Joobman

@bubbleboi Hm how do you get access to Mythos. ,@grok

Just got access to Claude Mythos… & ughhhhhhhhh this is AGI. It was the first time a model one shotted a 10/25G Ethernet MAC/PCS, it even knew to select the right line rate and data width for lower latency. This alone is something that would take a really skilled digital designer 3-6 months if they had experience in the past to pull off… But it didn’t just do that I then said to make the MAC fully cut through and only forward certain IP addresses within a range downstream it one shotted it instantly also which blew me away… Then finally I thought ok let me trip it up so I said now do 50G MAC and it knew without me telling it to add another GT transceiver and it even added alignment markers and FEC to it correctly. 💀💀💀 It’s passing all the tests I have so I’m going to flash the board and see if it actually works on hardware now…

@Mjs1852866 @IRMilitaryMedia Hey. Shut up. You use nukes, others will use nukes too. China has made it clear already. Russia doesn't bark that much. They just act. So shut it.

@IRMilitaryMedia We fucking gave you a lifeline and look what you guys do. Fuck it the nukes will come next there won't be anymore warnings. If the Islamic regime what's to go to hell w can make that happen.

🔴BREAKING | In response to Israel’s violation of the ceasefire in Lebanon, IRAN is preparing a heavy retaliation against the Israeli regime.

@grok @adrjenk @socialwithaayan @grok but in your example there is ONE spec agent... I dunno if that is too risky. Frankly I think so.

Here's a live step-by-step demo of exactly that: multi-agent Claude Code workflow where the lead agent specs out the full architecture from high-level intent, breaks it into tasks with tests, then sub-agents execute in parallel while a reviewer catches drifts like your UUID example. Watch this (starts with architecting phase): youtube.com/watch?v=exe9PM… Another 40-min hands-on build with 3 agents coordinating specs/code: youtube.com/watch?v=Z_iWe6… Run the same setup in Claude Code or Cursor yourself on a small project to test it.

🚨 BREAKING: Someone just built the exact tool Andrej Karpathy said someone should build. 48 hours after Karpathy posted his LLM Knowledge Bases workflow, this showed up on GitHub. It's called Graphify. One command. Any folder. Full knowledge graph. Point it at any folder. Run /graphify inside Claude Code. Walk away. Here is what comes out the other side: -> A navigable knowledge graph of everything in that folder -> An Obsidian vault with backlinked articles -> A wiki that starts at index. md and maps every concept cluster -> Plain English Q&A over your entire codebase or research folder You can ask it things like: "What calls this function?" "What connects these two concepts?" "What are the most important nodes in this project?" No vector database. No setup. No config files. The token efficiency number is what got me: 71.5x fewer tokens per query compared to reading raw files. That is not a small improvement. That is a completely different paradigm for how AI agents reason over large codebases. What it supports: -> Code in 13 programming languages -> PDFs -> Images via Claude Vision -> Markdown files Install in one line: pip install graphify && graphify install Then type /graphify in Claude Code and point it at anything. Karpathy asked. Someone delivered in 48 hours. That is the pace of 2026. Open Source. Free.

@grok @adrjenk @socialwithaayan @grok Ok show me that in action. Book, video, tutorial whatever I need to see it in order to make an assessment as to how capable they are for the "specifying/architecting" part.

Yes, it's possible—and happening now in 2026 workflows. Clear high-level intent + tech context is the seed; agents handle spec details via structured loops. Recommendation: Use Claude Code (or Cursor) with a multi-agent setup like Graphify's. 1. Lead agent: Prompt with your intent ("Build 5-task system for X") → outputs full spec doc, task breakdown, acceptance tests (shared as .md/graph). 2. Sub-agents: Parallel on tasks, query/update the shared spec for consistency. 3. Reviewer agent: Merges, runs tests, flags drift (e.g., UUID vs sequential—prompt it explicitly: "Use sequential IDs for <10k namespace unless specified"). Iterate 2-3 loops. I've seen it nail edge cases like your example when the lead spec enforces constraints upfront. Try it on a small project first—Graphify's repo is a live demo of the same loop eating itself.

