Dishant Ghai

Founder,Coder (read Dreamer,Doer) for Zoodify

[Haiyu Wu @HaiyuWu1]

We are trying to build a free LeetCode for AI researchers. You can learn how to code core ideas from important papers across various areas (e.g., LLMs, JEPA, RL, diffusion), contribute to the problem bank, and importantly, post your thoughts to find like-minded people for collaboration or job opportunities.

Of course, the current version is still far from mature, so anyone who wants to help us with this project is welcome!

[deezzex @deezzex]

spaceXAI engineer leaked a complete guide on reverse engineering 

so i pointed a local LLM at an app i had zero source code for.

no cloud. no api key. no internet. just my machine, a quantized model and a binary.

40 minutes later it handed me back:

→ every function renamed to something a human would write
→ the login flow mapped step by step
→ the hidden debug menu the devs forgot to strip
→ a spec md clean enough to rebuild the whole thing from scratch

reverse engineering used to be a craft you spent years on. reading assembly at 2am. guessing what sub_401a3c does. losing a whole weekend to one obfuscated function.

now the model reads it in seconds and explains it like a senior dev doing a code review.

what actually made it work (keep this):

1. don't dump the whole binary in. feed it one function at a time, smallest first
2. let it rename as it goes. every good name makes the next function easier to read
3. make it write down a guess before it explains anything, then check the guess against what the program really does at runtime
4. keep a running notes file the model reads every time. that file becomes the spec
5. run it locally. your target never leaves your machine, so there are no logs and nobody else sees it

the crazy part isn't that it works.
it's that it runs on a laptop.

the gap between "i wonder how this app works" and "i know exactly how this app works" is now one prompt.

i'm going deeper into local ai + reversing every week. stay tuned if you want to see what i break next.

and show this to the one friend who still thinks reverse engineering is only for hackers in hoodies.

real question: if any app can be read like open source now, is closed source already dead?

[Joying.eth @Joyingeth]

Memori hits 87% on LoCoMo at 721 tokens per query, 2.8% of full context, per README

Graphiti v0.31.1 tagged Oct 15. Temporal fact graphs with validity windows.

9 open-source GitHub repos for agent memory that survives sessions and keeps the prompt short:

212,966 stars. Pushed within 31 days. 6 Apache-2.0, 3 MIT. Oct 18.

EXTRACT WHAT MATTERS

01. mem0 ★67,374: https://t.co/WiO0IMQTUH -> extracts facts, recalls per user
02. langmem ★1,708: https://t.co/iIIeZ2SGmP -> agent learns from past chats
03. MemMachine ★3,073: https://t.co/2IRDjGetmh -> one store, any agent
04. memU ★14,554: https://t.co/HbadZV9AJN -> shared memory across agents
05. Memori ★17,151: https://t.co/HinrcYR31C -> chats into structured state

LINK IT IN A GRAPH

06. graphiti ★31,950: https://t.co/xRZyFjKlmA -> facts with validity windows
07. agent-memory ★585: https://t.co/MVsjHmSKak -> memory graph on Neo4j

RETRIEVE A SLICE

08. LightRAG ★40,200: https://t.co/yn7KG7XBTW -> graph retrieval, no heavy index
09. graphrag ★36,371: https://t.co/DgYEBR55Sc -> graph summaries for retrieval

turn → facts → graph → slice → call

pip install mem0ai

not more tokens. better recall ↓

[DeepSeek @deepseek_ai]

🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!

Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.

🔗 https://t.co/JGAik1kgk0

1/4

[morph @morpphhhaw]

THIS IS A CLEAN-ROOM SETUP THAT ACTUALLY WORKS

Anthropic just showed a way to make Opus 5.5 rebuild a feature it was never allowed to see

Two rooms. One hard rule: room B never sees the original

Here's the split:

→ REA sits in room A. It reads the app (strings, xrefs, IPC, runtime trace) and writes SPEC.md. Every rule must cite evidence. No evidence, no rule
→ JEV is the gate. Proof crosses. No proof goes back to REA. Rule 4 came back, REA dug again, nothing to cite, dropped. It never reached room B
→ Opus 5.5 in room B gets a fresh session and SPEC.md only. It rebuilds the feature from behavior alone
→ Haiku 5.5 writes the tests and three deliberately broken versions of the rebuild. The real one must pass. All three broken ones must fail. If a broken version passes, your tests were decoration
→ /goal keeps looping until all three fail

