Загрузка видео...

Не удалось загрузить видео

На главную

Anthropic shipped 125 settings for Claude The official docs cover 40 One developer found the other 85 and his API bill dropped from $340 to $87 - not by using a cheaper model - not by writing shorter prompts just by moving one line in a config file to...

doublenickk's profile picture

Shadow Nick

12,682 subscribers

379,914 просмотров • 2 месяцев назад •via X (Twitter)

Наука и технологии
LIVE

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

Claude Code Head invented a word for what most config files aren't doing: tokenmaxxing. 90% of it lives in one place: ~/.claude/settings.json. 125 keys in v2.1.105. official docs cover 40. 4 of mine aren't anywhere in the docs. i audited the whole file - 18 settings actually decide what your month costs. > enabledPlugins - mine had 14 on, 4 actually used. each idle plugin loads hooks and SKILL.md at session start. > mcpServers.enabled - false instead of delete. 9 unused servers cost ~30K tokens of schema every session. > cache_control breakpoint at the static/dynamic boundary moved my Opus month from $340 to $87. tokenmaxxing isn't a vibe. it's a file with 125 keys most people never open. watch this tonight instead of the new Rick and Morty drop.
55:39

Claude Code Head invented a word for what most config files aren't doing: tokenmaxxing. 90% of it lives in one place: ~/.claude/settings.json. 125 keys in v2.1.105. official docs cover 40. 4 of mine aren't anywhere in the docs. i audited the whole file - 18 settings actually decide what your month costs. > enabledPlugins - mine had 14 on, 4 actually used. each idle plugin loads hooks and SKILL.md at session start. > mcpServers.enabled - false instead of delete. 9 unused servers cost ~30K tokens of schema every session. > cache_control breakpoint at the static/dynamic boundary moved my Opus month from $340 to $87. tokenmaxxing isn't a vibe. it's a file with 125 keys most people never open. watch this tonight instead of the new Rick and Morty drop.

Mnimiy

83,403 просмотров • 2 месяцев назад

20-YEAR-OLD GUY SHOWS HOW CLAUDE API CAN TURN 500-VIEW CONTENT INTO A $40K/MONTH CLIENT WORKFLOW Not from a big agency. Not from hiring writers. Just taking the same service people sell every week and wiring Claude API into the boring middle. A topic goes in. It pulls the research, angles, outline, draft, edits, and a clean file. That is normally $500-$1,500 of freelance work per client project. The screen says around $40K/month, but the real number is the time saved. A $3K project stops being 8 hours of rewriting and becomes 45 minutes of checking before delivery. Most people use Claude to write one answer. He is showing what happens when the whole process becomes the product.
1:18

20-YEAR-OLD GUY SHOWS HOW CLAUDE API CAN TURN 500-VIEW CONTENT INTO A $40K/MONTH CLIENT WORKFLOW Not from a big agency. Not from hiring writers. Just taking the same service people sell every week and wiring Claude API into the boring middle. A topic goes in. It pulls the research, angles, outline, draft, edits, and a clean file. That is normally $500-$1,500 of freelance work per client project. The screen says around $40K/month, but the real number is the time saved. A $3K project stops being 8 hours of rewriting and becomes 45 minutes of checking before delivery. Most people use Claude to write one answer. He is showing what happens when the whole process becomes the product.

Gipp 🦅

19,254 просмотров • 2 месяцев назад

Peter Steinberger Quietly Started A Shift That Makes Claude 35x Cheaper To Run The way Claude Code talks to tools matters more than people think. Most use MCP, and it quietly eats tokens, it loads everything into context every time, even tools you never call. One benchmark: MCP burned 35x more tokens than a CLI on the same task. Peter Steinberger, the OpenClaw guy, got annoyed by this and started building lean CLIs himself. That kicked off a tool called Printing Press. You point Claude Code at any website, even ones with no API like ESPN or Craigslist, and it builds a small command tool for it in about 10 minutes. In one demo, a request pulled 132,000 tokens of raw data, but the tool processed it locally and only handed Claude a 2,000-token summary. The rest never touched the context window. It also comes with ~50 ready-made tools you can grab right away. To start, point Claude Code at the links from and ask it to set it up. Bookmark this.
14:45

Peter Steinberger Quietly Started A Shift That Makes Claude 35x Cheaper To Run The way Claude Code talks to tools matters more than people think. Most use MCP, and it quietly eats tokens, it loads everything into context every time, even tools you never call. One benchmark: MCP burned 35x more tokens than a CLI on the same task. Peter Steinberger, the OpenClaw guy, got annoyed by this and started building lean CLIs himself. That kicked off a tool called Printing Press. You point Claude Code at any website, even ones with no API like ESPN or Craigslist, and it builds a small command tool for it in about 10 minutes. In one demo, a request pulled 132,000 tokens of raw data, but the tool processed it locally and only handed Claude a 2,000-token summary. The rest never touched the context window. It also comes with ~50 ready-made tools you can grab right away. To start, point Claude Code at the links from and ask it to set it up. Bookmark this.

