Sabitlenmiş Tweet
0xDipper
1K posts
0xDipper
@Dipper_pol
Researcher AI x finance. builders and legends, in their own words @zscdao
United States Katılım Şubat 2022
191 Takip Edilen5.7K Takipçiler
Outtake built a long-running agent on Claude that investigates cyber attacks for up to 2 hours unattended. their lead engineer gave away the two hardest lessons:
"prompts are suggestions."
"as this agent runs longer, every single word in that prompt eventually will probably be ignored."
the fix: anything the agent must do every time gets pulled OUT of the prompt and built into the harness - so it never has to think about it, and that attention goes where it can actually improvise.
second lesson, before you write any code:
"you have to be the agent before you can build the agent."
they ran the investigations by hand first to learn what good looks like.
and the part nobody expects - a cyber investigator needed to be half coding agent:
file system + bash → it writes its own tools mid-investigation → routes around broken APIs → finishes the job anyway
Anthropic's Applied AI take: the agent loop has converged to something simple. the power is now entirely in what you give the model - memory, tools, skills - then you get out of the way.
median run: 16 minutes. longest: about 2 hours.
Watch it today ↓
0xDipper@Dipper_pol
Andrej Karpathy builds GPT from scratch in one 2-hour video - before he led pretraining at Anthropic: he starts with an empty file and tiny Shakespeare, and ends generating infinite fake Shakespeare, character by character - the exact same way ChatGPT runs, just token by token. "now we get to the crux of self-attention - this is probably the most important part of this video." then he explains the whole engine: every token emits a query (what am I looking for) and a key (what do I contain), and attention is just them finding each other. by the end he connects it to the real thing: "to train ChatGPT there are roughly two stages: pre-training, then fine-tuning." this one video is the clearest path from "I use ChatGPT" to "I understand how it's built." Watch it today ↓
English
My friend applied to 200 tech jobs in two years. No MIT. No Stanford.
Last month Anthropic offered him $750,000.
I asked him how he broke in from zero.
He sent me the exact video that got him in. A 4-hour course on mastering Claude Code.
I watched it last night.
Halfway through, I realized I've been using Claude Code completely wrong for a year.
Bookmark this and read the article below.
• 00:00 - Claude Code setup
• 49:34 - building apps with Claude Code
• 2:07:52 - prompting Claude Code
• 2:46:16 - Claude Code for production
Avid@Av1dlive
English
Andrej Karpathy builds GPT from scratch in one 2-hour video - before he led pretraining at Anthropic:
he starts with an empty file and tiny Shakespeare, and ends generating infinite fake Shakespeare, character by character - the exact same way ChatGPT runs, just token by token.
"now we get to the crux of self-attention - this is probably the most important part of this video."
then he explains the whole engine: every token emits a query (what am I looking for) and a key (what do I contain), and attention is just them finding each other.
by the end he connects it to the real thing:
"to train ChatGPT there are roughly two stages: pre-training, then fine-tuning."
this one video is the clearest path from "I use ChatGPT" to "I understand how it's built."
Watch it today ↓
Clodex@0xClodex
English
Andrej Karpathy just reveales how LLMs actually thinks:
"GPT-4 knows it failed. It just won't tell you unless you ask."
>80% of GPT-4 errors are recoverable - the model already knows it screwed up.
It has 80 transformer layers and spends the SAME compute on every single token as your brain
In a 20-minute speach at Microsoft Build, Karpathy reveals the full psychology of LLMs.
Worth more than any $500 prompting course you've seen on your timeline.
Ricker@0xRicker
Ex-Google AI Agent Architect just launched a full course on " Harness, Loop, AI Agent Memory system" 90% AI agents are wasting 60%+ of their tokens wrong 00:00 - Ai Agent Memory 12:06 - Harness, Loop 31:18 - Ai Agent System This 50-minute will replace 10 paid courses and guides on Harness and Looping Watch it today, then read how to build proper agent loops in the article below
English
the math nobody wants you to see:
AI visibility agency: $10K/month, 3 month contract, reports every Friday
CrowdReply AI Agent: you literally just ask it "why am I invisible in ChatGPT" and it goes and fixes it
same outcome. one is 30x the price.
this is why agencies are scared of this launch.
