Graham

NEW: Replit founding data scientist and engineer @giansegato: “I’ve begun to see product managers developing business financial models; designers writing commercial ads; barbershops building custom booking systems; and restaurant owners creating advanced pricing tools... These…

One of the most freeing realizations & gifts you can give yourself is realizing that most people have next to zero knowledge of who you are. Bosses, peers, investors, acquaintances, even parents and friends—most are reading “you” through a very narrow lens that is mostly…

Been test driving this and it's really solid! Much more polished than the old CLI experience. Although I would have kept using it just for the Oracle (o3) tool call capability if we're being honest.

criminally underrated talk from @devanshtandon_ Gemini powers YouTube's Large Recommender Model by **tokenizing every video on youtube** (SemanticID) - a vocabulary several OoMs larger than English, CONTINUOUSLY PRETRAINED every day. can reason across titles/descriptions, throws…

“Let's take a guitar as an example. Everyone knows what a guitar is, and everyone knows that if you put deliberate, intentional practice into it, you can become good at the guitar. Still, it takes time, effort and experimentation. In the circles around me, the people who are…

Optimize for energy

Funny how our instinct with LLMs is to argue with them when they go off track. We've learned the hard way that trying to 'fix' a derailed conversation is less effective than just using a message checkpoint to rewind and give a better prompt from the start. more thoughts 👇

it's funny how many people wrote up huge predictions for MCP without even looking into how LLM performance degrades when you add even 10 tools to them

been playing with o3 in @AmpCode for planning and anecdotally its been way better than sonnet. this is the adv of being model-agnostic. amp is on par with claude code, but can swap out providers when it makes sense

I genuinely hate the words I'm about to say... Prompting is a skill you can improve at over time, especially for agentic code use. More reps = better prompts = better code output.

The real unlock imo is that it’s scriptable/programmable and massively parallelizable in containers. It’s not a competitor to cursor. It’s complementary. Cursor is my primary IDE for human in the loop work and I dispatch teams of Claudes in the background.

+1 for "context engineering" over "prompt engineering". People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window…