Ridark

32,351 просмотров • 1 месяц назад

Anthropic just gave OpenClaw the green light Anthropic staff confirmed that CLI-based usage is sanctioned. We’ve successfully bridged the Claude Code CLI directly into the OpenClaw gateway, letting you run your entire agent fleet on your Pro/Max plan and no additional API tokens required. The Implementation: - Auth: Log in to Claude Code locally: claude auth login - Path: Grab your binary location: which claude - Bridge: Drop the path into your openclaw.json under agents.defaults: "cliBackends": { "claude-cli": { "args": [ "-p", "--output-format", "json" ], "command": " ", "input": "arg", "output": "json", "serialize": true, "sessionArg": "--session-id", "sessionMode": "existing" } }, "model": { "primary": "claude-cli/claude-opus-4-6" }, "models": { "claude-cli/claude-opus-4-6": {}, "claude-cli/claude-opus-4-7": {}, "claude-cli/claude-sonnet-4-6": {} }, The Result: OpenClaw reuses your local session IDs and subscription limits. You get the full power of Opus and Sonnet without the restricted OAuth endpoints or the API tax. They sanctioned the path. We built the bridge. Here is a video how I done it step by step:
1:09

Anthropic just gave OpenClaw the green light Anthropic staff confirmed that CLI-based usage is sanctioned. We’ve successfully bridged the Claude Code CLI directly into the OpenClaw gateway, letting you run your entire agent fleet on your Pro/Max plan and no additional API tokens required. The Implementation: - Auth: Log in to Claude Code locally: claude auth login - Path: Grab your binary location: which claude - Bridge: Drop the path into your openclaw.json under agents.defaults: "cliBackends": { "claude-cli": { "args": [ "-p", "--output-format", "json" ], "command": " ", "input": "arg", "output": "json", "serialize": true, "sessionArg": "--session-id", "sessionMode": "existing" } }, "model": { "primary": "claude-cli/claude-opus-4-6" }, "models": { "claude-cli/claude-opus-4-6": {}, "claude-cli/claude-opus-4-7": {}, "claude-cli/claude-sonnet-4-6": {} }, The Result: OpenClaw reuses your local session IDs and subscription limits. You get the full power of Opus and Sonnet without the restricted OAuth endpoints or the API tax. They sanctioned the path. We built the bridge. Here is a video how I done it step by step:

Ziwen

182,975 просмотров • 3 месяцев назад

FREE Fable 5, Opus 4.8, Sonnet 5, and GPT-5.6 in one IDE. Alibaba built it, 520 credits at signup - one Sonnet 5 message eats 4. It's called Accio. Web or Windows desktop, both free. Registration takes 30 seconds via Google or email. Inside: → Fable 5 for creative and long-context work. → Opus 4.6 through 4.8 - the full Anthropic frontier lineup. → Sonnet 5 for speed and reasoning at pennies per message. → GPT-5.5 for OpenAI native tasks. → Gemini 3.1 Pro and 3.5 Flash for multimodal. Real math: 520 credits divided by 4 = 130 Sonnet 5 messages in your first week. Cheaper models drop that even further. Anthropic charges $20/month for the same access with weekly limits. Custom MCP support built in. Profile > Settings > MCP. Every agent workflow you've built for Claude Desktop ports over instantly. Setup: → Go to accio(.)com → Sign up with Google → 520 credits land automatically → Pick your model and start firing prompts Bookmark this before Alibaba puts it behind a paywall.
0:45

FREE Fable 5, Opus 4.8, Sonnet 5, and GPT-5.6 in one IDE. Alibaba built it, 520 credits at signup - one Sonnet 5 message eats 4. It's called Accio. Web or Windows desktop, both free. Registration takes 30 seconds via Google or email. Inside: → Fable 5 for creative and long-context work. → Opus 4.6 through 4.8 - the full Anthropic frontier lineup. → Sonnet 5 for speed and reasoning at pennies per message. → GPT-5.5 for OpenAI native tasks. → Gemini 3.1 Pro and 3.5 Flash for multimodal. Real math: 520 credits divided by 4 = 130 Sonnet 5 messages in your first week. Cheaper models drop that even further. Anthropic charges $20/month for the same access with weekly limits. Custom MCP support built in. Profile > Settings > MCP. Every agent workflow you've built for Claude Desktop ports over instantly. Setup: → Go to accio(.)com → Sign up with Google → 520 credits land automatically → Pick your model and start firing prompts Bookmark this before Alibaba puts it behind a paywall.

Atenov int.

31,157 просмотров • 15 дней назад

THIS GUY TURNED ONE OBSIDIAN VAULT INTO A SHARED BRAIN FOR 5 CLAUDE AGENTS BUILT TO HANDLE 80% OF A SOLO DEVELOPER’S WORK. 00:03 he zooms out and hundreds of connected notes appear on screen. it is the perfect visual for Claude working from one shared project memory instead of starting from zero every session. most developers give Claude 1 file and 1 prompt. this setup gives it the architecture, coding rules, tests, client briefs, and previous decisions before it writes anything. obsidian becomes the memory layer, storing everything in markdown files Claude can read alongside the codebase. every agent works from the same context instead of repeating the same questions. then 5 specialists split the job. one researches, one plans, one codes, one reviews, and one deploys while the developer only controls the important decisions. one vault, 5 agents, 2 model tiers, and a $50 to $150 monthly stack can turn Claude from a chatbot into a small development team.
0:30

THIS GUY TURNED ONE OBSIDIAN VAULT INTO A SHARED BRAIN FOR 5 CLAUDE AGENTS BUILT TO HANDLE 80% OF A SOLO DEVELOPER’S WORK. 00:03 he zooms out and hundreds of connected notes appear on screen. it is the perfect visual for Claude working from one shared project memory instead of starting from zero every session. most developers give Claude 1 file and 1 prompt. this setup gives it the architecture, coding rules, tests, client briefs, and previous decisions before it writes anything. obsidian becomes the memory layer, storing everything in markdown files Claude can read alongside the codebase. every agent works from the same context instead of repeating the same questions. then 5 specialists split the job. one researches, one plans, one codes, one reviews, and one deploys while the developer only controls the important decisions. one vault, 5 agents, 2 model tiers, and a $50 to $150 monthly stack can turn Claude from a chatbot into a small development team.