CrowdReply@Crowdreply_io
Today we're introducing Claude for Marketing. The first AI agent that doesn't just track your brand in AI answers but ranks it all from your chat. Just ask. It reads your visibility, pinpoints what's wrong and handles the fix.
English
A HOMELAB BUILDER JUST PAIRED AN AMD EPYC 77 CHIP WITH 512GB OF DDR5 AND LLAMA.CPP TO RUN DEEPSEEK V3 ENTIRELY FROM RAM WITHOUT A SINGLE GPU
00:03 "my very own AI super computer, not exactly, but it should be quite capable. the goal is to run large LLMs entirely from RAM with a decent context window"
the stack is minimalist. one AMD EPYC 77 series processor, 8 memory channels, 512GB of DDR5 ECC, llama.cpp compiled for CPU inference. no GPU. no CUDA. no waiting for a used 3090 to appear on eBay
running DeepSeek V3 and Llama 4 from system memory means inference happens at CPU speeds, slower per token but unlocks model sizes that would need a $40,000 8xH100 server to fit on GPU. 405B parameters loaded straight into RAM
the article ranks local AI compute from $180 Tesla P40 up to $4,199 Mac Studio. this build sits in a different lane entirely. not GPU acceleration, but raw memory capacity. RAM is the new frontier for anything above 200B parameters
most people wait out the GPU shortage. he built an inference machine that never needed a GPU in the first place
save this before every homelab realizes their server chassis was already an LLM machine waiting for enough RAM ↓
beamnxw ./@beamnxw
English
A free 1-billion-parameter model on an RTX 5080 just made every $20 chatbot subscription look slow.
The screen recording shows one PowerShell window. One command: ollama run llama3.2:1b. One request: JavaScript code that reads a file.
The answer finishes in 987 milliseconds. Then the logs print the part worth recording.
Prompt swallowed at 9,663.51 tokens per second. Response streaming at 267.61 tokens per second. Working code, 2 approaches, formatted, done before a cloud tab connects.
Every question after that: same speed, $0, nothing sent anywhere. No spinner. No rate limit. No usage meter draining in a billing dashboard.
The model is the smallest in its family, tuned for exactly the everyday questions people burn paid credits on 30 times a day. The same GPU holds models 20 times bigger.
The setup is 1 free install and 1 command. The whole demo fits in 20 seconds because there's nothing to wait for.
Somewhere a server farm processed nothing and billed nothing.
The fastest AI answer you've ever seen costs exactly $0 per token.
Moysei@0xMoysei
English
Boris Cherny, the engineer behind Claude Code, and Anthropic CEO Dario Amodei, in one interview:
Boris: "for me personally, Claude has been writing 100% of my code for at least six months. I feel like I suddenly have superpowers - I have a jetpack, and engineering has never been this fun."
he built Claude Code and Claude Cowork after leaving a slow life making miso in rural Japan - one chatbot took his breath away and pulled him back in.
then Dario on where this goes:
"you automate 90% of the job, people are 10x more productive in the other 10%. but eventually it gets close to 100% - and then you have to find something else for them to do."
the man building the tool that writes the code, and the man warning about what happens when it writes all of it.
Watch it today ↓
Clodex@0xClodex
English
Andrew Ng x Anthropic built an agent that runs your computer on its own - from scratch, in a 1-hour course:
→ 0:00 - what computer use actually is: Claude reads the screen, clicks and types
→ 3:49 - multimodal: feed Claude images and stream the responses back
→ 16:06 - prompting that separates a toy demo from a production agent
→ 33:22 - prompt caching: cut your API bill up to 90% and latency up to 85%
→ 46:09 - tool use: give Claude real functions to call, not just text
free. and more useful than most $500 agent courses.
bookmark and watch ↓
sunick@Serantych
English
Andrew Ng stood on NVIDIA's stage and waved off the fear of dangerous AI with one line:
"I don't work on preventing AI from turning evil for the same reason that I don't work on the problem of overpopulation on the planet Mars."
he called the whole worry "an unnecessary distraction."
one of the researchers who'd worked under Ng at Baidu was Dario Amodei. a year later, on a Google Brain stage, Dario answered his former boss directly:
"it may not be worth worrying about overpopulation on Mars today, but we can and should study overpopulation on Earth. and if we do that right, a lot of what we learn may someday apply to Mars."
