Damián Catanzaro ☕️

Damián Catanzaro ☕️

94,445 Twitter followers

Follow

☕️ Founder and CTO of @cafecito_app 🚀 CTO of @fundit_protocol ✉️ Los DMs se mandan siempre con pregunta incluida.

Posts

Damián Catanzaro ☕️

Chatting with friends who aren't developers and have started tinkering with AI, I realized why they spend so many tokens and burn through the limit so quickly: - They use everything in the same chat, so the chat keeps growing and growing, and you're sending it more and more context. When planning a project, you have to break it down into smaller functions. Each function will be a development in itself. For example, in OpenCode, you use `/new` to start a new session. OpenCode itself is smart enough to know where to look in the current directory. I saw some people using Cursor and spending 900k to 3M tokens per request. It's insane! My context windows didn't exceed 150k tokens across the entire chat. Always think of each chat as a feature, not as a project. Not only will you have more tokens for development, but the model will become increasingly unreliable as the number of tokens increases beyond a certain threshold.

Damián Catanzaro ☕️

I stayed up programming until 4 AM and woke up at 10 AM to continue. My plan for this long weekend was to rest, but your enthusiasm was stronger. Here's a preview of the world and how it's coming along. twitter.com/DamianCatanzaro/st...

Damián Catanzaro ☕️

This is unbelievable, I've read it 20 times already and I still can't believe it. Someone swapped $50 million USDT to buy the AAVE token. When you want to make large trades and there isn't enough liquidity, a message appears saying you accept it. The person received $32,000 for the $50 million. I still can't believe it.

Earlier today, a user attempted to buy AAVE using $50M USDT through the Aave interface. Given the unusually large size of the single order, the Aave interface, like most trading interfaces, warned the user about extraordinary slippage and required confirmation via a checkbox.

Damián Catanzaro ☕️

How close are we to OpenAI having datacenters of human brain cells running GPT?

🚨: A petri dish of human brain cells just learned to play DOOM

Damián Catanzaro ☕️

I love it! For those who want the specs, it's the Xiaomi YU7 Max: 760km range, 0-100 km/h in 3.23 seconds, 690HP twitter.com/DamianCatanzaro/st...

Damián Catanzaro ☕️

Who would have thought you have to read the AI code before pushing it to production? This is what I've been saying and explaining continuously: LLMs aren't deterministic; they can have hallucinations, and if just one hallucination passes, it can destroy your business.

Guillermo Rauch

A Vercel user reported an issue that sounded extremely scary. An unknown GitHub OSS codebase being deployed to their team. We, of course, took the report extremely seriously and began an investigation. Security and infra engineering engaged. Turns out Opus 4.6 *hallucinated a

Damián Catanzaro ☕️

We’ve identified industrial-scale distillation attacks on our models by DeepSeek, Moonshot AI, and MiniMax. These labs created over 24,000 fraudulent accounts and generated over 16 million exchanges with Claude, extracting its capabilities to train and improve their own models.

Damián Catanzaro ☕️

I don't understand Anthropic's decision. They released Claude Opus 4.6 Fast, which is 2.5x faster than the regular Opus 4.6, but at 6x the price. I don't understand why anyone would use it. $150 USD for a million tokens for 2.5x speed seems like a lot. twitter.com/DamianCatanzaro/st...

Damián Catanzaro ☕️

Maybe I'm missing something, but if they're saying multiple bots are organizing to conquer the world through a forum that prevents you from running multiple agents on your machine and letting them communicate with each other to organize and conquer the world? twitter.com/DamianCatanzaro/st...

Damián Catanzaro ☕️

I paid for a year of Z AI to use GLM 4.7 directly from their subscription, $28 USD annually. Obviously, it's not a Claude Opus, but it can be used with multiple agents in parallel and you can see how far the model's power goes by giving it a good context with OpenCode.

Loading..