Gipp 🦅

10,776 просмотров • 28 дней назад

10 repos that cut your ai agent token bill by up to 80% 1. microsoft/LLMLingua → cuts prompt size by up to 95% compresses prompts before the api call. 20x compression. published at EMNLP + ACL. near-zero quality loss. 6,100 stars 2. mem0ai/mem0 → replaces full conversation history in context stores what matters. retrieves only what's needed. 10,000 token history → 200 token memory. per agent. 54,800 stars 3. BerriAI/litellm → routes each call to the cheapest model simple task → haiku. complex task → sonnet. tracks cost per agent, per call, per day. 45,700 stars 4. run-llama/llama_index → replaces sending full documents rag: 100-page doc → 3 relevant chunks → same answer. 98% fewer tokens per query. 49,100 stars 5. chroma-core/chroma → replaces keyword search in full context vector store. finds the closest match. feeds only that. 50-200 tokens per query instead of thousands. 27,800 stars 6. letta-ai/letta → replaces infinite context window crashes paged memory for agents. loads only relevant memory. stops your agent from hitting limits and retrying. 22,400 stars 7. guidance-ai/guidance → cuts output token bloat by 30-50% structured generation. constrains model output natively. no more 100-token prompts to get json back. 21,400 stars 8. Aider-AI/aider → replaces pasting entire codebases builds a repo map. sends only files relevant to the task. not your whole project. just what the agent needs. 44,300 stars 9. openai/tiktoken → count tokens before you send know the exact cost before the api call happens. not after the bill arrives. 18,100 stars 10. simonw/ttok → hard cap on what gets sent cli tool: count tokens, truncate to budget limit. pipe any text in. get truncated output back. 389 stars most agents are expensive not because the model is expensive. because nobody checked what was being sent to it.
0:51

10 repos that cut your ai agent token bill by up to 80% 1. microsoft/LLMLingua → cuts prompt size by up to 95% compresses prompts before the api call. 20x compression. published at EMNLP + ACL. near-zero quality loss. 6,100 stars 2. mem0ai/mem0 → replaces full conversation history in context stores what matters. retrieves only what's needed. 10,000 token history → 200 token memory. per agent. 54,800 stars 3. BerriAI/litellm → routes each call to the cheapest model simple task → haiku. complex task → sonnet. tracks cost per agent, per call, per day. 45,700 stars 4. run-llama/llama_index → replaces sending full documents rag: 100-page doc → 3 relevant chunks → same answer. 98% fewer tokens per query. 49,100 stars 5. chroma-core/chroma → replaces keyword search in full context vector store. finds the closest match. feeds only that. 50-200 tokens per query instead of thousands. 27,800 stars 6. letta-ai/letta → replaces infinite context window crashes paged memory for agents. loads only relevant memory. stops your agent from hitting limits and retrying. 22,400 stars 7. guidance-ai/guidance → cuts output token bloat by 30-50% structured generation. constrains model output natively. no more 100-token prompts to get json back. 21,400 stars 8. Aider-AI/aider → replaces pasting entire codebases builds a repo map. sends only files relevant to the task. not your whole project. just what the agent needs. 44,300 stars 9. openai/tiktoken → count tokens before you send know the exact cost before the api call happens. not after the bill arrives. 18,100 stars 10. simonw/ttok → hard cap on what gets sent cli tool: count tokens, truncate to budget limit. pipe any text in. get truncated output back. 389 stars most agents are expensive not because the model is expensive. because nobody checked what was being sent to it.

self.dll

39,475 просмотров • 2 месяцев назад

SOMEONE IS CHARGING $5,000/MONTH FOR SOCIAL MEDIA CONTENT THAT CLAUDE WRITES IN 6 HOURS A WEEK no team, no office, just Claude and a scheduling app mid-market companies pay $2,000-$5,000 per month for social media management, and most of them are getting a junior freelancer posting generic content or an agency that barely touches their account one of them runs the whole thing in a weekly workflow: brief Claude on Monday, write posts Tuesday and Wednesday, schedule Thursday, pull numbers Friday another one charged three clients the same $5,000 rate before hiring a single person, that's $180,000/year running on one tool there's a prompt that makes Claude write every post in the client's exact voice by analyzing 5 examples of their existing content, 10 min per post instead of 60 you can land the first client with one outreach message, no portfolio, no case studies, just a specific offer and a 30-minute call the model is already working, the only question is whether you build it or watch someone else do it full breakdown of the workflow, the prompts, and how to find clients in the article below
0:42

SOMEONE IS CHARGING $5,000/MONTH FOR SOCIAL MEDIA CONTENT THAT CLAUDE WRITES IN 6 HOURS A WEEK no team, no office, just Claude and a scheduling app mid-market companies pay $2,000-$5,000 per month for social media management, and most of them are getting a junior freelancer posting generic content or an agency that barely touches their account one of them runs the whole thing in a weekly workflow: brief Claude on Monday, write posts Tuesday and Wednesday, schedule Thursday, pull numbers Friday another one charged three clients the same $5,000 rate before hiring a single person, that's $180,000/year running on one tool there's a prompt that makes Claude write every post in the client's exact voice by analyzing 5 examples of their existing content, 10 min per post instead of 60 you can land the first client with one outreach message, no portfolio, no case studies, just a specific offer and a 30-minute call the model is already working, the only question is whether you build it or watch someone else do it full breakdown of the workflow, the prompts, and how to find clients in the article below

