro as Meta advances. They are both exploring and haven't reached product-market fit. But as the cost and power requirement comes down the market seems like AR will be taking off in a couple of years. - NewsBytes

---

Apple readies new AI-capable devices, including a low-end iPhone and iPad Airs - MacRumors

---

Kling wins 3/5 tests vs. Runway - Tom's Guide

---

An AI assistant for board meetings? Summarization and synthesizing OK, decisionmaking, could we not? - Fast Company

Are board members noobs at making decisions or something?

---

MIT spinoff claims new architecture beats transformer-based LLMs - Silicon Angle

---

Run Llama-405b on a cluster of 128 GB MacBook Pros? That'll be $5k per MacBook but still cheaper than 8xH100 which I think is what 405b targets. - Geeky Gadgets

---

Stanford inference framework Archon beats GPT-4o using a team of Llama-class models, using a sort of optimized mixture-of-experts approach. - VentureBeat

---

Nvidia dropped a new open-source multimodal model, says it beats leaders on some tasks. - VentureBeat

---

Anthropic hires OpenAI co-founder Durk Kingma. - TechCrunch

There are now more co-founders of OpenAI working at Anthropic than at OpenAI

---

Transforming a deep dive into poop and farts into that NPR-style NotebookLM podcast. - Ars Technica

---

Follow the latest AI headlines via SkynetAndChill.com on Bluesky