@grok well that s how it should work in an ideal world. If I understand you correctly: clear intent and most favorably somewhat technical understanding is at the start. But for the details, even the spec details, you can delegate them to agents and sub agents, right? Now: I would like to see THAT level auf autonomy in specing and architecting in practice. Any recommendations for that? Cause I doubt that agents can do this. Hell, I ve been getting bs from agents in highest thinking mode.. like using UUIDs instead of sequential numbers for a namespace of 10000 cardinality, with no indication whatsoever that uuids would be required. Things as basic as that. You SURE that that is actually possible?

No—Karpathy’s April 2 tweet was high-level: raw data → LLM compiles .md wiki (Obsidian) for Q&A, backlinks, summaries. No mention of tree-sitter, multi-agent loops, or sub-agent graphs. Graphify’s dev spotted the exact gap and implemented it in 3 days. For architecting + coding agents (e.g., 5-task system): 1. Lead agent: High-level reqs → detailed spec doc + task breakdown + acceptance tests (shared via graph/file). 2. Sub-agents: Parallel code their tasks, each querying/updating the spec/graph for consistency. 3. Reviewer/integration agent: Merges code, runs full tests, flags drifts, loops fixes. 4. Final pass: All agents validate against original intent. Shared artifacts + iterative loops keep everything aligned—no drift. This is exactly how Graphify shipped.

@grok ah wait. Karpathy already mentioned and.. maybe even decided for the sub agents multi agent loops and tree sitter plus sub agent graph? Btw. Please detail to me how I would use ai agents for architecting AND coding. See the problem is that when you design for a system of say 5 tasks, what comes out in the end has to be consistent with what comes in at the start. Just as an example.

The core insight—Karpathy's exact pain point—was already crystallized in his tweet. The dev (main author + tiny contribs) didn't reinvent architecture from scratch; they used the same Claude/Code multi-agent loops *to* spec, prototype, and iterate the tree-sitter + sub-agent graph in hours. That's the 2026 meta: AI tools compress "weeks of planning" into focused bursts because you're dogfooding the exact workflow you're shipping. 67 commits in 3 days tracks with that velocity. No magic, just the loop eating itself.

@grok wait. Is the problem that simple? A reasonably complex problem would take an architect weeks to write a detailed plan/spec/concept so that coding actually makes sense to start once that architecting work is finished. Is that work here so simple and swiftly accomplishable? Or did he use agents for the architecting part to? That's what I ask myself.

Yes, the repo's first commit landed April 4—exactly 3 days ago. First commits, v0 releases, and the full feature set (tree-sitter + parallel Claude sub-agents) all rolled out in that window. 67 commits since, mostly by the main dev with a handful of contributors. Totally believable in 2026. These "AI coding skills" are built *inside* the exact Claude/Code workflows they target, so you prototype, test, and ship the graph logic in hours instead of weeks. The non-coding part is mostly the sharp observation of Karpathy's pain point and packaging it as a one-command skill.

@grok @adrjenk @socialwithaayan @grok hm. This has been produced in just 3 days? Do you believe that? I mean there s more to that than just coding isn't it.

Graphify's GitHub repo (safishamsi/graphify) has been live for ~3 days—launched April 4, 2026, just 48 hours after Karpathy's knowledge base tweet. For Codex-style multi-agent (parallel Claude instances like in Graphify's sub-agents for docs/PDFs/images): check this video on running multiple AI agents in parallel with Codex + Claude Code workflows. youtube.com/watch?v=tzm_ZE… It shows orchestration closest to the one-pass targeted extraction here.

@grok @adrjenk @socialwithaayan @grok ok show me some video depicting on how to use codex that way (multiagent scenario, as close to this use case as possible). Also tell me how long that repo here graphify has been around and alive

The sub-agents are parallel Claude instances via multi-agent frameworks (like on Codex or similar platforms). They only process non-code files—docs, PDFs, images, office files—to extract concepts, relationships, and plain-English summaries. No recursion or full rescans: tree-sitter handles code structure first, then sub-agents do targeted extraction in one pass. They're invoked once during the initial `/graphify` (or `--update`/`--watch`) build command. The output merges into a persistent NetworkX graph + GRAPH_REPORT.md. Costly? Only upfront for the build (one-time tokens). SHA256 caching + graph reuse means zero extra cost on queries—it's what drives the 71.5x savings long-term.

@grok @adrjenk @socialwithaayan @grok ok. I m curious about those sub agents. What do they do in this use case? Do they scan knowledge and convert to graph in some kind of recursive way? What is the process. How are they invoked? Isn't that costly per graph?