Chrome

13,769 просмотров • 2 месяцев назад

This guy replaced an $8,000 survey crew with one drone and a pipeline on Claude Code that digitizes a whole site in a single trip. Inside it is not one drone but a whole pipeline of 5 modules on Claude Code each with its own job all answering to a single orchestrator. And he built them himself with no team. He draws one line on the map from the controller and the whole site starts turning into data at once. Pilot flies the DJI Matrice 350 RTK along the route and reaches where a crew drags tripods: the parking lot and the roofs and the grading. Scanner on the Zenmuse L2 records the geometry with a laser down to the centimeter. Builder in DJI Terra stitches the point cloud into a digital twin. Analyst on Claude builds what the client needs out of the finished model: a volume report and a progress diff and a tour behind a link. Mobile lives in his iPhone and hands the developer a link to the model while he drives to the next site. The site is ready as a file in about an hour and the developer rotates it in the browser himself and measures distances. No crew. No tripods. No week on site. Just him and a pickup and a drone and one API key. The whole processing pipeline lives in a folder at /Users/dev/site-capture. But he did not stop at a one time scan. The pipeline raises him by voice only when a flight misses a patch or when a sag on site goes past tolerance. And this is not made up: the DJI Matrice 350 RTK and the Zenmuse L2 are off the shelf enterprise hardware. The specs are one search away and the L2 really does write around 240,000 points per second. A crew for the same job charges $8,000 and three trips plus a week of desk work. His cost is tokens and subscriptions: one battery charge per flight and around $300 a month to host the finished models. In the end he draws one line on the map and an operator that does not exist scans the site and calculates the volumes and builds the report while he never leaves his truck. There is a huge market of everyone who needs to measure construction sites and warehouses and roofs and parking lots over and over while they still send out a photographer with a camera and wait a week. And he built this whole pipeline himself: one drone and one scanner and Claude Code that turns a flight into a file the developer pays for.
0:41

This guy replaced an $8,000 survey crew with one drone and a pipeline on Claude Code that digitizes a whole site in a single trip. Inside it is not one drone but a whole pipeline of 5 modules on Claude Code each with its own job all answering to a single orchestrator. And he built them himself with no team. He draws one line on the map from the controller and the whole site starts turning into data at once. Pilot flies the DJI Matrice 350 RTK along the route and reaches where a crew drags tripods: the parking lot and the roofs and the grading. Scanner on the Zenmuse L2 records the geometry with a laser down to the centimeter. Builder in DJI Terra stitches the point cloud into a digital twin. Analyst on Claude builds what the client needs out of the finished model: a volume report and a progress diff and a tour behind a link. Mobile lives in his iPhone and hands the developer a link to the model while he drives to the next site. The site is ready as a file in about an hour and the developer rotates it in the browser himself and measures distances. No crew. No tripods. No week on site. Just him and a pickup and a drone and one API key. The whole processing pipeline lives in a folder at /Users/dev/site-capture. But he did not stop at a one time scan. The pipeline raises him by voice only when a flight misses a patch or when a sag on site goes past tolerance. And this is not made up: the DJI Matrice 350 RTK and the Zenmuse L2 are off the shelf enterprise hardware. The specs are one search away and the L2 really does write around 240,000 points per second. A crew for the same job charges $8,000 and three trips plus a week of desk work. His cost is tokens and subscriptions: one battery charge per flight and around $300 a month to host the finished models. In the end he draws one line on the map and an operator that does not exist scans the site and calculates the volumes and builds the report while he never leaves his truck. There is a huge market of everyone who needs to measure construction sites and warehouses and roofs and parking lots over and over while they still send out a photographer with a camera and wait a week. And he built this whole pipeline himself: one drone and one scanner and Claude Code that turns a flight into a file the developer pays for.

Blaze

10,681 просмотров • 1 месяц назад

THIS GUY SPENT 5 MINUTES SETTING UP A $20 CLAUDE WORKSPACE AND TURNED 30 MINUTES OF DAILY PROMPT CHAOS INTO A SYSTEM Most people open Claude, type 1 messy prompt, rewrite it 10 times, get the same generic answer and call the model “bad.” He opens Claude Cowork, adds the project rules, past context, working files, examples, decisions and a clean place for the model to remember what it is doing. Same Claude. Same $20 tool. Completely different output. The funny part is there is no magic prompt here. No secret template. No “10x Claude” trick. Just the boring layer Karpathy keeps pointing at: context, memory and structure before output. A setup like this can save 30 minutes a day, which is 15 hours a month. If you use Claude for client work at $50 to $100/hour, that is $750 to $1,500/month in time you stop burning on repeated prompts. The model is real. The output is real. The edge is in not making Claude start from zero every single time.
0:37

THIS GUY SPENT 5 MINUTES SETTING UP A $20 CLAUDE WORKSPACE AND TURNED 30 MINUTES OF DAILY PROMPT CHAOS INTO A SYSTEM Most people open Claude, type 1 messy prompt, rewrite it 10 times, get the same generic answer and call the model “bad.” He opens Claude Cowork, adds the project rules, past context, working files, examples, decisions and a clean place for the model to remember what it is doing. Same Claude. Same $20 tool. Completely different output. The funny part is there is no magic prompt here. No secret template. No “10x Claude” trick. Just the boring layer Karpathy keeps pointing at: context, memory and structure before output. A setup like this can save 30 minutes a day, which is 15 hours a month. If you use Claude for client work at $50 to $100/hour, that is $750 to $1,500/month in time you stop burning on repeated prompts. The model is real. The output is real. The edge is in not making Claude start from zero every single time.