Passing tests prove nothing. Tests that catch a planted bug prove they're watching

The whole thing is one command. Paste this into Claude Code:

"Set up a clean-room rebuild"

1. REA in room A reads the target feature (strings, xrefs, IPC, runtime trace) and writes SPEC.md where every rule cites evidence.
2. JEV gates every rule. No proof, send it back. Drop anything that can't be cited.
3. Opus 5.5 in room B gets a fresh session and SPEC.md only. It never sees the original. Rebuild from the spec.
4. Haiku 5.5 writes tests plus 3 deliberately broken versions. Tests must pass the real rebuild and fail all 3 broken ones.
5. /goal loops until all 3 broken versions fail.
Show me the SPEC.md and the three broken diffs before anything merges."

One command. Two rooms. The model that rebuilds never sees the code it is supposed to match

Save it

[virgilxbt @virgilxbt]

This is f*cking gold..

Chinese researchers just published a paper with a brutal title: "The End of Software Engineering".

The argument fits in one line: writing code is finished as a skill.

A human brain can juggle only so much state before it locks up. An agent hits no such wall, and every round of compute pushes its limit further out.

Typing stops being your part of the job.

You become the intent architect: you decide what gets built and you sign off on what comes back.

The next raise goes to whoever can tell a solid agent run from a confident wrong one.

Syntax turns disposable. Judgment turns into the job.

So I wrote the guide below on building an agent system of your own. The paper behind it is in there too ↓

[Promix @Promixyz]

10 GitHub repos can turn Claude Code into an entire AI development team

Give it better code navigation, live documentation, reusable skills, browser debugging, and automated PR workflows

(1/10) Claude Code
the foundation for an AI coding agent that works directly inside your codebase
▸ https://t.co/SkjKYCjDnK

(2/10) Superpowers
add structured planning, test driven development, and systematic debugging
▸ https://t.co/LTKFfkvzMr

(3/10) Serena
navigate large codebases using symbols and semantic code intelligence
▸ https://t.co/fyAknbmyhb

(4/10) Context7
pull current library documentation instead of guessing outdated APIs
▸ https://t.co/1CnhgAIGo7

(5/10) Anthropic Skills
equip Claude with reusable instructions, scripts, and specialized workflows
▸ https://t.co/92AxiUAo9k

(6/10) Playwright
automate browser tests and catch UI regressions
▸ https://t.co/6leZ4QgPiQ

(7/10) Chrome DevTools MCP
inspect browser errors, network requests, and performance traces
▸ https://t.co/IiOUmLfcTB

(8/10) Claude Code Action
bring Claude into GitHub issues and pull requests
▸ https://t.co/YtEj3OJm3L

(9/10) Semgrep
scan code for security issues and risky patterns
▸ https://t.co/GKTMvjvTHr

(10/10) Langfuse
trace and evaluate LLM workflows when your project uses model APIs
▸ https://t.co/hTgXhmn1Ul

The interesting part?
The biggest problem with AI coding isn't generating code

It's knowing whether the code actually works
Give Claude Code a repeatable process:

1 → Read the codebase before making changes
2 → Retrieve current documentation
3 → Write a failing test before fixing the bug
4 → Implement the smallest working change
5 → Run tests and inspect the result before opening a PR

Start with Claude Code + Serena + Context7, then add

Superpowers for structured development

You don't need all 10 repos on day one

Save these 10 repos before your next Claude Code project

[Daily Dose of Data Science @DailyDoseOfDS_]

The next step after Karpathy’s agentic engineering:

(built by Google; bookmark this)

Karpathy outlined the disciplines that separate production agent development from vibe coding: precise specifications, evaluation loops, and security oversight.

What was missing was a workflow connecting these ideas from the initial specification to a published agent.

The diagram below maps that complete lifecycle.

We tested it by building a RAG agent from scratch with Google’s Agents CLI and Claude Code.

It scaffolded the ADK project, generated 20 evaluation scenarios, found an instruction loophole before deployment, and produced a quantitative scorecard.

It then deployed the agent to Agent Runtime and registered it with Gemini Enterprise.

Everything happened in one continuous workflow, from specification to a production agent.

