RT @darrenangle: *sniff* *pulls shirt* You know, this is perfect - *gestures wildly* - this is the ultimate perversity of capitalism at its….

RT @LemmySmackett: "Okay, so imagine a magic button.". "I'm imagining the button.". "If you press the button—". "What color is it?". "It's….

Me tweeting at Delta to not lose my bags

Hey @GeminiApp , is the python-genai github being watched? I think grounding/citations is still broken and it doesn't seem like stuff is getting triaged? .

RT @JohnHolbein1: Timeless advice

Today I’ll be releasing my new eval called “naming things.” @StreamOnMax and @OpenAI.

I think @ESYudkowsky 's AI safety argument gets decently validated when xAI can't rollback an obvious problem and intentional bias in their system prompt quickly before everyone knows it. Whatever you think about what's happening in AI safety generally, it's clear that Grok needs.

There may have been some signs with Kanye in retrospect….

OpenAI has this, I just want something like this.

"Why don't you just Google it?" . No information on how it's priced in the forum:

First one's up: How can you know how much Gemini costs for tool use?.

It’s weird how Uncle Bob gets a twitter account and then everyone realizes that if you do what he suggests your company dies.

It seems like MLX doesn't actually increase speeds relative to GGUF for decently sized models on macs? > 20B params it definitely is slower at least on LM Studio.

I put together a small benchmark of LocalLLMs running on an m4 pro-max with a lot of ram to give a qualitative sense of how running local models feels/speeds, let me know if you want any other info.

Although Github link currently isn't found--I've emailed one of the authors (@ypwang61) and so I hope it'll be up soon!.

Full details, code, and prompts are in the paper. Worth reading if you are interested in data-efficient RL methods for language models and in understanding how exploration bonuses interact with reasoning ability.

Key takeaway: large language models already encode much of the required knowledge; a single well-selected, auto-verifiable example can unlock it through reinforcement learning.

Cross-domain evaluation shows spill-over benefits: training on a single Geometry question also improves Algebra, Number Theory and Combinatorics accuracy.