Gipp 🦅

14,871 просмотров • 2 месяцев назад

Everyone uploading PDFs to Claude is burning 70,000 tokens before asking a single question. A 20-page document costs between 1,500 to 3,000 tokens per page. Formatting. Broken tables. Images. Junk metadata. Claude processes all of it before you even type “summarize this.” There’s a free fix. It’s called Markitdown. Microsoft built it. 10,000+ GitHub stars. Almost nobody in the Claude community is talking about it. It converts any file PDFs, Word docs, Excel sheets, PowerPoints, YouTube videos into clean Markdown text. Token usage drops by up to 70%. Answers get better, not worse. Claude was trained on millions of Markdown documents. It reads the format natively. You’re not compressing quality you’re removing noise. There’s an MCP server too. Connect it to Claude Desktop once. Every file you upload auto-converts to .md before Claude touches it. You do nothing. It just works. Old way: upload PDF → Claude burns 70k tokens on junk → mediocre answer New way: Markitdown → clean .md → 70% fewer tokens → sharper answer The tool is free. The MCP setup takes 10 minutes. The token savings are permanent. You’ve been paying for Claude to read garbage. Stop.
0:54

Everyone uploading PDFs to Claude is burning 70,000 tokens before asking a single question. A 20-page document costs between 1,500 to 3,000 tokens per page. Formatting. Broken tables. Images. Junk metadata. Claude processes all of it before you even type “summarize this.” There’s a free fix. It’s called Markitdown. Microsoft built it. 10,000+ GitHub stars. Almost nobody in the Claude community is talking about it. It converts any file PDFs, Word docs, Excel sheets, PowerPoints, YouTube videos into clean Markdown text. Token usage drops by up to 70%. Answers get better, not worse. Claude was trained on millions of Markdown documents. It reads the format natively. You’re not compressing quality you’re removing noise. There’s an MCP server too. Connect it to Claude Desktop once. Every file you upload auto-converts to .md before Claude touches it. You do nothing. It just works. Old way: upload PDF → Claude burns 70k tokens on junk → mediocre answer New way: Markitdown → clean .md → 70% fewer tokens → sharper answer The tool is free. The MCP setup takes 10 minutes. The token savings are permanent. You’ve been paying for Claude to read garbage. Stop.

Veltrx

108,922 просмотров • 1 месяц назад

Anthropic pays its ML engineers $500K+/year to deploy what Percy Liang just taught for free in Lecture 3 of CS221. 80 minutes. one whiteboard. everything between "a neuron" and the model running in your IDE right now. three things click after this one: / why two Claude prompts that "mean the same thing" land in different regions of weight space / why fine-tuning sometimes makes a model dumber, not smarter / what RLHF actually changes when Anthropic aligns Claude same $90K Stanford sequence. lecture 3 of ~20. free. the gap between "using Claude" and "understanding it" is exactly one lecture wide.
1:12:27

Anthropic pays its ML engineers $500K+/year to deploy what Percy Liang just taught for free in Lecture 3 of CS221. 80 minutes. one whiteboard. everything between "a neuron" and the model running in your IDE right now. three things click after this one: / why two Claude prompts that "mean the same thing" land in different regions of weight space / why fine-tuning sometimes makes a model dumber, not smarter / what RLHF actually changes when Anthropic aligns Claude same $90K Stanford sequence. lecture 3 of ~20. free. the gap between "using Claude" and "understanding it" is exactly one lecture wide.

dunik

24,381 просмотров • 2 месяцев назад

andrej karpathy spent two hours teaching one thing: tokens are the atom of llms. tokenization is at the heart of every llm weirdness you've ever debugged. [watch the 15-min clip below. then run the 7-day playbook] ↓ save this before everyone copies it learn how the tokenizer works. understand how your llm actually consumes input. then run the engineering roadmap that took one production agent from $4,800/mo to $620/mo in 7 days. 87% reduction. no model swap. no framework migration. no quality drop on the eval set. token cost in 2026 is an engineering discipline. every line of your system prompt is rent you pay forever. what was eating the budget: → a single forgotten cron job ate 47% of one team's bill. they turned it off on a tuesday and the bill dropped before they wrote any optimization code. → anthropic ships a 90% discount on cache reads. one config line, cache_control ephemeral, break-even after one hit. most teams cache the volatile parts of the prompt and watch their hit rate sit at 12%. → one production agent went from 14,500 tokens of context overhead per turn to 850. a 94% drop. output quality held within 2% of the uncompressed baseline. → 60% of agent calls are haiku-tier work running on opus rates. classify the task first. pick the model second. → retry loops are the silent killer. no MAX_STEPS bound, one bad search query, $14 burned in a single session. one team traced 38% of their bill to this single pattern. karpathy gave you the atom. the playbook below gives you the harness. watch the lecture. read the playbook ↓
15:00

