AI Advances (@AIAdvances) | Aguea

Photos and videos

Tweets

AI Advances @AIAdvances

May 16

Kenny Vaneetvelde @DeadlyPretzel 's “How the Internet Dies” reframes Dead Internet Theory for the AI era: not just bots and troll farms, but platforms that reward humans for behaving like machines. A timely read on AI & the slow hollowing-out of the web: ai.gopubby.com/how-the-inter…

How the Internet Dies

For years, “Dead Internet Theory” was framed as something done to us. Foreign bot armies. State-sponsored troll farms. Algorithmic…

1

31

AI Advances @AIAdvances

Apr 27

Your AI agent re-reads the same docs and never learns. @yanliliu has three fixes: 📷 RAG — retrieves at scale, never compounds 📷 LLM Wiki (Karpathy) — compiles knowledge that grows richer 📷 Fat Skills (Garry Tan) — acts autonomously on what it knows ai.gopubby.com/rag-llm-wiki-…

RAG, LLM Wiki, or Gbrain? How Your Agent Remembers Changes Everything

Karpathy’s compounding wiki, Garry Tan’s autonomous brain, and the decision framework most teams skip

1

63

AI Advances @AIAdvances

Apr 27

This is the latest post in @fabio_yanez_'s Titans series on Medium — the one where the story stops being about any single model and starts being about the hidden map behind all of them. ai.gopubby.com/miras-the-blu…

MIRAS: The Blueprint Behind Transformers, Mamba, and Titans

Google’s new framework reveals every modern sequence model is solving the same four-choice optimisation problem

38

AI Advances @AIAdvances

Apr 27

In this article, @UmairAliKhan81 demonstrates how to accurately extract/parse the content from documents with messy layouts. ai.gopubby.com/how-to-accura…

How to Accurately Extract Everything from Documents Using AI

How asking the AI models to output in a specific format changes everything

1

33

AI Advances @AIAdvances

Apr 26

From tabular RL to neural control: on-policy methods like Sarsa scale—but stability, features & tuning make or break performance. By @hrmnmichaels ai.gopubby.com/on-policy-app…

On-Policy Approximate Control Methods

Introducing Sarsa with Function Approximation

18

AI Advances @AIAdvances

Apr 20

What do we mean when AI “remembers”? Pooja Kashyap @poojakashyap unpacks why LLMs generate, not retrieve and why that shift changes how we build AI. #AI #LLM ai.gopubby.com/ai-doesnt-sea…

AI Doesn’t Search. It Remembers.

What if everything an AI understands about the world isn’t retrieved but remembered?

83

AI Advances @AIAdvances

Apr 10

Dr. @UmairAliKhan81 improved speech-to-text for low-resource languages with a 2-pass LLM method, reducing WER across multiple ASR models. ai.gopubby.com/how-i-improve…

How I Improved Speech-to-Text Accuracy with a 2-Pass LLM Pipeline

A post-processing method that reduced speech-to-text errors across multiple models

1

1

47

AI Advances @AIAdvances

Apr 6

60 clients. 0 employees. No coding. Powered by one file - CLAUDE.md. Not prompts - it is an AI operating manual. The real shift? Moving from using tools to designing workers. The real skill? Translating work into structure. By @calvindong15 ai.gopubby.com/the-most-impo…

The Most Important Employee at a Japanese Firm Was a Markdown File

How a Japanese tax accountant uses CLAUDE.md to serve 60 clients and what it reveals about future of knowledge work.

27

AI Advances retweeted

Manjunath Janardhan

@manjunath_shiva

Apr 4

I just published my first technical deep-dive: adapting @Google 's #TurboQuant for weight quantization on @Apple Silicon using MLX. The headline result on Qwen 2.5-7B: • Standard 2-bit quantization: perplexity 2199 (garbage) • TurboQuant 2-bit QJL: perplexity 15.93 (usable) • Zero training data required How it works: 1. Randomized Hadamard rotation makes weights Gaussian 2. Lloyd-Max codebook (optimal since 1960) replaces uniform quantization 3. 1-bit sign correction cleans up the residual — 43-45% MSE reduction At 4-bit QJL, we're within 4% of FP16 quality. The advantage grows with model scale — bigger models = wider layers = better Gaussian approximation. Published in @AIAdvances on @Medium : medium.com/ai-advances/i-bui… Also preparing an @arxiv preprint for cs.LG — if you can endorse, DMs are open.