We documented the complete build in an article. It covers persistent state, resumable execution, human approval, cloud identity, and the application logic required to handle repeated requests safely.

Read it below.

[Huda @AlAssaf_H]

10 GitHub repos for AI-driven offensive security.

Autonomous pentesting, vuln discovery, exploit development, prompt injection, and MCP security.

Strix — Multi-agent pentesting: recon, exploitation, PoC validation.
https://t.co/MIAeyCGeGP
Shannon — Source-aware web/API pentesting with exploit-backed findings.
https://t.co/GQc9LmDhXH
PentAGI — Multi-agent orchestration for automated offensive workflows.
https://t.co/LFsD8Sii1k
Buttercup — AI-assisted fuzzing, vuln discovery, and automated patching.
https://t.co/mhSFOtQWLX
Vulnhuntr — LLM-driven source analysis and attack-relevant call-chain tracing.
https://t.co/1cxslohlTZ
ExploitGym — Exploit-development benchmark for AI agents: userspace, V8, Linux kernel.
https://t.co/JlXDaeMpDT
AgentDojo — Prompt-injection attack/defense benchmark for tool-using agents.
https://t.co/tgKSGPlT7s
Snyk Agent Scan — Security analysis of AI agents, MCP servers, and agent skills.
https://t.co/hlKzKdw3Da
Promptfoo — LLM/agent red teaming: injection, data leakage, unsafe tool use.
https://t.co/hCuISgXWt3
PyRIT — Automated adversarial testing of generative AI systems.
https://t.co/Q515j81GMA

Two attack surfaces: AI agents as offensive operators, and AI agents as exploitation targets.

[Jonghae Park @jonghae717]

No prior data. No rewards. No pretrained models.

Can an RL agent learn to solve extremely long-horizon tasks (e.g., 5-cube stacking) from scratch?

We show it can! C-Slerp learns hyperspherical contrastive representations and infers waypoints from them in closed form 🧵👇

[Rob Pruzan @RobKnight__]

Introducing Pixel - a library for building GUIs in the terminal

- 120 FPS pixel-perfect graphics
- Embeddable inside existing TUIs
- Built in chromium webview component

https://t.co/rZJ6lyHgJD

[Deliriumm @0xDeliriumm]

yann lecun, french-american researcher born in paris in 1960, has been chief AI scientist at Meta since 2013. Meta's market cap grew from $60B to $1.5T on the back of transformers and LLMs. he spent those same years publicly telling anyone who would listen that the entire architecture was fundamentally broken.

nobody listened. the transformer became a $4.6T industry.

this october 2026 paper - independent research, not affiliated with any institution - is the mathematical proof that he was right.

standard transformers compute in a single pass. H^(l) = Attention(H^(l-1)). one shot. no persistent state. no memory between tokens. no ability to revise reasoning mid-computation.

at 1K context: transformer accuracy 40%, NCM 60%. at 8K: transformer 42%, NCM 70%. at 32K: transformer 38%, NCM 75%. at 128K tokens - the scale every enterprise deployment now requires - transformer collapses to 35%. NCM holds at 80%.

the architecture replacing it: h_{t+1} = f(h_t, x_t, θ). a continuous internal state. information doesn't pass through once and disappear. it persists. it evolves. it gets refined across multiple computation steps.

state update: h_{t+1} = h_t + Δt · F(h_t, x_t, θ). a gated continuous function. the model doesn't process the next token and move on. it updates its entire internal representation based on everything it already computed.

transformer: quadratic complexity. static weights. implicit memory. reasoning: emergent, unreliable.

NCM: linear complexity. dynamic state. explicit memory h_t. reasoning: built-in. architectural.

lecun said in 2023: "LLMs cannot reason. they cannot plan. they are not going to get us to AGI."

the research community responded by raising another $500B in LLM funding.

Meta paid him $50M+ over a decade to be the loudest critic of the technology that made them $1.4T in market cap.

in 2026, anonymous independent research notes are publishing the architecture he predicted four years ago.

the paper ends with one sentence: "NCM may not simply improve LLMs - it could mark the next generation of language models."

lecun has been writing that sentence since 2022.

save this.