andrej karpathy spent two hours teaching one thing: tokens are the atom of llms. tokenization is at the heart of every llm weirdness you've ever debugged. [watch the 15-min clip below. then run the 7-day playbook] ↓ save this before everyone copies it learn how the tokenizer works. understand how your llm actually consumes input. then run the engineering roadmap that took one production agent from $4,800/mo to $620/mo in 7 days. 87% reduction. no model swap. no framework migration. no quality drop on the eval set. token cost in 2026 is an engineering discipline. every line of your system prompt is rent you pay forever. what was eating the budget: → a single forgotten cron job ate 47% of one team's bill. they turned it off on a tuesday and the bill dropped before they wrote any optimization code. → anthropic ships a 90% discount on cache reads. one config line, cache_control ephemeral, break-even after one hit. most teams cache the volatile parts of the prompt and watch their hit rate sit at 12%. → one production agent went from 14,500 tokens of context overhead per turn to 850. a 94% drop. output quality held within 2% of the uncompressed baseline. → 60% of agent calls are haiku-tier work running on opus rates. classify the task first. pick the model second. → retry loops are the silent killer. no MAX_STEPS bound, one bad search query, $14 burned in a single session. one team traced 38% of their bill to this single pattern. karpathy gave you the atom. the playbook below gives you the harness. watch the lecture. read the playbook ↓

Rohit

73,258 просмотров • 2 месяцев назад

NVIDIA just dropped free API keys for every top AI model You don't need your own GPU and you don't pay per token. GLM-5.2, MiniMax, Kimi, DeepSeek, OpenAI, all running on NVIDIA's servers, called through a normal API. Link: How to use one: 1. Create a free NVIDIA account. 2. Pick a Free Endpoint model and open its Build tab. You'll see ready-to-copy code with the base URL 3. Hit Generate API Key, copy it and paste that base URL and key into Claude Code, Cursor, or Cline. Bonus: NVIDIA also dropped 237 official skills that install into Claude Code and Codex in one command. Bookmark this.
0:24

NVIDIA just dropped free API keys for every top AI model You don't need your own GPU and you don't pay per token. GLM-5.2, MiniMax, Kimi, DeepSeek, OpenAI, all running on NVIDIA's servers, called through a normal API. Link: How to use one: 1. Create a free NVIDIA account. 2. Pick a Free Endpoint model and open its Build tab. You'll see ready-to-copy code with the base URL 3. Hit Generate API Key, copy it and paste that base URL and key into Claude Code, Cursor, or Cline. Bonus: NVIDIA also dropped 237 official skills that install into Claude Code and Codex in one command. Bookmark this.

Yarchi

62,979 просмотров • 24 дней назад

Someone ran Claude Code on a beach where any device overheats and that spot suddenly turned out to be the best home for the most powerful AI in the world. This is the reMarkable Paper Pro. A paper tablet for notes with no browser and no social media and not a single app. He sat down right on the sand in the open sun and brought up Claude Code on Opus 4.6 over the Claude API on the paper screen and opened his project ~/repos/webs while the waves broke a few steps away. For years every device had the same trouble outside. In direct sun the screen glares and washes out and heats up and instead of your work you see your own reflection. But e-ink does not blast its own light into your face. It reflects the sunlight like the page of a book. And here is what came out of it. The very thing that kills any normal screen outside turned into fuel for this one. The brighter the sun the sharper the picture because it has nothing to glare with and nothing to wash out. And then comes the thing no laptop on a beach will give you. Your eyes do not get tired. You can watch Opus think on max effort for an hour and it reads like a book in the sun and not a backlight you squint into. The picture only comes alive. In bright light it does not fade but turns sharper and higher in contrast than it ever was in a room. The charge lasts for days. E-ink barely touches the battery so there is no outlet anywhere on the sand and the tablet does not care. It weighs as much as a notebook. The whole setup folds into a beach bag like a pad with a pen on top. Everything on the screen is for real. Claude Code v2.1.110 and Opus 4.6 on the Claude API and the project ~/repos/webs open right on the e-ink in the middle of the sand. In my opinion this is the most unexpected home for an AI this year. Not an office with the blinds drawn and not a monitor cranked to full brightness but a quiet sheet of paper on the sand that open sun only makes better and on it the most powerful Claude writes code right on the page like a pen.
0:33

Someone ran Claude Code on a beach where any device overheats and that spot suddenly turned out to be the best home for the most powerful AI in the world. This is the reMarkable Paper Pro. A paper tablet for notes with no browser and no social media and not a single app. He sat down right on the sand in the open sun and brought up Claude Code on Opus 4.6 over the Claude API on the paper screen and opened his project ~/repos/webs while the waves broke a few steps away. For years every device had the same trouble outside. In direct sun the screen glares and washes out and heats up and instead of your work you see your own reflection. But e-ink does not blast its own light into your face. It reflects the sunlight like the page of a book. And here is what came out of it. The very thing that kills any normal screen outside turned into fuel for this one. The brighter the sun the sharper the picture because it has nothing to glare with and nothing to wash out. And then comes the thing no laptop on a beach will give you. Your eyes do not get tired. You can watch Opus think on max effort for an hour and it reads like a book in the sun and not a backlight you squint into. The picture only comes alive. In bright light it does not fade but turns sharper and higher in contrast than it ever was in a room. The charge lasts for days. E-ink barely touches the battery so there is no outlet anywhere on the sand and the tablet does not care. It weighs as much as a notebook. The whole setup folds into a beach bag like a pad with a pen on top. Everything on the screen is for real. Claude Code v2.1.110 and Opus 4.6 on the Claude API and the project ~/repos/webs open right on the e-ink in the middle of the sand. In my opinion this is the most unexpected home for an AI this year. Not an office with the blinds drawn and not a monitor cranked to full brightness but a quiet sheet of paper on the sand that open sun only makes better and on it the most powerful Claude writes code right on the page like a pen.

