Sabitlenmiş Tweet
shae wang
36 posts
shae wang
@shaeyuwang
head of core data @ripple built XRPL data on @dune building production-grade web3 x AI infra writing about web3, AI, and neuroplasticity on substack
San Francisco, CA Katılım Mart 2014
230 Takip Edilen805 Takipçiler
@GaryMarcus AI is not deterministic, (b) will never happen, build your own gates properly and actually study engineering if you want to vibe-code
English
This is totally wrong.
Blaming the user is missing the point that (a) coding agents have been overhyped and (b) can’t reliably obey the rules given to them in system prompts and other guardrails.
John A De Goes@jdegoes
Sorry, @lifeof_jer, but this is YOUR failure: 1. Your failure to demonstrate extreme ownership for AI generated code; instead, you abdicated your responsibility and blamed the AI. 2. Your failure to have an adequate and predictive mental model for how LLMs work. 1/2
English
@triathenum @jdegoes @lifeof_jer It's entirely on the vibers to understand the ceiling and risks here, disclaimers and cautionary tales are everywhere if they aren't too lazy to check and invest in learning the basics
English
@jdegoes @lifeof_jer I find the divide between the nontechnical vibe-coders and professional engineers interesting.
Engineers are calling out that the user failed at basic infrastructure while the vibers are saying the footgun should have had more safeties.
English
Sorry, @lifeof_jer, but this is YOUR failure:
1. Your failure to demonstrate extreme ownership for AI generated code; instead, you abdicated your responsibility and blamed the AI.
2. Your failure to have an adequate and predictive mental model for how LLMs work.
1/2
JER@lifeof_jer
English
shae wang retweetledi
I taught Claude to talk like a caveman to use 75% less tokens.
normal claude: ~180 tokens for a web search task
caveman claude: ~45 tokens for the same task
"I executed the web search tool" = 8 tokens
caveman version: "Tool work" = 2 tokens
every single grunt swap saves 6-10 tokens. across a FULL task that's 50-100 tokens saved
why does it work? caveman claude doesn't explain itself. it does its task first. gives the result. then stops.
no "I'd be happy to help you with that." no "Let me search the web for you" no more unnecessary filler words
"result. done. me stop."
50-75% burn reduction
with usage limits getting tighter every week this might be the most practical hack out there right now
English
I've been using Obsidian for a week. Lesson for beginners: if you're just starting with Obsidian, build first, organize later
There's no right or wrong structure, and you'll find your favorite one by spending time with it yourself
Andrej Karpathy@karpathy
Wow, this tweet went very viral! I wanted share a possibly slightly improved version of the tweet in an "idea file". The idea of the idea file is that in this era of LLM agents, there is less of a point/need of sharing the specific code/app, you just share the idea, then the other person's agent customizes & builds it for your specific needs. So here's the idea in a gist format: gist.github.com/karpathy/442a6… You can give this to your agent and it can build you your own LLM wiki and guide you on how to use it etc. It's intentionally kept a little bit abstract/vague because there are so many directions to take this in. And ofc, people can adjust the idea or contribute their own in the Discussion which is cool.
English
instead of detecting deepfakes, what if we redesign the trust/identity system so they don’t matter?
Modulate@modulate_ai
We built the world's best deepfake detection model, per @huggingface. Then priced it at $0.25/hour. The competition charges up to $150/hour for worse accuracy. Turns out "enterprise security" was just a pricing strategy. We blew it up. 🧵↓
English
Tighter scope -> lower variance -> better calibration -> better performance on these specialized tasks
Specialized agents are also inevitable economically, because cheaper compute + easier to finetune/debug
The orchestration layer does not need to be an LLM to work extremely well
Santiago@svpino
I really like the idea of having multiple specialized agents instead of a "general purpose" agent that tries to do it all. A few days ago, I read (sorry, I don't remember where) a study claiming that specialized agents, even when they are all using the same model, beat general agents by a mile. These guys are doing precisely that with an army of hyper-specific agents. And of course, they are following the ---Claw theme.
English
shae wang retweetledi
I’m really sorry to @FTX_Official users. I cannot sugarcoat the situation: your money is stuck and you won’t be getting it back soon.
NO ONE is going to acquire FTX. There are too many lawsuits and too many crimes performed on their books. Acquiring that liability won’t happen ,
English
shae wang retweetledi
1/6 SV irony🤦🏾♀️😆Picture this. You’re on CH, in Big Ideas After Party w/ @eriktorenberg @PatrickJBlum (rec) where the convo is on edu, credentials & corps. Q asked 'Do corps/schools heavily marketing diversity efforts actually do a disservice to those they’re trying to attract?
English