RT @sharifshameem: Blindly following your curiosity is deeply underrated, creating demos is how we excavate model capabilities, and you hav….

wrote about it here.

if you don't want your product to look like 1000 others out there, you need to care about every detail

here's a 6 second summary of the paper's result

when someone asks me next week why i have a bowl of rice on my mac

whats amazing is i just realized this scene is from a show that involves two superintelligent AIs fighting.

man leading the effort to pass the Great Filter (spacex, tesla) ends up being the catalyst (grok companions) for their failure.

me to claude code when i have 1% context until auto-compact.

all fun and games until its 2028 when you walk up to the location and there's sonnet 4.2.069 in a humanoid waiting for you.

i was curious how far the car drove itself, it was a 15-mile, 25min drive . on a related note, o3 blew my mind again today.

@OpenAIDevs opus has some thoughts on this

using claude code and claude tried to access openai docs using WebFetch and keeps getting a 403. can't tell if @OpenAIDevs - an AGI company, prevents access to AIs trying to access the site, or if this is specific to claude.

claude artifact apps? i wonder what they've got. *goes on to spend the next 11 minutes learning how to play happy birthday after figuring the sound of each key*

When SpaceX had an anomalous explosion during static fire prep in 2016, they went for weeks without an answer for why it happened - gathering every tiny fragment, running through 1000s of scenarios - to finally learn that it exploded because they moved too fast, quite literally.

#3 the barrier to entry in the field of memory is low!. this is personally what i find most exciting. so long as you're willing to experiment and dogfood constantly, you can make real impact in the field in a relatively short amount of time. forget about doing things the "right".

#2 bitter lesson for memory. gpt4 had an 8k token context window ~2 years ago. now we have a smarter, cheaper gemini-2.5-flash with 125x the context. paraphrasing what @sjwhitmore said, the complex architecture that was right 1 year ago today may not be needed today, instead

#1 stop trying to make perfect memory architecture happen, its not going to happen. the memory architecture that's right depends on the use case you're building for. the primitives that make @newcomputer work are very different from what works for @cursor_ai

really enjoyed the talks from @WHinthorn @sjwhitmore @yash_anysphere and @NicoleHedley3 - thanks for organizing it @GregKamradt and @LangChainAI! . here's 3 things i found very useful from what they had to share 🧵.

seeing how good claudecode is at searching for code relevant to the task at hand makes me realize that we need long context windows for models to be useful, but not for the reason many think we do. we don't need a model w/ 10m context window so that the model can ingest the.