Blaze

89,297 просмотров • 1 месяц назад

this model runs coding agents for $0.30 per million tokens and someone just wired it into an entire company most people arguing about AI agents online have never watched one actually run a business types → seeded world → store → heartbeat → memory → cockpit → gate → command line → runtime nine files, one shared state, every screen just reads from it, nobody’s faking their own copy fail a policy check and the approve button turns into “override” instead, so nobody clicks past a red flag by accident that gate is the only reason an agent gets to touch money without someone standing over it bookmark this and watch it run, then read the article below
0:22

this model runs coding agents for $0.30 per million tokens and someone just wired it into an entire company most people arguing about AI agents online have never watched one actually run a business types → seeded world → store → heartbeat → memory → cockpit → gate → command line → runtime nine files, one shared state, every screen just reads from it, nobody’s faking their own copy fail a policy check and the approve button turns into “override” instead, so nobody clicks past a red flag by accident that gate is the only reason an agent gets to touch money without someone standing over it bookmark this and watch it run, then read the article below

leopardracer

28,047 просмотров • 5 дней назад

A 26-YEAR-OLD AI ENGINEER WHOSE BIRMINGHAM THESIS OUTPERFORMED GOOGLE SCHOLAR BY 50% JUST SHIPPED GRAPHIFY: ONE COMMAND TURNS ANY FOLDER INTO A CLAUDE CODE SECOND BRAIN Safi Shamsi built Graphify 48 hours after Karpathy posted his LLM wiki idea. It turns any folder, codebase, docs, PDFs, into a knowledge graph Claude reads instead of grepping. Up to 43x fewer tokens per query. The trick almost nobody is using yet: one flag exports the entire graph as a fully-linked Obsidian vault. Repo: /safishamsi/graphify Setup, end to end: 1. Install: uv tool install graphifyy (or pipx install graphifyy). Verify with graphify --version. 2. Install the skill in Claude Code: graphify claude install. This wires Graphify into Claude Code so you can call it as a skill. 3. Open the folder you want mapped in Claude Code. In the terminal: graphify . It extracts every concept and builds the graph in graphify-out/. 4. Export to Obsidian: graphify . --obsidian. Writes one note per concept, every relationship as a wikilink, every node linked back to its source. 5. Open the new vault in Obsidian (Manage vaults → Open folder as a vault), or drag it as a subfolder into your existing one. That's it. Your Claude Code instance now has a navigable map of the codebase that loads instantly instead of re-reading files every session. Full step-by-step build of the Claude + Obsidian second brain in the article below. Bookmark this
16:36

A 26-YEAR-OLD AI ENGINEER WHOSE BIRMINGHAM THESIS OUTPERFORMED GOOGLE SCHOLAR BY 50% JUST SHIPPED GRAPHIFY: ONE COMMAND TURNS ANY FOLDER INTO A CLAUDE CODE SECOND BRAIN Safi Shamsi built Graphify 48 hours after Karpathy posted his LLM wiki idea. It turns any folder, codebase, docs, PDFs, into a knowledge graph Claude reads instead of grepping. Up to 43x fewer tokens per query. The trick almost nobody is using yet: one flag exports the entire graph as a fully-linked Obsidian vault. Repo: /safishamsi/graphify Setup, end to end: 1. Install: uv tool install graphifyy (or pipx install graphifyy). Verify with graphify --version. 2. Install the skill in Claude Code: graphify claude install. This wires Graphify into Claude Code so you can call it as a skill. 3. Open the folder you want mapped in Claude Code. In the terminal: graphify . It extracts every concept and builds the graph in graphify-out/. 4. Export to Obsidian: graphify . --obsidian. Writes one note per concept, every relationship as a wikilink, every node linked back to its source. 5. Open the new vault in Obsidian (Manage vaults → Open folder as a vault), or drag it as a subfolder into your existing one. That's it. Your Claude Code instance now has a navigable map of the codebase that loads instantly instead of re-reading files every session. Full step-by-step build of the Claude + Obsidian second brain in the article below. Bookmark this

Yarchi

140,729 просмотров • 1 месяц назад

THIS GUY AUDITED 926 CLAUDE CODE SESSIONS AND FOUND MOST OF THE TOKEN WASTE WAS ON HIS SIDE everyone is blaming anthropic for the limits, so he decided to actually look at the data 858 sessions, 18,903 turns, and $1,619 estimated spend across 33 days here's what he found: 1\ one default setting was burning 14,000 tokens per turn Claude Code loads the full JSON schema for every tool into context at session start. whether you use them or not. 20,000 tokens of tool definitions sitting there on every single turn. the fix: one line in your settings.json "ENABLE_TOOL_SEARCH": "true" context dropped from 45K to 20K instantly. across 858 sessions that one setting was wasting an estimated 264 million tokens 2\ cache expiry is the single biggest waste 54% of his turns came after a 5+ minute idle gap. every one of those turns re-processed the entire conversation at full price which caused a 10x cost jump you go grab coffee. come back 5 minutes later. type your next message. everything rebuilds from scratch. the context didn't change. you didn't change. the cache just expired. 12.3 million tokens wasted on idle gaps alone 3\ 42 skills loaded. 19 of them used twice or less across 858 sessions. every one of those skill schemas sat in context on every turn eating tokens for nothing. 4\ 1,122 redundant file reads where the same file was read 3+ times one session read the same file 33 times. he ALSO built a full token auditor dashboard that shows you exactly where your waste is coming from 19 charts, opens in your browser, free AND open source
0:20

