Mastering AI coding, tips

@ChrisGPT not even close to call of duty. the hand is fugly. not the worst kind of slop.

As I’ve been saying for the past 6 months the end of indie gaming studios is nigh. Soon it will be in the people’s hands. This is why I love Matt as well, he doesn’t produce slop. He pushes AI models to the max and makes all other demos look like absolute slop.

@mattshumer_ the hand is ugly

Claude Opus 5 one-shotted this game. EVERYTHING you see in this demo is custom code... not a single external asset was used. AI games are going to be amazing. (sound on)

@abhinavflac Max effort?

opus 5 is something else man vest 3d voxel generation I have ever seen from any model so far.

How should we educate and warn people of the risks of using long context (400k-1M+ tokens in a single thread)? And should we make it opt-in before you can use it? Context: We want to add (back) a `large` mode to Amp with 1M tokens of context, using Sonnet 4.5 for now. This `large` mode would guarantee 1M tokens even when `smart` mode has less, such as now with Opus 4.5. Long context is helpful in some rare cases, such as a large refactor with related and dependent steps. But it can get very expensive, like $100-1000+ per thread (model pricing, not our fault). And if it's used for a bunch of different topics throughout the day, like a stream-of-consciousness chat, then quality goes way down. We want Amp to be opinionated about the right way to use it. We feel a responsibility to educate and warn people of these risks before they use `large` mode. But I fear some people will use `large` mode, ignore our warnings, and then have a bad and costly experience. I want to figure out how we can reach those people first. We already have ampcode.com/200k-tokens-is… and ampcode.com/guides/context…, we were early to replace compaction with something better (ampcode.com/news/handoff), and we often repeat our recommendation to use small threads. But the world is big, and we need to repeat this recommendation a lot and build it into our product to reach everyone. So: what else do you all think we can do to educate and warn people about long context usage, and nudge them to use it the right way? (BTW, I think long context up to ~350k is great. It's 350k-1M+ tokens that I am nervous about.)

@Motion_Viz I'm still waiting for your full prompt library. Can you DM me now?

@Motion_Viz GEMINI

Spent 50+ hours mastering Gemini 3.0 prompts. Here's what most people get wrong: They write prompts like requests. I write prompts like technical specs. The difference: ❌ "Make this icon look like glass" ✅ "Transform into translucent glass with glossy reflections, realistic refraction, liquid-glass material, iridescent edges in electric red/orange/gold, intense highlights, visible depth and curvature. Camera: Canon EOS R5, 85mm f/1.2, f/2.8." Specificity isn't optional. It's the entire game. Like + reply "GEMINI" for my full prompt library. (must be following for DM)

@Thom_Wolf Putnam is more knowledge-based whereas IMO requires more creativity and more time per problem, so Putnam is easier for LLMs.

OpenAI added the codex-cconger-test model to the API

😭

silentnova in LMArena is a Google model

OpenAI testing img-1p-t-edit5-ev3 model in API

I accidentally paid for 2 Gemini Ultra accounts. I have 1 month of Gemini Ultra (can let you log into my account) DM me (obviously not free but I paid $250/m. Will give access for $50-75 depending how well I know you 🙏🏼

The model is too good, the harness is too good. Anthropic isn't done winning. OpenAI GPT-5.2 better have another big jump

Gemini 3 Deep Think is about to solve the hardest Olympiad math problem: constructing the floor pattern at Sunshine Airport. Let's fucking go! Don't die. Prompt: pastebin.com/Vy5YWfsp

@chrisuniverse @GeminiApp All AI models failed to solve IMO 2025 Problem 6 pastebin.com/Vy5YWfsp Gemini 3 Pro got it wrong (answered 2048). Thanks!

Finally got @GeminiApp Ultra! Send me your toughest prompts ever, and I will reply with a link to the response. I am super hype about this!

people act like they already have access to gemini 3. as an A/B tester, share something new we don't know.

Gemini 3 (Lithiumflow) Checkpoint on KingBench: The new Gemini 3 checkpoint on LMArena scores higher than ECPT on KingBench. The older X58 and 2HT checkpoints are still better, but this one isn't bad either. Costs are based on $2.5/m & $15/m Generations are in the thread.

So next week we get Cursor’s first own model, Cheetah, personally more excited about Gemini 3 tho

👨‍💻Prompt engineering with Gemini 3👨‍💻 You can get much more out of Gemini 3 if you can make the model output more code. The SVGs on the left are the lazy ones; the right ones use prompt engineering to make the model output more code. Top = Pelican SVG Bottom = Robot SVG

@diabrowser linux wen? been waiting patiently for so long

Less invites, more ships. Dia is now open to everyone on macOS. Download now: 🔗 in bio