[Argona @Argona0x]

f*cking insane. I hired an AI chief of staff for $0, gave it one rule: "find me $50,000 a month or you're fired", went to sleep, and Opus 5.5 had it on my desk by morning

a SpaceXAI engineer said one line that hit 5 million views:
"i have a Chief of Staff agent. he knows about all my other bots and manages everything"

a human one costs $200,000 a year. mine runs on Opus 5.5 inside Viktor, on his own computer in the cloud, with my laptop shut

i give the goal, he does the steps, i only make the calls

the rule:
"find me $50,000 a month of new business. public data only. contact nobody without my yes"

the night:
→ 9pm broke the goal into 4 jobs and gave each a day and a time, i wrote none of them
→ 11pm pulled 240 dental clinics in Austin from Google Maps and cut every chain
→ 1am read every site that was left and kept 25, each with one thing he noticed and where he saw it
→ 2am wrote an opener for every one of them, sent zero
→ 3am built the offer on one page, no invented results, price left blank for me instead of guessing
→ 8am i typed $2,000 a month. 25 clinics × $2,000 = $50,000 a month on the table

honestly the blank price is what i'd pay for, an AI that guesses your price loses you the deal

and he does it again every Monday: 100 new clinics a month, and he never forgets to start

240 clinics → 25 worth it → 1 offer → $50,000 a month on the table → 0 sent without my yes

second 12, he leaves the price blank and hands it to me

someone's chief of staff is already pitching the clinics in your city

i left the whole setup below ↓ free, steal it

[Arindam Majumder 𝕏 @Arindam_1729]

this is gold for anyone building coding agents.

Someone wrote a guide to building a Jev harness.

the promise:
→ 200x faster
→ 400x cheaper
and the best part?
you don't need a better model.

the bottleneck is often the harness around it: how your agent routes tasks, uses tools, manages context, and decides what deserves a full LLM call.

here's how to try it:

→ grab the PDF and the accompanying article
→ drop both into Claude Code or Codex
→ ask it to rebuild its own harness based on the guide

one evening of setup could give your coding agent a serious upgrade.

same model. smarter harness. less waiting, lower bills.
most teams are still trying to solve every problem by switching models.
start by fixing the system around them.

[Antonio Lupetti @antoniolupetti]

"Mathematical Theory of the Transformer" is an extensive, 825-page mathematical monograph exploring the theoretical foundations of Transformer architectures.

It covers a remarkable range of topics, including embeddings and positional encoding, self-attention and multi-head attention, matrix decompositions, linear algebra, spectral analysis, information geometry, training dynamics, redundancy, optimization, and incremental Transformer architectures.

This is an advanced, mathematically rigorous read for anyone interested in understanding the mathematical structures underlying Transformer-based AI systems, including modern generative AI.

https://t.co/wwa1ucTYiF

[hoshino @hoshinoxbt]

This is f*king insane. Andrej Karpathy spent 8 years at OpenAI and Tesla

Last week, he packed what he learned into one free 2-hour lecture

Agents → Harness → Loops → Graphs

Some bootcamps charge $15K and teach less than this

You might not have 2 hours today

Bookmark and watch it it before it disappears from your feed

Then read the guide below to build your first agent harness

[Hamel Husain @HamelHusain]

New blog post 🎉: How To Use Decision Models For Evals

In this post, we show you how to choose a decision model for your task, tune its threshold (don't assume the 0.5 default is good!), and interpret its scores.

Lots of cool interactive charts here. https://t.co/uYAfte9Kb5

[JosephZhou @openmys]

We plugged Google Research’s Harness² into JEV 1.14.

Zero weights touched, yet driving massive gains on vision benchmarks round after round.

Under the hood, the system:
⚡️ Rewrites its own prompts
🛠️ Builds its own tools
🧠 Distills its own experience

JEV-Harness² repo drops soon. Stay tuned.

[The New Stack @thenewstack]

A better reranker can't solve a retrieval problem. Watch the on-demand session to explore why finding the right candidates in the first place is critical for building accurate and cost-effective AI search.

Watch now: https://t.co/Aj0NkL7kOc

[Doug Turnbull @softwaredoug]

Reminder that @treygrainger and I are beginning our online conference "The Future of AI Search"

Lots of great names giving talks like Yonik Seeley (@lucene_solr ), @jobergum, and more

Today we're hearing from @AmelieTabatta on whats going on with late interaction in 2026

https://t.co/tySWrIW651