THIS GUY AUDITED 926 CLAUDE CODE SESSIONS AND FOUND MOST OF THE TOKEN WASTE WAS ON HIS SIDE everyone is blaming anthropic for the limits, so he decided to actually look at the data 858 sessions, 18,903 turns, and $1,619 estimated spend across 33 days here's what he found: 1\ one default setting was burning 14,000 tokens per turn Claude Code loads the full JSON schema for every tool into context at session start. whether you use them or not. 20,000 tokens of tool definitions sitting there on every single turn. the fix: one line in your settings.json "ENABLE_TOOL_SEARCH": "true" context dropped from 45K to 20K instantly. across 858 sessions that one setting was wasting an estimated 264 million tokens 2\ cache expiry is the single biggest waste 54% of his turns came after a 5+ minute idle gap. every one of those turns re-processed the entire conversation at full price which caused a 10x cost jump you go grab coffee. come back 5 minutes later. type your next message. everything rebuilds from scratch. the context didn't change. you didn't change. the cache just expired. 12.3 million tokens wasted on idle gaps alone 3\ 42 skills loaded. 19 of them used twice or less across 858 sessions. every one of those skill schemas sat in context on every turn eating tokens for nothing. 4\ 1,122 redundant file reads where the same file was read 3+ times one session read the same file 33 times. he ALSO built a full token auditor dashboard that shows you exactly where your waste is coming from 19 charts, opens in your browser, free AND open source

Om Patel

298,746 просмотров • 3 месяцев назад

Boris Cherny (Claude Code's creator): I deleted my IDE in November and hasn't opened it since. I writes more code than ever. his actual words on stage: I don't prompt Claude anymore. I have loops that are running. They're the ones prompting Claude. My job is to write loops. the progression he described: → a year ago: writing code by hand with autocomplete → then: running 5-10 Claudes in parallel, prompting each one → now: he doesn't prompt at all - he writes the loops that prompt Claude what it's already done inside Anthropic: • ~3x more code shipped per engineer (he says that number is already outdated and higher) • Cowork - a full product - built in ~8 days, 100% with Claude Code • designers, the finance guy, even the chief of staff now ship code his advice to founders: give everyone as many tokens as possible, and underfund every project on purpose - 2 engineers + a pile of tokens instead of 4. the roles are melting into one "builder." Watch this, than read the article below
28:49

Boris Cherny (Claude Code's creator): I deleted my IDE in November and hasn't opened it since. I writes more code than ever. his actual words on stage: I don't prompt Claude anymore. I have loops that are running. They're the ones prompting Claude. My job is to write loops. the progression he described: → a year ago: writing code by hand with autocomplete → then: running 5-10 Claudes in parallel, prompting each one → now: he doesn't prompt at all - he writes the loops that prompt Claude what it's already done inside Anthropic: • ~3x more code shipped per engineer (he says that number is already outdated and higher) • Cowork - a full product - built in ~8 days, 100% with Claude Code • designers, the finance guy, even the chief of staff now ship code his advice to founders: give everyone as many tokens as possible, and underfund every project on purpose - 2 engineers + a pile of tokens instead of 4. the roles are melting into one "builder." Watch this, than read the article below

Morty

136,622 просмотров • 1 месяц назад

THIS DEVELOPER CONNECTED 8 NVIDIA DGX SPARKS INTO ONE CLUSTER - AND RAN AN 800GB MODEL THAT MADE HIM 10X MORE PRODUCTIVE 21:47 he says it straight - "this is a terabyte of VRAM - we ran Quen 3.5, 800GB on disk, a model that doesn't even fit on a single Mac Studio - 24 tokens per second - I'd say that's a win" 8 Sparks connected through a $1,300 switch via RDMA over Ethernet - each node adding 128GB of memory into one unified pool of 1TB started with one Spark at 3 tokens per second - every added node doubled the speed - and eight together deliver 24 tokens on a model that physically cannot run anywhere else Kimi K2 at 600GB loaded in 15 minutes, 115GB per node, 13 tokens per second - a model that simply cannot run on anything smaller Claude helped configure the entire cluster - SSH mesh across all 8 machines, network config, jumbo frames, QSFP port speeds - all from one terminal most people rent cloud compute for models this size at $2,000+/month - he built the cluster once and now every token costs 20x less
24:17

THIS DEVELOPER CONNECTED 8 NVIDIA DGX SPARKS INTO ONE CLUSTER - AND RAN AN 800GB MODEL THAT MADE HIM 10X MORE PRODUCTIVE 21:47 he says it straight - "this is a terabyte of VRAM - we ran Quen 3.5, 800GB on disk, a model that doesn't even fit on a single Mac Studio - 24 tokens per second - I'd say that's a win" 8 Sparks connected through a $1,300 switch via RDMA over Ethernet - each node adding 128GB of memory into one unified pool of 1TB started with one Spark at 3 tokens per second - every added node doubled the speed - and eight together deliver 24 tokens on a model that physically cannot run anywhere else Kimi K2 at 600GB loaded in 15 minutes, 115GB per node, 13 tokens per second - a model that simply cannot run on anything smaller Claude helped configure the entire cluster - SSH mesh across all 8 machines, network config, jumbo frames, QSFP port speeds - all from one terminal most people rent cloud compute for models this size at $2,000+/month - he built the cluster once and now every token costs 20x less

Noisy

185,148 просмотров • 1 месяц назад