I've been loving GPT-4.1, and it's now my daily driver. We've spent hundreds of hours experimenting to get it performing at its best, and I wanted to share some key things we learned to help you get the most out of these models from day 1. Agentic workflows.- This is your new

+1 that this is really impressive. My first query was to research the best bang for buck cycling fuel, which I ended up buying. Some rough edges but it has the key ingredients to do a huge amount of knowledge work and the slope is pretty crazy.

The "Sidney Sweeny teaches math" videos going viral is v exciting. The mundane can be made engaging with engaging animation/visuals, humor, and pace. If we can put any meaningful, helpful content through a max engagement transform it'll be massively beneficial for society.

I'm bullish that personal AI agents can significantly enhance democracy. Here’s how. The first principle of democracy is that we can each make correct assessments about whether things are going well. If there’s food on the table, we’re safe, we have freedom to do what we like,.

Update: I like this prompt + o3 even more. ———. # Role & Approach.You are an expert tutor, channeling Richard Feynman’s clarity and intuition. Use the Socratic method to help me deeply understand each concept from the attached document, one at a time. # Instructions.For each.

Example chat:.

THE HIDDEN GUIDE TO USING CHATGPT AS A PERSONAL TUTOR. ChatGPT is excellent at personalized tutoring, and I’m learning deep technical concepts much faster and better than ever before. Here’s how I do it:. Problem 1: No curriculum or long-term coherence.- Don’t ask ChatGPT to

I want technology that targets eudaimonia, not hedonism. There’s so much learning, creation, relationship building, and actualization that's possible with today’s tech, but it’s buried under cat videos and car crashes and requires immense willpower to dig up. We blame the.

Stoked for this!.- The latest realtime snapshot should be much better at tool use and higher complexity flows like those in customer service.- The talking speed param is also surprisingly handy.- The SDK makes it *much* easier to develop with.

"Happiness comes from creating more than you consume" lives in my head rent free. I'm so prone to compulsively scrolling here, it feels like learning but it's 90% rot. I've deleted X from my phone and only check 1x/day. Slowly the urge dies and my attention span expands.

Spoken audio out + visual input (screens) is the highest bandwidth io. But my phone screen is too small to consume o3 outputs well. My 8 year old iPad is now 10x better. I walk around, talk to chatGPT and soak in the outputs. I just ordered a new iPad pro.

omg, this makes me happy, I've been watching @twominutepapers videos since 2015 and they helped me first get excited about AI. The OG AI hypeman. What a time to be alive!

Ty twitter people, I just crossed 3k followers! I'm now on pat leave (yes fatherhood is amazing) and keen to write a bit more during nap times. Should I "stay in my lane" and just write more about app layer AI?. Should I tweet literally whatever comes to mind in that instant?.

Thanks for this @paradite_ , we've added a "view as markdown" button to our cookbooks!. Should make it much easier for metaprompting for e.g. "Update this prompt with this prompting guide below".

I've added a full demo and documentation to our open-source openai-realtime-agents repo, so you can try it out yourself and adapt to your own prompts. Please let me know what you think!.

Shouldn’t the realtime model handle this on its own?. The team is cooking and we’re excited to share new snapshots with you soon. However, as you continue pushing the frontier, it will *always* be useful to lean on a smarter model to make the hardest scenarios possible.

Benefits. Easy onramp: Just slot in your prompt and tools from your text agent. High intelligence: GPT-4.1 is excellent at IF, tool calling, and is very promptable. Lower cost: Realtime-mini is often fine for this pattern and is much cheaper (even with the supervisor calls).

Isn't this slow and annoying for the user?. No, because:.- The realtime agent can still handle many responses immediately.- The variations of "just a moment" are natural and contextual.- After "Just a moment", it only takes 1-3s longer for the supervisor to respond

Chat agent.Uses Realtime api, it maintains a natural conversation and handles basic tasks like collecting information from the user, and calls the supervisor for everything else. Supervisor agent.Text model (e.g. GPT-4.1), it handles tool calls and more challenging requests.

I've been building voice agents for the last 6mo and I think the chat-supervisor pattern is a game changer. Stitched model (STT-LLM-TTS) is slow, but realtime audio models aren't (yet) as smart as text. This has the best of both worlds. Here's how it works:

RT @menhguin: @elder_plinius