Ng saw a distraction. Dario saw the work.
five years later, Dario founded Anthropic ↓
Clodex@0xClodex
English
Paul Tudor Jones went to zero three times before he turned 25. his father begged him to quit. he ignored him and became one of the great traders alive.
$10,000 to nothing. then $20,000 to nothing. again.
"I had two phenomenal learning periods where I lost everything"
his father, on attempt three: "you've got to get in something safe like real estate." he said no.
he even got into Harvard Business School. loaded the U-Haul, then turned it around in the driveway: "they're not going to teach me anything about this."
but the sharpest thing he says isn't about his wins. it's what he watched happen to everyone like him.
100,000 floor traders. $5-10 billion in profits spread across them. then electronic trading arrived:
"a mass extinction of entire profession"
and the part that should stop you: "it happened so slowly and incrementally that no one ever saw it." the profits didn't vanish. three or four quant firms compressed 100,000 people's income down to about 1,000.
his words for it: the precursor of what's happening today.
~45 min, free. the trader who went bust three times on how an entire profession quietly disappeared ↓
0xDipper@Dipper_pol
Everyone calls it "Soros breaking the Bank of England." it was Stanley Druckenmiller's trade. and Soros's only contribution was telling him to bet more. 1992. Druckenmiller sees it: Britain needs low rates, Germany needs high ones, and the two currencies are chained together at a fixed rate. something has to break. he sizes the short at $5 billion - already enormous. he walks into Soros. Soros isn't impressed: "why are you only doing $5 billion?" "this is a one-way bet," Soros tells him. do $15 billion. the error wasn't being too aggressive. it was being too timid on a trade you're certain of. they pressed. the pound broke. the fund made 62% that year. his whole philosophy in one line: "diversification is really overrated" find one thesis you've truly done the work on, put your eggs in it, and watch the basket closely. and when it finally works, don't grab the profit just to feel good: "you can't just make yourself feel good by taking a profit" ~1hr, free. the trader behind the pound on why the real risk is betting too small ↓
English
Everyone calls it "Soros breaking the Bank of England." it was Stanley Druckenmiller's trade. and Soros's only contribution was telling him to bet more.
1992. Druckenmiller sees it: Britain needs low rates, Germany needs high ones, and the two currencies are chained together at a fixed rate. something has to break. he sizes the short at $5 billion - already enormous.
he walks into Soros. Soros isn't impressed:
"why are you only doing $5 billion?"
"this is a one-way bet," Soros tells him. do $15 billion. the error wasn't being too aggressive. it was being too timid on a trade you're certain of.
they pressed. the pound broke. the fund made 62% that year.
his whole philosophy in one line:
"diversification is really overrated"
find one thesis you've truly done the work on, put your eggs in it, and watch the basket closely. and when it finally works, don't grab the profit just to feel good:
"you can't just make yourself feel good by taking a profit"
~1hr, free. the trader behind the pound on why the real risk is betting too small ↓
0xDipper@Dipper_pol
Nassim Taleb saw 2008 coming. his warning sign wasn't chaos - it was calm. everyone watched markets go quiet before the crash and called it stability. Taleb saw the opposite: risk quietly piling up where nothing moves. "it sinks but doesn't fluctuate" his proof: 87% of what we call "fat tails" traces to a single day - the 1987 crash. the whole risk lives in the rare event nobody models. "we don't know how little we know about the rare event" so he never tried to predict the crash. he built to survive it, then profit. the tail-hedge fund he advises returned up to 115% in the 2008 crash. "winning is surviving" and the flip side people miss: done right, you don't just dodge the blowup - you get "more robust and more open to upside." lose small, win huge. ~19 min, free. the man who called 2008 on why the calmest markets hide the most risk ↓
English