@aidan_mclau pls fix for gpt5🙏.

OpenAI o3 problem: if initial searches don't confirm user claims, it locks into 'user is wrong' mode and defends that position relentlessly. Takes extraordinary persistence to break through its first (mis)interpretation. Should default to curiosity, not skepticism.

Claude Code native Windows support! Huge day for .NET developers

Kimi K2, specifically for agentic coding?! This might be another R1 moment!

So Grok 4 just feels similar to o3? What's supposed to be the selling point here? Did not watch the livestream, just tested it, am I missing what makes it special?.

"I don't see much difference between Opus & Sonnet.". True: single edits feel similar. But agentic coding isn't about single edits. It's about compounding accuracy. A "small" 3pp benchmark improvement means the difference between code that meets requirements, and hours spent

I don't even want a Claude Opus 4.1, all I want is Opus with a 500k context window.

what? guess SQL for production orders is too dangerous for the constitutional classifier? please don't shut down my claude account tho

wtf it also works with image generation!? the canvas is calling the imagen-3.0 api here, all inside the canvas!

Wait you can now integrate the Gemini api directly into the canvas by pressing this button?! gemini changed my example to generate new quotes using 2.0 Flash directly in the canvas using an api key provided by the gemini app! This opens so many new cool ideas!

If the voice mode does not include this setting, the chances that I will ever use it drop by 90%. Slow-talking voice modes are the most annoying part of AI apps. I'm looking at you, Google and Perplexity

Guide to build an AI slide generator that actually works:. 1. Claude Opus 4 - prompt it to generate static html + css, no javascript. each slide is its own div with a specific id. each slide div has exact css dimensions. 2. script that does static analysis to check if any text.

Can we please just get a setting for thinking token in Claude code? The think, think harder, ULTRATHINK is a fun idea but it gets old pretty fast when you have to type ULTRATHINK after each message to get the best performance. In fact, opus started ending its messages with.

Either ChatGPT has gotten way worse with the "it's not (only) x–it's (also) y" writing pattern in the past few months, or many more ppl are using unedited ChatGPT output for serious posts and articles. This is honestly turning me into an AI boomer, come on, at least edit the em.

Gemini Flash 2.5 just proved performance isn't the only thing following an exponential curve - billing is scaling beautifully too. Truly next-level AI innovation

Interesting to see, the SOTA model price was the lowest with the GPT-4o release. After that, getting the best model on the market has been getting more and more expensive with every release for all AI companies. The popular narrative seems to be that AI prices will fall over.

Claude Code is the best way to code right now. The way it manages the context is significantly better than Cursor; Cursor clutters the context window way too much, leading the model to forget about important stuff all the time. Claude Code can just do its thing; you need to hover.

In my headcanon o3 is a frat bro at a party, laptop open, freshly snorted Adderall, screaming over blasting music: "Whatever bro, no one's gonna check sources! Just write that shit confidently!".

Ppl still don't understand how model pricing works. No, you don't need to distill o3 to discount it by 80%; it's not a huge model anyways. They cut it because of strong competition. The same way new novel drugs have huge margins but lose them as soon as generics hit the market.

RAG will be here past AGI. Models will always search for info; you will always need a retrieval system for it. You can't put the full internet in context, even if we figure out infinite context linear compute attention.