silentguy

I TRIED RUNNING MY BUSINESS WITH CLAUDE COWORK FOR A WEEK went from 8 hours of emails, decks, formatting, research down to 2 claude handled the other 6 without me touching anything the anthropic pm showed exactly this workflow in their demo - sql queries, data analysis, test cases generated in seconds his words: "if i had to do this on my own, it would take me hours, if not more" one person + claude = team of five busywork solved. strategy left. what's the first task you'd automate with an ai coworker?

WISH I SAW THIS BEFORE SPENDING $4K ON THE WRONG CARD bought a 3090 because "more VRAM" that's what everyone said in every thread, every discord, every build guide ran the same 70B model on 4080 super it's inference was 2.3x faster. same quantization. same prompt. same llama.cpp build the difference? memory bandwidth 4090 has 1.01 TB/s. 3090 has 936 GB/s. doesn't sound like much but on autoregressive inference where every single token reads the entire model weights from VRAM, that gap compounds on every forward pass bandwidth > VRAM for inference speed. vram just determines what fits, bandwidth determines how fast it actually runs what GPU are you running local llms on and what's your actual tok/s?

@slash1sol the real cost isn't subs, it's maintaining 50 foss tools

When I first saw this I didn't believe it either: > Replacing Netflix for free? > Replacing Notion for free? > Replacing Games for free? > $1,000+/year saved forever? -- Then I opened GitHub. -- Then I installed 50 tools. -- Then I stopped paying rent on software list in the article ↓

THE REAL STORY BEHIND GPT-5.6 not that it's smarter. Every model drop says that the interesting part is the tiers. Sol, terra, luna openai finally admitted one model doesn't fit every problem you don't need frontier reasoning for a to-do list you don't need speed when you're debugging a codebase tiered intelligence for tiered problems the announcement calls it "scales with your ambition" i call it "finally, common sense" which tier are you actually using most?

THE MISSING LAYER IN EVERY AI WORKFLOW everyone obsesses over the perfect prompt nobody builds the context layer karpathy explained this years ago prompts are programs in english but a one-line program never works on first try the difference between generic and gold output: - your brand voice in a file - examples of what good looks like - constraints before instructions the 5% figured this out the 95% keep blaming the model when was the last time you changed your prompt vs your context?

EVERY AI AGENT YOU'VE EVER USED HAS AMNESIA you spend 3 hours teaching it your codebase, your style, your preferences it nails the task. you close the tab next session: starts from zero. asks the same questions. makes the same mistakes. doesn't remember you exist this isn't intelligence. this is a stateless function with a really good text generator attached most ai products are still on step 0.5 they charge $200/mo for a model that forgets everything about you the second you close the browser fable 5, mem0, the memory bank approach. These are attempts at step two but the real question is why step two wasn't built into the architecture from day one.

@0xCodila most people don't restructure prompts so it's ggs for tokens

Anthropic and Andrew Ng built an agent that uses 90% fewer tokens from scratch: they dropped the entire book of Frankenstein into a prompt - 108,000 tokens asked one question the input dropped from 108,000 tokens to 11 here's how: step 1 → put everything that never changes at the top - tools, then system, then docs step 2 → mark where the static part ends. everything above it gets cached step 3 → one stray space breaks it - and you pay full price again step 4 → the cache dies in 5 min - every read resets the clock step 5 → cached tokens don't count against your rate limits. free headroom most people never touch this - it pays for itself on day one watch & bookmark - this 1-hour brilliant course ↓

THE AI COLD WAR HAS A FREE TIER everyone debates claude vs chatgpt vs deepseek meanwhile chinese labs give away more free compute than any western provider not because they're generous developer mindshare is the new oil i've been running both ecosystems for months quality gap is shrinking fast access gap is already zero the real moat isn't the model it's who got you hooked on free credits first

THE BEST AI TRADING TOOL DOESN'T MAKE MONEY it kills bad ideas before you lose any one sentence → full backtest → "this strategy loses over 4 years" that's not failure. it's years of losses you never took jensen said traders with AI will replace traders without it but the real edge isn't alpha generation it's speed of falsification most people spend months convincing themselves a bad premise is solid ai collapses that to one conversation what's something you believed too long because nobody pushed back?

YOUR DUSTY MAC MINI IS THE BEST AI AGENT you have a device right now running 23 hours a day doing nothing. old laptop, spare mac, retired pc. one openclaw setup later it runs 24/7. no cloud costs. no session resets. the cloud ai model was a rental. local agents are ownership: - $200/mo vs $0/mo (hardware you already own) - persistent memory by default - your data never touches a datacenter the edge agent era doesn't need new hardware. it needs you to open that drawer. what's collecting dust in your house that could be building?

BINARY WAS A CHOICE, NOT A LAW every computer you've ever programmed thinks in two states: 0 or 1 that was an engineering shortcut from 1957, not a law of physics google's willow chip proved quantum error correction at scale. 20 year problem, solved calculation that takes willow 5 minutes would take the fastest supercomputer 10 to the 25 years - longer than the age of the universe that's not "faster". that's a fundamentally different kind of computation for ai builders this matters: - quantum solves optimization problems classical compute can't touch - hybrid quantum-ai pipelines are closer than most people realize - the assumption that "everything is bits" dies this decade Demis Hassabis said quantum + AI will unlock science we can't model today.

YOUR AGENT DOESN'T GET WORSE you stop feeding it everyone ships an agent once and expects magic forever new tools drop, the context goes stale workflows shift under your feet the ones that keep delivering aren't built differently they're maintained differently 1. feedback loops 2. context refreshes 3. prompt audits treat your agent like a product not a one-time script when did you last update your agent's context

@misat0x 💯

@silentguyy66 context beats clever wording most days

THE REAL CLAUDE PLAY every "print money with ai" post sells prompts Boris Cherny, creator of claude code, does most of his work from his phone model writes 100% of his code now the leverage isn't in prompts. it's in artifacts.

THE GUY WHO BUILT OPENCLAW STOPPED PROMPTING AGENTS peter steinberger created the most popular open-source agent framework now at openai. his exact words: "you shouldn't be prompting coding agents. you should be building loops" the creator uses his own tool differently than everyone else are you still prompting or already looping?

THE $5 SMART HOME IS ALREADY HERE off-the-shelf "smart" gadgets are just on/off switches in fancy plastic $50 for a motion sensor that detects one thing: motion a $5 esp32 + temp/presence/light sensors + ai logic = context it checks the time, the weather, your habits, not just "something moved" building your own isn't harder anymore it's just smarter the smart home industry is selling you overpriced relays build your own. it costs less. it works better.

THE MODEL ISN'T THE BOTTLENECK anthropic engineers ship 8x more code per day same model. same hardware. same team size. what changed? what claude sees before it writes context engineering > prompt engineering always was. the best AI coding setup isn't a better LLM it's giving your current one the right context.