I Built a Quantization Method That Beats Standard 4-bit on a 7B Model — With Zero Training Data

Adapting Google’s TurboQuant for weight compression on Apple Silicon using MLX, and what the perplexity numbers actually tell us

1

2

141

AI Advances @AIAdvances

Apr 4

Tesseracts as Data Structures for Understanding High-Dimensional Computing. By @poojakashyap ai.gopubby.com/tesseracts-as…

Tesseracts as Data Structures for Understanding High-Dimensional Computing

There is a scene in Interstellar that I have seen multiple times and still not over it. Cooper, having plunged into a black hole, finds…

1

94

AI Advances @AIAdvances

Mar 24

Vibe coding looks great until it reaches production. Durable code needs engineering. By @kzdeb ai.gopubby.com/beyond-vibe-c…

Beyond Vibe Coding: The Organizational Redesign

New roles, better metrics, the apprenticeship problem, and governance that actually works — what changes when agents build and humans…

1

27

AI Advances @AIAdvances

Mar 21

Anomaly detection fails where signal breaks: low volume, cold start, false confidence. Lessons from a T20 World Cup pipeline. By @sengorajkumar ai.gopubby.com/i-predicted-t…

I Predicted the T20 World Cup 2026 Final With an Anomaly Detection Pipeline.

Building an anomaly detection system on ball-by-ball cricket data, testing it blind against a live event, and discovering why the three…

27

AI Advances @AIAdvances

Mar 17

Prompts disappear. Specs stay in the repo. That's why agentic engineering needs durable artifacts. Read this article by @kzdeb ai.gopubby.com/beyond-vibe-c…

Beyond Vibe Coding: The Artifacts Layer

Specs, plans, guidance, agent skills, and verification gates: the durable intent layer that separates responsible delegation from guessing.

1

21

AI Advances @AIAdvances

Mar 11

From Iron Man to Spatial Intelligence and the Future of Human-Centered XR by @poojakashyap ai.gopubby.com/what-happens-…

What Happens When AI Understands Space Like You Do?

Remember that scene in Iron Man, where Tony Stark stands in his workshop, surrounded by holographic interfaces, gesturing fluidly through…

1

98

AI Advances @AIAdvances

Mar 4

Most people use Claude Code as a chatbot with a terminal. @0xAedelon turned it into an operating system: 6 layers, 17 hooks, 32 skills, zero trust. The model is the engine. Everything else is deterministic code. Blueprint article: ai.gopubby.com/claude-code-s…

I Turned Claude Code Into an Operating System. Here’s the Blueprint.

How to configure Claude Code with hooks, skills, agents, and zero-trust permissions. A field-tested architecture that eliminates…

74

AI Advances @AIAdvances

Feb 28

The Most Expensive Thing in AI Training Is Waiting: Taming the Long Tail in RL. By @poojakashyap ai.gopubby.com/the-most-expe…

The Most Expensive Thing in AI Training Is Waiting: Taming the Long Tail in RL

There’s a scene in Interstellar that I always think of, which goes like this… the crew is aboard the Endurance, orbiting Miller’s planet…

1

52

AI Advances @AIAdvances

Feb 26

AI that writes the rulebook: a transformer model now generates Python code that builds entire families of quantum experiments, not just single solutions. By @poojakashyap #AI #QuantumComputing ai.gopubby.com/what-if-machi…

What if Machines Learn Patterns Humans Can’t See?

Let’s start with a simple analogy (the above image), imagine you walk into a room and see three Lego towers:

2

49

AI Advances @AIAdvances

Feb 23

What if a neural network knew the laws of physics before seeing any data? The AI revolution nobody's watching. ai.gopubby.com/when-deep-lea…

When Deep Learning Meets Physics

The quiet revolution happening inside neural networks, and why it matters more than ChatGPT

1

1

26

AI Advances @AIAdvances

Feb 18

Why do most AI pilots in finance fail while a few actually scale? This post by @lak_luster uncovers the hidden infrastructure and domain expertise needed to move from impressive demos to reliable, production-ready AI. ai.gopubby.com/the-platform-…

The platform and the people you need to build Vertical AI

Why AI in Finance requires Specialized Infrastructure and an interdisciplinary team

1

38

AI Advances @AIAdvances

Feb 17

AI is already reshaping what ends up on your plate. By 2030, algorithms will control food systems you never see. New investigation by @Yiannis_A_med: ai.gopubby.com/the-ai-algori…

The AI Algorithm Deciding What You’ll Eat in 2030

AI promises to help feed the world, but it’s placing decisions about global food security into algorithms that can fail catastrophically

22