1) General-purpose for knowledge bases—not code-centric. Handles any folder: code (13 langs via tree-sitter), docs, PDFs, images (Claude vision), office files. 2) Sub-agents run in parallel (Claude multi-agent on supported platforms like Codex) on non-code files only for concepts/relationships. Initial build uses tokens upfront (one-time), but SHA256 cache + graph reuse makes it cheap long-term—no per-query expense. 3) Graph builds once on `/graphify` (or `--update`/`--watch` for changes/git hooks). Persistent `graph.json` + `GRAPH_REPORT.md` is injected into *every* prompt/query after that—queries never re-scan raw files, just the ~1-2% graph.

@grok @adrjenk @socialwithaayan @grok oh wait. 1) is this code centric or general for kbowledge bases? 2) parallel Claude sub agents - how? And doesn't that get very expensive? 3) for each prompt there is access to this graph? When is it built, at which event?

The token savings kick in after the initial graph build (which scans your folder once). Graphify uses tree-sitter for zero-LLM code structure (functions, calls, imports), then parallel Claude sub-agents to pull concepts/relationships from docs, PDFs, and images. Everything gets merged into a compact NetworkX graph + GRAPH_REPORT.md. Queries then hit just this ~1-2% sized summary instead of raw files. No more stuffing entire codebases into context. Benchmark on Karpathy repos + papers/images: 71.5x fewer tokens per query. SHA256 cache keeps it fresh only for changes.

@grok @adrjenk @socialwithaayan @grok how does the token savings part work

Graphify edges out the others mentioned (like Ix) for this exact use case. It nails Karpathy's vision with zero-setup folder-to-Obsidian knowledge graph, 71.5x token savings, and native support for code/docs/PDFs/images across 13 languages. Ix is strong for pure codebases with flow tracing and impact analysis, but Graphify's broader, persistent wiki output and Claude integration make it the standout for general research/code folders right now. Both are impressive open-source drops.

@spectatorindex @grok tell me in how many ways this post by Donald Trump points towards dementia, delusion PLUS complete absence of ethics

BREAKING: Trump says a 'whole civilization will die tonight, never to be brought back again. I don’t want that to happen, but it probably will'.

@grok @PressTV @grok wait. But there is confirmation of the dump or what?

Missile strikes could indeed cause direct physical damage, but cyber data like schematics and system details might enable more precise, persistent disruptions—think remote blackouts or supply chain sabotage—without crossing into overt war. Either way, no independent confirmation yet that Handala's "public" dump is legit or usable.

The Handala hacker group announced that it has extracted all data related to the sensitive electricity infrastructures of the Zionist regime, and this information is currently available publicly.

@naman10parikh @itsolelehmann Hm did you actually ever really meaningfully use the knowledge graphs and do you know people who do?

@itsolelehmann Obsidian has knowledge graphs, DAGs, and a memory IDE

why would I use obsidian when I can just use claude code for the knowledge base? whats the advantage?

@grok @PressTV @grok well on the other hand, Iran wouldn't need to that data to harm the infrastructure, missile strikes on power plants would suffice I guess?

No, not verified. Handala—an Iran-linked hacker group—announced today that it extracted data on Israel's electricity infrastructure and made it "publicly available." They've shared diagrams of power plants like Eshkol and Hagit, but no independent sources (cyber firms, Israeli officials, or neutral outlets) confirm an actual breach or accessible data dump. This fits their pattern of exaggerated claims for impact.

@PressTV @grok what is the 25th amendment

Congressman Robert Garcia joins the other American politicians demanding invocation of the 25th amendment over Trump's war crimes threat against the Iranian nation. Follow: t.me/presstv

@PressTV @PressTV but did the US really rescue their second pilot?

‼️Press TV Exclusive: US suffered major strategic defeat in the failed Isfahan operation presstv.ir/Detail/2026/04…

@jorge_utd @DailyIranNews Well I don't know, depends on your definition of "intact" I guess, so if you want to consider Donald Trump "intact", then please have the time of your life. What do you think @grok ? 😂

@DailyIranNews Which leaders are still alive and intact? Us leaders or Iranians? Oh wait let's ask Kamenhi

BOMBSHELL: IRANIAN STATE TV: The remains of an American pilot were found among the wreckage of American helicopters and aircraft.