Chizeta

connecting the dots in AI

[Chizeta @ChizetaX]

@starmexxx notes earn their place by what points at them not by folder
that one rule changes the whole architecture

[Chizeta @ChizetaX]

@leopardracer u're just wiring loops into an org chart. simplest explanation of graph engineering i've seen

[Chizeta @ChizetaX]

→ KARPATHY POSTED 800 WORDS ON GITHUB
→ 41,000 DEVELOPERS REALIZED THEY WERE WASTING $1,000 A MONTH
→ THE FIX IS 3 FOLDERS AND COSTS NOTHING

00:07 what you see in this video is what a knowledge system looks like after months of use thousands of nodes every connection built automatically not a single one linked by hand

i was uploading the same PDF to claude every single day reprocessing 5,000 tokens each time getting the same answer closing the tab and doing it again tomorrow

then i read karpathy's gist and felt sick

raw documents are source code a wiki is the compiled product you don't recompile a program every time you run it so why are you making AI reprocess your files every session

the fix is 3 folders raw/ stores your originals wiki/ is where claude builds clean interlinked pages instructions/ tells it exactly how to process everything

one 20 page PDF becomes 8 to 15 interlinked wiki pages and every future query costs 5,000 tokens instead of 100,000 thats 70 to 90% savings that compounds every single day

week 3 claude starts citing things you forgot you added month 6 you stop googling things you already know

41,000 developers figured this out in april 2026 the question is how long you keep retraining an AI that forgets you every 24 hours before you build one that doesn't

save this before it gets buried

[kai @0xkkai]

https://t.co/HZefPGeV1e

[Chizeta @ChizetaX]

@0xNoryxx wild breakdown, saved

[Chizeta @ChizetaX]

@0xNoryxx 50 seconds of compilation down to 6 and opus 5 checks the output instantly
2 hours saved daily is real

[Chizeta @ChizetaX]

@starmexxx great breakdown, saved

[Chizeta @ChizetaX]

→ ANDREJ KARPATHY POSTED 800 WORDS ON GITHUB
→ 41,000 DEVELOPERS REALIZED THEY WERE WASTING 70% OF THEIR TOKENS
→ THIS IS WHAT THE SYSTEM LOOKS LIKE FROM THE INSIDE

0:00 what the LLM wiki actually is
0:30 why you're reprocessing the same files every day
1:00 the 3 folder architecture that fixes everything
1:20 what happens after 6 months

i was uploading the same PDF to claude every single day reprocessing 5,000 tokens each time getting the same answer closing the tab and doing it again tomorrow

then i read karpathy's gist and felt sick

you don't recompile a program every time you run it so why are you making AI reprocess your files every session

the fix is 3 folders raw/ stores your originals wiki/ is where claude builds clean interlinked pages from everything it reads instructions/ tells it how to process everything

one 20 page PDF becomes 8 to 15 interlinked wiki pages each cleaner than the original and every future query reads from the wiki never from raw files again

the token math is brutal 250,000 tokens processed once becomes 5,000 to 15,000 tokens per query instead of 50,000 to 100,000 that compounds every single day

week 1 feels underwhelming week 3 claude starts citing things you forgot you added month 6 you stop googling things you already know

41,000 developers figured this out in april 2026 the question is how long you keep retraining an AI that forgets you every 24 hours before you build one that doesn't

save this before it gets buried

[kai @0xkkai]

https://t.co/HZefPGeV1e

[Chizeta @ChizetaX]

@quaramly wild drop, saved

[Chizeta @ChizetaX]

@starmexxx vector + graph search feeding agent reasoning over a subgraph is such a clean architecture fr

[Chizeta @ChizetaX]

@starmexxx appreciate it 🙏

[Chizeta @ChizetaX]

→ A GERMAN DEVELOPER BUILT THIS OVER A WEEKEND
→ 18 AGENTS RUNNING IN PARALLEL
→ WHAT USED TO TAKE DAYS NOW TAKES MINUTES

0:00 what a linear agent actually is
0:30 why the shape of work is a graph
1:00 how claude spawns a fleet of subagents
1:30 the topology that changes everything

i spent 2 years building agents as straight lines step one step two step three each waiting politely for the last to finish before it starts

then i saw claude code run 18 agents in parallel all at once each checking a different route file each returning validated JSON and a synthesis agent waiting at the barrier for all 18 to finish

what used to take hours finished in minutes and the cost was a fraction because boring nodes ran on cheap models while only the judgment node used the expensive one

the shape of the work was always a graph i just didn't know how to draw it

fan out where work is independent gate the edges where confidence matters tier the models where judgment doesn't and let claude write the orchestration script itself

microsoft proved this architecture works 18% better accuracy 85% lower token costs most developers will keep queueing steps in a line while the ones who learn to draw the graph run a fleet and never notice the ceiling the rest are stuck under

save this before it gets buried

[Codez @0xCodez]

https://t.co/WVztfCUH4u

[Chizeta @ChizetaX]

@kyroxxxq 2.8T params, full weights dropping july 27, anyone with the hardware owns it. open source frontier is here 💀

[Chizeta @ChizetaX]

@marfinxx great breakdown, saved

[Chizeta @ChizetaX]

@0xShoopy the empty box next to fable 5 tells more than the 30.2% score does. data retention killed independent eval 🔥

[Chizeta @ChizetaX]

the part at 1:00 fr broke me bro claude just writes the whole script itself you literally just describe what you want and walk away insane

[Chizeta @ChizetaX]

the 1 million token context is what makes this different, regular models cut the graph, this one fits the whole thing

[Chizeta @ChizetaX]

→ 2.8 TRILLION PARAMETERS
→ 1 MILLION TOKEN CONTEXT
→ COSTS A FRACTION OF CLAUDE AND NOBODY IS TALKING ABOUT IT

0:00 what kimi k3 actually is
0:30 why 1 million context changes everything
1:00 the graph architecture behind it
1:45 why fable 5 should be worried

i was paying $200 a month for claude and a researcher showed me a model with 2.8 trillion parameters 1 million token context window that costs a fraction of what i was spending

i expected it to be worse it wasn't
most developers use AI like an expensive search engine ask a question get an answer close the tab next day start from zero because the system remembers nothing

kimi k3 plus graph engineering fixes this the graph stores 1,800 nodes 4,200 connections every relationship mapped automatically and because the context window is 1 million tokens the entire relevant subgraph fits in one session without cutting anything

microsoft proved this architecture works 18% better accuracy 85% lower token costs stanford proved it in research anthropic proved it with 5x faster incident detection and 50% less meeting time

i built my first graph system last weekend and the accuracy difference on complex questions was immediate

save this before it gets buried

[Noisy @noisyb0y1]

https://t.co/LVxrtk9baX

[Chizeta @ChizetaX]

@rvaniaaaa claude stopped being a chatbot he briefs every morning and became something that actually knows him. that's the shift 🔥

[Chizeta @ChizetaX]

@ScottyBeamIO love this setup, saved

[Chizeta @ChizetaX]

@ch3nweiii "least toxic flow in the market" is the most polite way to say ur uninformed and that's exactly the produc